Datasets:

KenLuo

/

RC-Sum-PLOS

like 0

10.1371/journal.ppat.1004874

Experimental Evolution of an RNA Virus in Wild Birds: Evidence for Host-Dependent Impacts on Population Structure and Competitive Fitness

Within hosts, RNA viruses form populations that are genetically and phenotypically complex. Heterogeneity in RNA virus genomes arises due to error-prone replication and is reduced by stochastic and selective mechanisms that are incompletely understood. Defining how natural selection shapes RNA virus populations is critical because it can inform treatment paradigms and enhance control efforts. We allowed West Nile virus (WNV) to replicate in wild-caught American crows, house sparrows and American robins to assess how natural selection shapes RNA virus populations in ecologically relevant hosts that differ in susceptibility to virus-induced mortality. After five sequential passages in each bird species, we examined the phenotype and population diversity of WNV through fitness competition assays and next generation sequencing. We demonstrate that fitness gains occur in a species-specific manner, with the greatest replicative fitness gains in robin-passaged WNV and the least in WNV passaged in crows. Sequencing data revealed that intrahost WNV populations were strongly influenced by purifying selection and the overall complexity of the viral populations was similar among passaged hosts. However, the selective pressures that control WNV populations seem to be bird species-dependent. Specifically, crow-passaged WNV populations contained the most unique mutations (~1.7× more than sparrows, ~3.4× more than robins) and defective genomes (~1.4× greater than sparrows, ~2.7× greater than robins), but the lowest average mutation frequency (about equal to sparrows, ~2.6× lower than robins). Therefore, our data suggest that WNV replication in the most disease-susceptible bird species is positively associated with virus mutational tolerance, likely via complementation, and negatively associated with the strength of selection. These differences in genetic composition most likely have distinct phenotypic consequences for the virus populations. Taken together, these results reveal important insights into how different hosts may contribute to the emergence of RNA viruses.

Viruses are constantly emerging into new areas and pose significant challenges to public health. Chikungunya and West Nile viruses (WNV), both mosquito-borne RNA viruses, are quintessential examples of how increased globalization has facilitated the expansion of viruses into new territories. Rapid evolution of both of these agents has contributed to their rapid spread and health burden. Thus, characterizing how selection shapes zoonotic RNA viruses in their natural hosts is important to understand their emergence. As an ecological generalist able to infect hundreds of bird species, WNV is an excellent tool to study how different animal hosts can differentially drive virus evolution. We examined the genetic composition and fitness of WNV produced during replication in wild-caught American crows, house sparrows and American robins, species that range in mortality following WNV infection (crows the highest, robins the lowest). We demonstrate host-dependent effects on WNV population structure and fitness. Our study provides insights on how different virus-animal interactions can influence the success of a virus in the next host and ultimately the success of virus emergence into new host systems.

RNA viruses pose some of the most complex, persistent and challenging problems facing public health and medicine. The ongoing outbreaks of avian influenza A(H7N9) virus (Orthomyxoviridae) in China [1], Ebola virus (Filoviridae) in West Africa [2], and chikungunya virus (CHIKV, Togaviridae, Alphavirus) and West Nile virus (WNV, Flaviviridae, Flavivirus) in the Americas [3,4] highlight the health and societal impacts imposed by RNA virus-induced diseases. Several factors contribute to the emergence of these agents and the continued burdens they impose on human health. Among these is their ability to undergo rapid evolution in new and/or changing environments. Well documented examples of RNA virus evolution leading to increased virus transmission include WNV and CHIKV. In both cases, small, conservative amino acid substitutions (residues with similar physiochemical properties) to the viral envelope proteins resulted in more efficient transmission by mosquito vectors [5,6]. Adaptive changes to RNA virus genomes first arise as minority components within a genetically complex population of related but non-identical virus variants. The genetic diversity present in naturally occurring RNA virus populations has been clearly shown through a large and expanding body of observational and experimental studies to be critical to their biology. For example, several studies have demonstrated that the diversity of an intrahost viral population, rather than the fitness of individual variants, correlates with pathogenesis, disease progression and therapeutic outcome [7–9]. Moreover RNA viruses have the capacity for rapid evolutionary change because within infected hosts, all single nucleotide mutations may be generated. This has been particularly clear in the case of WNV, an arthropod-borne virus (arbovirus) that persists in nature in enzootic cycles between ornithophilic mosquitoes (mainly Culex spp.) and birds. After its initial identification in the New York City area in 1999, WNV spread throughout the continental United States, producing the largest outbreaks of flaviviral encephalitis ever recorded in North America. The explosive spread of the virus was accompanied by the displacement of the introduced genotype by a derived strain that is more efficiently transmitted by local Culex mosquitoes [10]. Studies of intrahost population dynamics of WNV demonstrated that genetic diversity is greater in mosquitoes than in birds [11]. The selective basis for the host-specific patterns of WNV genetic diversity is that the strong purifying selection that predominates in birds is relaxed in mosquitoes [11,12]. In addition, the RNA interference-based antiviral response in mosquitoes creates an environment where negative frequency-dependent selection may drive rare variants to higher population frequency [13]. Moreover, WNV maintains both adaptive plasticity and high fitness by alternating between hosts that impose different selective forces on the virus population [14]. Nonetheless, important gaps remain in our understanding of how error-prone replication interacts with selective and stochastic reductions in viral genetic diversity under natural conditions. This is particularly the case for arboviruses, which tend to cause acute infection in vertebrates, with transmission occurring before the development of a neutralizing antibody response. Therefore, well-described mechanisms of immune selection such as those that occur during chronic hepatitis C and human immunodeficiency virus infections are comparatively weak during acute arbovirus infection of vertebrates. Thus, the ways that ecologically relevant, natural hosts can influence arbovirus genetic diversity remain poorly understood. WNV in particular provides an excellent experimental system to study the influences of natural vertebrate hosts on viral evolution. The virus infects a large number of wild bird species [15] with a wide-range of infection outcomes [16]. In addition, several studies have provided evidence that particular WNV variants may arise through adaptation to birds [17,18]. Therefore, we sought to determine whether different wild bird species may have distinct impacts on WNV population structure. Specifically, we allowed WNV to replicate in wild-caught American crows (Corvus brachyrhynchos), house sparrows (Passer domesticus), and American robins (Turdus migratorius), bypassing the mosquito portion of the arbovirus cycle in order to focus on the impact of different vertebrate environments on virus populations during acute infection. Virus was passaged in individuals of each species five times in order to amplify host-specific patterns of selection that may remain cryptic after a single passage. Bird species were selected on the basis of ecological relevance and resistance to WNV-induced mortality. American crows experience high viremia and mortality following inoculation with WNV [19] and can directly transmit virus to roost mates without mosquito involvement [20]; house sparrows experience high viremia and intermediate mortality [21] and are frequently involved in WNV perpetuation [22]; and American robins experience intermediate viremia but very low mortality [23] and can be drivers for human WNV risk [24]. Virus populations were characterized using next generation sequencing (NGS) and through in vivo fitness competition studies in birds and mosquitoes. Our findings demonstrate that relevant vertebrate hosts with varying levels of disease susceptibility differentially shape WNV population structure with direct impacts on fitness during host shifts. The WNV used in these studies was derived from an infectious clone of the NY99 genotype and is described in detail elsewhere [25]. Clone-derived WNV was passaged five times in wild-caught American crows, house sparrows and American robins. To avoid systematically selecting high- or low-replicating strains and population bottlenecks during passage, and since titers are highly variable in wild-caught birds, the sera from the individuals with the intermediate viral load were passed into the next cohort at a standard dose of 1000 plaque forming units (PFU). Virus titer was variable but did not change significantly or consistently during the course of passage (Fig 1A). Further, five passages in wild birds did not alter viremia production or mortality in crows and sparrows (S1A and S1B Fig). WNV replication and fitness after passage was assessed using young chickens and Culex quinquefasciatus mosquitoes to directly compare the viral populations in hosts not used for passaging and to remove the variability of wild-caught birds (e.g. age and infection history) (Fig 1B and 1C). Passaged virus (p5) was similar to the WNVic (p0) in peak viremia production in chickens (i.e. at 2 and 3 dpi) (Fig 1B). Fitness assays were used to directly compare passaged viruses to a standard reference WNV in head-to-head competition. These assays can detect subtle fitness differences that are inapparent in comparative studies. Competitive fitness of all wild-bird p5 WNV was significantly enhanced in chickens. Crow-passaged virus had the smallest fitness gains and robin-passaged virus the largest (Fig 1C). Fitness studies conducted in wild birds produced the same results as those in chickens (S1C Fig). Competitive fitness was slightly increased in mosquitoes, but no bird-specific differences were noted (Fig 1C, S1D Fig). At each passage virus was examined by NGS to determine whether the consensus sequence changed during passage and to characterize the diversity of intrahost viral populations (S1 Table, S2 Fig). WNV genome coverage was variable across the genome and between samples (S2A Fig), and positively correlated with viral population size (S2C Fig). The lower relative WNV genome coverage from robin sera can in part be explained by smaller intrahost viral population sizes and smaller virus to host RNA ratios. Approximately 68%, 29% and 7% of NGS reads aligned to the WNV genome from crow, sparrow and robin sera, respectively. Comparatively, 20% and 0.5% of the NGS reads aligned to the WNV genome from chicken sera and mosquito bodies, respectively. Three nucleotide mutations that led to consensus amino acid substitutions were detected though passaging in birds, but none became fixed (i.e. frequency = 1) in the population. In contrast, three consensus amino acid substitutions were detected after a single mosquito passage. All intrahost single nucleotide variants (iSNVs) > 0.02 frequency are listed in S2 Table. We estimated intrahost variation from NGS data to determine whether WNV population diversity was bird species-dependent. The mean number of unique iSNVs in each virus population was relatively constant between passages, but differences were apparent among bird species (Fig 2A). WNV populations passaged in crows five times (p5) had significantly more unique iSNVs than WNV passaged in sparrows and robins. In addition, the frequency of individual iSNVs increased during passage in a species-dependent manner: The mean iSNV frequency after p5 in robins was significantly higher than after p5 in crows or sparrows (Fig 2B). Despite these differences, the viral populations had similar Normalized Shannon entropies (SN), Hamming distances (i.e. SNVs per coding sequence) and amino acid substitutions per coding sequence after p5 in different species (Fig 2C). We examined the ratio of viral genome equivalents (GE) to PFUs and intrahost single nucleotide length variants (iLVs, including both insertions and deletions) to assess defective viral genomes in WNV populations during passage. Crow-passaged WNV had the highest GE:PFU ratio (Fig 3A) and the most unique iLVs (Fig 3B). In addition, a greater proportion of the iLVs in crows were found in subsequent passages compared to sparrows and robins (Fig 3C). The number of iLVs per coding sequence was positively correlated with the titer of infectious virus (Fig 3D). We then evaluated the possibility that greater levels of iLV carry though in crows, which can only occur via complementation (Fig 3C), were due to sampling artifacts. To do this, we used a hypergeometric test implemented in R that indicated that selecting 400 common iLVs in two samples of 600 from the total pool of available single-nucleotide iLVs (n = 51,490) was 0. Simulation studies confirmed that it is extremely unlikely that random sampling produced the observed data. Evidence for natural selection was assessed in WNV populations using intrahost neutrality tests. The proportion of mutations in each population that were nonsynonymous (pN) and the ratios of nonsynonymous to synonymous variants per site (dN/dS) were highest in the input p0 WNV population and decreased significantly during passage in each bird species (Table 1). Separate analysis of dN and dS shows that dN did not significantly increase during passage while dS increased significantly at p5 in all bird species, a hallmark of purifying selection. The Fu and Li’s F and Fay and Wu’s H statistics were obtained from reconstructed haplotypes. The F statistic at p1 and p5 was consistently negative, indicating that the haplotypes contained excessive amounts of rare SNVs, again indicative of purifying selection (Table 1). The H statistic measures an excess of high compared to intermediate frequency SNVs. The insignificant H values suggest that the deviations from neutrality were due to natural selection rather than selective sweeps (Table 1). Analysis of reconstructed haplotypes that arose during passage and high frequency iSNVs (i.e. frequency > 0.02) was conducted to minimize the impact of differences in sequencing coverage and to assess positive selection. 0.02 was selected as a cutoff for “high frequency” mutations because it includes the top 5% of a gamma distribution of all VPhaser2-accepted iSNVs. The proportion of iSNVs that were high frequency after p5 was the greatest within robin-passaged WNV populations (16.5%) compared to sparrows (4.9%) and crows (4.8%) (Fig 4A). Reconstructed haplotypes from high frequency iSNVs were then used to assess the selective pressures that lead to haplotype replacement during passage (Fig 4B). The ancestral p0 virus population was composed of a single dominant haplotype that remained dominant after a single passage in all bird species. After p5, the ancestral haplotype remained dominant in crows, but not in sparrows and robins. Furthermore, high frequency iSNVs from crows contributed significantly fewer amino acid substitutions per coding sequence compared to robins after p5 (Fig 4C). Examination of dN/dS, amino acid diversity and high frequency nonsynonymous iSNVs across the WNV genome demonstrated that, in general, selection was the strongest in the structural protein coding regions (Fig 4D and 4E). Specifically, passage in robins imposed significant selective pressures on the envelope (E) protein coding region that heavily targeted ectodomains (ED) I and II. The apparent selection of the nonstructural protein 4B (NS4B) from sparrow passaging is the result of a single high frequency nonsynonymous iSNV (S2 Table). Individual high frequency iSNVs fluctuated in frequency through passaging and all nonsynonymous high frequency iSNVs were unique to its passage lineage (i.e. no “signature mutations” were detected that served as markers for replication in any particular bird species, see S2 Table). The standardized variance in iSNV frequencies (FST) was then estimated from the coding sequence to determine the degree of genetic divergence among replicates within a passage and between passages (Fig 5). Viral populations from robins were more divergent compared to those from crows and sparrows. FST from WNV passaged once in young chickens was similar to wild-caught birds, but WNV passaged once in mosquitoes was much more divergent. These results are supported by analysis of haplotypes (S3 Fig). The p0 haplotype was still dominant in chicken p1 populations with a small minority of haplotypes containing single iSNVs, similar to wild birds (Fig 4B). In mosquitoes the ancestral haplotype became a minority after a single passage. We examined WNV genetic diversity during the course of passage in birds that experience varying mortality due to WNV infection to assess how different hosts influence virus population structure and fitness. Passage in each host was accomplished in three concurrent biological replicates in order to control for the impact of individual wild-caught birds that may vary in several ways that could impact virus replication. Titers during passage were highly variable between individuals. However, mean titers did not significantly change during the course of passage, indicating that replication competence was retained and that overt increases in competitive fitness were not selected through our passage strategy. Wild-bird passaged virus was similar to unpassaged WNV in viremia production. Only when more sensitive in vivo competitive fitness assays (i.e. comparative replication of the passaged and reference WNV in the same host) were conducted were changes apparent. Note that our definition of fitness here is restricted to the specific competition environment (within the bird or mosquito) and does not consider the larger ecological fitness required for maintenance in a complex arbovirus transmission cycle. Passage in all birds resulted in significant competitive fitness gains during replication in chickens. Interestingly, the fitness gains were smallest after WNV was passaged in the host that experiences the most mortality (crows), and largest in the most disease-resistant avian host (robins). Fitness gains were far less clear when virus competition was measured in mosquitoes. A limitation to our mosquito studies is that competition was conducted via intrathoracic inoculation, which bypasses the midgut, a major physiological barrier in mosquitoes. Intrathoracic inoculation was used because the volume of blood available and the virus titers would have likely made oral infection highly inefficient. Importantly, our results on WNV replication and fitness are supported by previous observations [14] indicating that high fitness is maintained through purifying selection in vertebrates, and that no tradeoff occurs when the virus is re-introduced into mosquitoes. Moreover, replicative fitness increases occur during passage in ecologically relevant wild birds, and these gains occur in a species-specific manner. To investigate the viral genetic and population determinants of the observed fitness gains, we characterized WNV at each passage using NGS. Our data suggests that although the overall complexity of the virus population was similar among different bird species, its composition, and the selective pressures that produced it appear to be bird species-dependent. Interestingly, WNV replication in the most disease-susceptible bird species seems to be positively associated with the number of unique iSNVs (i.e. mutational tolerance) and negatively associated with iSNV frequency (i.e. strength of selection). This observation requires further investigation using additional resistant and susceptible birds, but may provide important insights into which bird species are most likely to drive virus evolution toward fitness gains. Our data thus far suggests that more disease resistant birds such as robins would be most likely to fill this role as long as they produce sufficiently high titers to infect mosquitoes. In this study we used various neutrality tests to determine whether intrahost WNV populations from each bird species were evolving non-randomly through purifying selection. While these tests all measure slightly different aspects of genetic diversity, all clearly demonstrate purifying selection in birds. This result confirms previous studies of WNV passaged in young chickens [11], and indicates that our approaches to sequencing and analysis, although they differ significantly from those reported previously, produce results consistent with other methods. Our studies also provide some evidence for positive selection during bird infection. We found that WNV passage in robins resulted in more amino acid substitutions that reach high frequency compared to crows. In addition, the ancestral haplotype tended to be displaced by novel mutants that arose during passage in sparrows and robins. These data suggest that positive selection within hosts is stronger in less susceptible bird species [26]. Examination of patterns of variation across the WNV genome provides additional evidence for differences in host selective environment. We found, consistent with previous reports on dengue virus populations [27], the highest variant frequencies in ectodomains I and II of the E coding sequence of WNV passaged in robins. The mechanisms that lead to the emergence of these variants are not currently clear. Although the E protein contains most neutralizing epitopes, the earliest neutralizing antibody responses observed in birds generally occur at around 5 to 7 days post infection [23,28]. Other mechanisms that could impact selection on the E protein include resistance to the early antiviral states induced by type I interferon [29,30] and alternate methods for virus entry and uncoating of the viral RNA [31]; though these mechanisms need further investigation, especially in birds. Our results suggest that in relatively resistant hosts, novel variants may rise to high frequency within the context of purifying selection. The notion that positive selection occurs in robins is further supported by our data showing that virus diverged most during replication in them. It is, however, balanced by a lack of evidence of a selective sweep, i.e. a rapid reduction in genetic diversity as a novel variant becomes very prominent in the population. Clearly further studies are needed to confirm whether and how positive selection contributes to WNV population structure in birds. Compared to other RNA viruses, arboviruses have low long-term rates of amino acid substitution [32]. This is at least partially due to the fact that most mutations are deleterious because of evolutionary constraints on arbovirus genomes [33]. We provide evidence that accumulation of deleterious mutations, or defective viral genomes, is unequal between hosts; WNV populations replicating in wild-caught crows accumulate the most defective genomes, and WNV replicating in robins accumulate the least. Defective genomes are often found during laboratory and natural virus infections [17,34] and can persist through multiple rounds of transmission [35,36]. Using both bioassays (i.e. GE:PFU) and sequencing data (i.e. iLVs per coding sequence), we found that the accumulation of WNV defective genomes during infection was positively correlated with viral load. This apparent density-dependent selection of deleterious mutations likely occurs via functional complementation, which becomes more efficient as effective multiplicity of infection (MOI, i.e. intrahost viral load) increases [37,38]. In addition, high MOI environments tend to tolerate neutral mutations that can become deleterious in a new environment [39]. Taken together, these studies provide a framework to understand how WNV replication in high-viremic crows leads to a broader network of potentially deleterious mutations and limited selection for adaptive amino acid substitutions, especially when compared to WNV replication in robins. The rather modest fitness gains experienced by crow-passaged WNV support this observation. The results presented here shed light on the selective forces that shape WNV populations in nature. We demonstrate that selective pressures that control WNV populations seem to occur in a species-specific manner (Fig 6). All three bird species evaluated have been suggested to be significant drivers of WNV outbreaks, with robins receiving particular attention due to findings indicating that this species is more frequently fed upon by mosquito vectors [24]. During intrahost WNV replication, our studies suggest that disease-susceptibility is positively associated with mutational tolerance and negatively associated with the strength of selection. This means that robins also may better maintain high fitness in WNV populations than do birds that are more susceptible to disease. While it is tempting to speculate that robins are significant generators of WNV genetic diversity, we also confirm herein that mosquitoes are much more efficient in generating mutational diversity in the WNV system. Moreover, these data suggest that intrahost virus evolutionary dynamics are associated with host resistance to disease in several ways and provide an important insight towards the genetic and ecological factors that influence RNA virus emergence. Wild birds were collected from under US Fish and Wildlife Service (#MB91672A-0) and Colorado Parks and Wildlife (#13TRb2106) permits and with permissions from landowners. No endangered or protected species were caught or harmed during the study. Experiments involving animals were conducted in accordance with protocols approved by the Colorado State University (CSU) Institutional Animal Care and Use Committee (#12-3694A) and the recommendations set forth in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. A WNV infectious clone (WNVic) was previously constructed from an American crow kidney isolate collected during the 2000 outbreak in New York City [25,40]. The WNVic contains a naturally selected proline at amino acid site 249 in nonstructural protein 3 (NS3) allowing it to replicate to high titers in wild birds [18,41]. Wild birds were collected in Northern Colorado from 2013 to 2014 using mist nets (house sparrows and American robins) and cannon nets (American crows). All birds were bled prior to inoculation and serum was tested by plaque reduction neutralization test to confirm that all birds used for subsequent studies were WNV seronegative. The virus strain used to initiate the passage series was derived from a WNVic as previously described [25]. Virus was harvested from the supernatant of BHK cells transfected with linearized plasmid, stored at -80°C and used without further passage. Viruses were administered to birds by subcutaneous inoculation to the breast region with 1,000 WNV PFU/100 μl, a dose similar to mosquito transmission [42], in inoculation medium (endotoxin and cation-free phosphate buffered saline with 1% FBS). Birds were bled from the jugular vein at the time of peak viremia on 3 days post-infection (dpi). Serum was titered by standard plaque assay on African green monkey kidney cells (Vero, ATCC CCL-81) and stored at -80°C until used for subsequent passage or sequencing as described below. The first passage series utilized seven birds for each wild-caught species and the three birds with the median viral titers were used to start three independent replicate lineages, each including three naïve birds (i.e. replicates ‘a’, ‘b’, and ‘c’). From each group of three birds, the serum with the median viral titer was used to continue passaging to another cohort until five serial passages were completed. The WNVic derived virus was also passaged once in three young chickens for 3 dpi and two individual Cx. quinquefasciatus mosquitoes for 14 dpi to compare viral populations from commonly used laboratory vertebrate host and invertebrate vector models, respectively. See S1 Text for information about housing and care of wild-caught birds, chickens and mosquitoes. The infection phenotype of each WNV lineage after five passages (p5) in wild-caught birds was compared to the unpassaged (p0) WNV in the same bird species as virus passage, young chickens (two-days old), and Cx. quinquefasciatus mosquitoes (4–7 days post emergence). Viremia and survival was measured from birds were inoculated with 1,000 PFU of p5 or p0 WNV (n = 4–5 birds/virus) for up to 6 dpi. As defined here, competitive fitness compares the replication of a competitor virus (i.e. serial passaged p5 WNV) and a standard WNV reference (WNV-REF) during infection of the same host. Competitive fitness is quantified by the proportion of competitor to WNV-REF genotypes using sequence chromatograms (i.e. quantitative sequencing) [43]. WNV-REF was generated from an infectious clone as described above and in S1 Text and is indistinguishable from the parental virus in replication in cells and relevant organisms [44]. Competitive fitness assays of co-inoculated birds and mosquitoes with equally mixed WNV-REF and p5 competitor virus was conducted as described in S1 Text. Virus libraries were prepared for RNA sequencing on the Illumina HiSeq 2000 platform (Beckman Coulter Genomics, Danvers, MA) using the NuGEN Ovation RNA-Seq System V2 and Ultralow Library kit (San Carlos, CA) (See SI Text for more details). Fastq files containing read data were demultiplexed using CASAVA and custom scripts that impose high stringency (0 mismatches) in the barcode region of each read. The sequence of the input WNV strain was determined from three independent biological sequencing replicates of the input virus using the Trinity assembler [45]. 100 nt paired-end reads were then aligned to this “input” sequence using MOSAIK [46]. Duplicate reads were removed using the MarkDuplicates tool within Picard to limit the influence of PCR artifacts and multiply sequenced clusters on variant calling with Vphaser2 [47]. Variants with significant strand bias were removed to reduce the potential for false-positives [48]. Variants called using Vphaser2 were used for subsequent data analysis unless otherwise specified. Analysis was limited to the protein coding sequences; and iSNVs and iLVs (includes both insertions and deletions) were analyzed separately. Hamming distances from the p0 “input” virus were calculated for each population by dividing the total number of polymorphisms by the average coding sequencing coverage. Mean viral population complexity was calculated by the SN at each site using the following equation [49]: SN=−pi(Lnpi)+(1−pi)(Ln(1−pi))/LnN where p is the frequency of the iSNV at site i and N is the coverage at that site. At a single nucleotide position, a SN score of 0 indicates a single nucleotide was present (i.e. no polymorphism) while a score of 1 represents maximum complexity (i.e. equal numbers of alternate nucleotides). The SN at all protein coding sequence nucleotides loci were averaged to estimate the viral population complexity. High frequency iSNVs were subjected to an additional analysis to reduce the possibility that conclusions drawn from the complete dataset were dependent on extremely rare variants. To establish a threshold for “high frequency” iSNVs, all of the Vphaser2 accepted variants detected in this study (n = 6052) were log10 transformed, increased by 3.75 (to make all of the values positive) and fit to a gamma distribution, where α = μ2/s2 and β = E[μ]/s2, using R (data did not fit a beta distribution). An iSNV frequency >0.02 was determined to be in the upper 5% of the gamma distribution and was used to define high frequency SNVs detected through WNV passage in birds (n = 341 individual SNVs). The sequencing reads from p0, p1 and p5 were aligned to the WNV genome using mpileup from the VarScan2 software package [50] and haplotypes were reconstructed using QuasiRecomb 1.2 [51] with the flags ‘-r 97–10395’, to reconstruct haplotypes from the entire coding sequence with respect to reference genome numbering, ‘-K 1–10’, to use a bigger interval of generators and ‘-noRecomb”, to disable the recombination process because it was not expected from the viral population and to reduce the runtime. To increase haplotype specificity, the flag ‘-conservative’ was employed and analysis was restricted to haplotypes containing high frequency SNVs (i.e. >0.02). pN and dN/dS were used to test for intrahost selection [33]. DnaSP (version 5) [52] was used to determine the number of nonsynonymous and synonymous sites to calculate dN/dS using the Nei-Gojorori method [53] with the following modifications for NGS data. Nd and Sd (i.e. the numbers of detected nonsynonymous and synonymous mutations, respectively) were calculated for each viral population by the sum of individual nonsynonymous and synonymous VPhaser2 accepted iSNV frequencies and the passage consensus sequence was used to determine the number of nonsynonymous and synonymous sites. The number of nonsynonymous (7843.67) and synonymous (2455.33) sites in the ancestral p0 consensus sequence were used to determine that pN prior to selection is ~ 0.76. In addition, 50 most frequent haplotypes reconstructed from p1 and p5 from each bird species were analyzed using the Fu and Li’s F [54] and Fay and Wu’s H [55] statistical tests of neutrality in DnaSP with a window length of 100, a step size of 25 and the p0 consensus sequence as an outgroup to infer the ancestral nucleotide state. FST was used to estimate the extent of interhost genetic divergence using a scale between 0 and 1, and the extent of FST change between populations represents the degree of genetic divergence. Specifically, in-house FORTAN scripts were used to calculate FST using equations 1, 2 and 4 by Fumagalli et al. [56]. Intrahost SNV frequencies determined by mpileup and readcounts from the VarScan2 software package [50] were used to estimate the per site heterozygosity in biological replicates compared to the total population (e.g. all biological replicates within passage) at a single passage (i.e. intra-passage) and the per site heterozygosity between passage replicates (i.e. inter-passage). For estimation of the probability of resampling for the iLV data, we used the phyper command in R (www.R-project.org). We calculated that a total of 51,490 single nucleotide iLVs were possible by multiplying the length of the coding sequence (10,299 nt) by the 5 different kinds of iLVs that could occur at each site (one deletion and four different nt insertions). We then used phyper to obtain the probability of sampling overlap of 400 iLVs out of 600 sampled (reflecting a reasonable approximation of our observed data for crows) given that 51,490 iLVs are possible. Simulation studies were conducted in R by randomly sampling 600 individuals, with replacement, from a set of 51,490 and comparing the sets. T-tests, Kruskal Wallis tests, and correlation statistics were obtained using R and GraphPad Prism (La Jolla, CA).

10.1371/journal.pntd.0007210

Evaluation of a novel West Nile virus transmission control strategy that targets Culex tarsalis with endectocide-containing blood meals

Control of arbovirus transmission remains focused on vector control through application of insecticides directly to the environment. However, these insecticide applications are often reactive interventions that can be poorly-targeted, inadequate for localized control during outbreaks, and opposed due to environmental and toxicity concerns. In this study, we developed endectocide-treated feed as a systemic endectocide for birds to target blood feeding Culex tarsalis, the primary West Nile virus (WNV) bridge vector in the western United States, and conducted preliminary tests on the effects of deploying this feed in the field. In lab tests, ivermectin (IVM) was the most effective endectocide tested against Cx. tarsalis and WNV-infection did not influence mosquito mortality from IVM. Chickens and wild Eurasian collared doves exhibited no signs of toxicity when fed solely on bird feed treated with concentrations up to 200 mg IVM/kg of diet, and significantly more Cx. tarsalis that blood fed on these birds died (greater than 80% mortality) compared to controls (less than 25% mortality). Mosquito mortality following blood feeding correlated with IVM serum concentrations at the time of blood feeding, which dropped rapidly after the withdrawal of treated feed. Preliminary field testing over one WNV season in Fort Collins, Colorado demonstrated that nearly all birds captured around treated bird feeders had detectable levels of IVM in their blood. However, entomological data showed that WNV transmission was non-significantly reduced around treated bird feeders. With further development, deployment of ivermectin-treated bird feed might be an effective, localized WNV transmission control tool.

West Nile virus (WNV) is a mosquito-borne virus that causes significant disease and death every year in humans, domesticated animals, and wildlife. Control of WNV transmission is focused on controlling the mosquito vector through applications of insecticides directly to the environment. In this study, we evaluate a novel control strategy for WNV transmission by targeting the main mosquito bridge vector in the Great Plains region, Culex tarsalis, through its blood feeding behavior. Because Culex tarsalis favor taking blood meals from particular bird species, our strategy aims to target these bird species with endectocide-treated bird feed that will result in lethal blood meals for Cx. tarsalis. In this study, we developed a safe and effective formulation of ivermectin-treated diet that resulted in increased mortality for Cx. tarsalis blood fed on birds consuming this treated diet as compared to mosquitoes feeding on control birds. We also conducted a pilot field trial in Fort Collins, Colorado to test this strategy in a natural transmission cycle, which demonstrated promising results.

West Nile virus (WNV) is an arthropod-borne flavivirus, and the leading cause of domestically acquired arboviral disease in the United States [1,2], resulting in significant disease and death every year in humans, domesticated animals, and wildlife. From 1999–2017, >48,000 cases of human WNV disease and >2000 deaths were reported to the CDC [3], but the total number of individuals in the U.S. who have been made ill from WNV is estimated to be greater than 1 million, or approximately 1 of every 5 persons infected (>5 million infected individuals) [4]. Control of WNV transmission remains focused on vector control through larvicide and adulticide applications [5]. Larvicide applications are generally preferred to adulticide applications as they are more cost-effective and less environmentally-damaging due to more direct and efficient targeting of mosquitoes [6,7]. While previous studies have demonstrated the effectiveness of larvicide applications to catch basins, a common Culex larval habitat, in reducing the number of mosquitoes [8,9], the efficacy may vary significantly with suboptimal catch basin design or environmental conditions [10,11]. Aerial spraying can be costly [12], but is effective in reducing target mosquito populations [13–16], and has been linked to reductions in human WNV cases in a treated area relative to an untreated area [15] and in entomological measures of WNV risk [16]. Similar ground ultra-low volume application of adulticides may reduce target mosquito populations under ideal conditions, but studies have provided inconclusive data on their effect on WNV infection rates in mosquitoes or subsequent virus transmission [17–20]. Additionally, off-target effects can occur despite optimal calibration of adulticide applications to host-seeking and active times for target vector species [21–23]. Insecticide applications also often face community opposition due to environmental and toxicity/allergenicity concerns [24–28] and are often restricted to urban and semi-urban communities that can afford to fund them [29,30]. WNV is maintained in an enzootic cycle between Culex mosquitoes and avian hosts. The highest WNV disease incidence occurs along the Great Plains region of the United States [31], as the irrigated agriculture provides a supportive habitat for the main WNV bridge vector of the region, Culex tarsalis [32].Therefore, blood meals by Cx. tarsalis from often-bitten avian species may be utilized to selectively target adult females through their blood feeding behavior. Given that the majority of Cx. tarsalis blood meals on the northern Colorado plains may come from select species during the WNV transmission season [33], effective targeting of these preferred hosts with endectocide-treated bird feed could result in control of WNV transmission. Previous studies have assessed the use of systemic endectocides provided to wild animals to control tick vector populations. Pound et al. evaluated ivermectin (IVM)-treated corn that was fed to white-tailed deer (Odocoileus virginianus) in a treatment pasture to control tick populations [34]. Amblyomma americanum collections from treatment pastures showed a 83.4% reduction in adults, 92.4% in nymphs, and 100.0% in larvae compared to control pastures [34]. IVM-treated feed provided to O. virginianus, which is the definitive host for the reproductive stage of Ixodes scapularis, has also been explored as a method for controlling this vector of Lyme disease. Rand et al. provided an island community of white-tailed deer with IVM-treated corn for 5 consecutive spring and fall seasons [35]. A treatment effect was observed in island deer that reached target IVM sera concentrations resulting in reductions in adult tick density, engorgement, and oviposition rates as well as reduced rates of larval eclosion from any laid eggs compared to collections from untreated deer on a control island [35]. Dolan et al. also conducted a field study that targeted the rodent reservoirs of Lyme disease to reduce the infection prevalence of Borrelia burgdorferi and Anaplasma phagocytophilum with antibiotic-treated bait. Between treated and control areas, they found that B. burgdorferi prevalence was reduced by 87% and A. phagocytophilum by 74% in small mammals, and in questing nymphal ticks, B. burgdorferi prevalence was reduced by 94% and A. phagocytophilum by 92% [36]. A field study testing the passive application of topical acaricide during bait consumption showed reductions of 68% and 84% of nymphal and larval I. scapularis found on white-footed mice, accompanied by a 53% reduction in the B. burgdorferi infection rate of white-footed mice and a 77% decrease in the questing adult I. scapularis abundance between control and treated properties [37]. Rodent baits with feed-through and systemic insecticide activity have also been evaluated to control the phlebotomine sand fly vectors of zoonotic cutaneous leishmaniasis and visceral leishmaniasis. A wide variety of insecticides have been tested for efficacy against multiple phlebotomine sand fly species using larval and adult blood feeding bioassays in multiple rodents. Methoprene, pyriproxyfen, novularon, eprinomectin, ivermectin, and diflubenzuron have been tested for efficacy within the lab [38–41], while fipronil has been additionally tested in field studies [40,42,43]. Systemic insecticides have also been used to target plague transmission, where field trials have assessed imidalcloprid-treated bait for controlling flea populations in California ground squirrels (Spermphilus beechyi), black-tailed prairie dogs (Cynomys ludovicianus), and other rodents [44–46]. To our knowledge, this strategy of endectocide-treated baits has not been evaluated in birds for arbovirus control. IVM use in birds is primarily off-label; however, IVM has been administered to treat multiple species of parasites that infest birds, including falcons, cockerels, and chickens [47–50]. Moreno et al. characterized the pharmacokinetics, metabolism, and tissue profiles of IVM in laying hens (Gallus gallus) with IVM delivered using intravenous (IV) and oral routes [51]. For both IV and oral routes, expected pharmacokinetic profiles and tissue distributions consistent for a highly lipophilic drug were observed [51]. Bennett et al. demonstrated transfer of IVM through crop milk when adult pigeon pairs were given 3.3 μg/mL IVM dosed in drinking water and housed with brooding squab, and IVM was subsequently detected in squabs following 3 days of daily adult pigeon IVM dosing [52]. In this present study, we evaluated endectocide-treated bird feed as a systemic endectocide to target Cx. tarsalis. 50% lethal concentrations for selamectin, eprinomectin, and ivermectin were determined in artificial blood meals. IVM-treated bird feed was evaluated for safety and consumption rates in chickens. Mosquitocidal effects in Cx. tarsalis fed on IVM-treated birds were also characterized. Lastly, we present the results of a pilot field trial conducted in Fort Collins, CO in 2017 that examined the safety of IVM-treated bird feed in the field and efficacy on entomological indices of WNV transmission. Animal research was done under CSU IACUC study protocol 16-6552A. Animal euthanasia was applied using sodium pentobarbital as approved in the IACUC study protocol. Field research was done under Colorado Parks and Wildlife Scientific Collection License #17TRb2104 and Fort Collins Natural Areas Permit #914–2017. Cx. tarsalis (Bakersfield colony) were reared in standard insectary conditions (28 ˚C, 16:8 light cycle). Approximately 150 larvae were reared in roughly 3 gallons of water and fed 2.5 grams of powdered Tetramin fish food daily until pupation. Adults were housed at approximately 300 per cage and fed ad libitum sugar and water until separated for bioassays. Mosquito bioassays were performed to determine the lethal concentrations resulting in 50% mortality (LC50) by adding drug (eprinomectin, selamectin, and IVM) into defibrinated calf blood (Colorado Serum Company) at serial dilutions for artificial membrane feeding. Following blood feeding, Cx. tarsalis were knocked down with CO2, and fully-engorged females were collected and held for 5 days in the same insectary conditions. For all bioassays, mosquito mortality was recorded every 24 hours and analyzed using Kaplan-Meier survival curves and compared using Mantel-Cox (log-rank) test. LC50 values were calculated using a nonlinear mixed model with probit analysis [53]. Artificial membrane blood feeds were also used to test the effects of IVM and WNV on Cx. tarsalis mortality. The WNV strain used was a 2012 Colorado isolate propagated in Vero cells. Negative controls were DMEM (Dulbecco’s Modified Eagle Media) and DMSO (dimethyl sulfoxide) at the same volumes as WNV and IVM, respectively. For the concurrent blood feed of WNV and IVM, IVM at 73.66 ng/mL (LC75) and WNV at low titer (5x105 PFU/mL) or high titer (107 PFU/mL) were fed in a membrane blood meal to Cx. tarsalis and mortality was observed as described above. For the WNV-exposure followed by an IVM blood feed, mosquitoes were fed a first blood meal containing 107 PFU/mL of WNV or DMEM for a mock-exposure. Fully engorged females were sorted and held for 10 days, then fed a second blood meal containing 73.66 ng/mL IVM, after which fully blood fed females were sorted and mortality observed. 4–6 weeks old white leghorn chickens were divided into groups (n = 4) that were housed separately, and which were provided clean water daily and control (untreated) diet consisting of a cracked corn mix (Chick Start and Grow, Northern Colorado Feeders Supply) mixed with any additives that were also added to IVM-treated diet for 3 or 7 consecutive days. IVM-treated diet consisted of two formulations: an Ivomec formulation where liquid Ivomec (Merial) was mixed directly into the cracked corn mix and a powder IVM formulation where powder IVM (Sigma-Aldrich) was mixed into all-purpose flour at 5% and then added to the cracked corn mixture to aid in even powder distribution. Chickens were fed ad libitum and feed consumed by each group was measured daily. Chickens were weighed daily and observed for clinical signs of toxicity, including diarrhea, mydriasis, ptosis, stupor and ataxia. The amount of chicken feed consumed was compared between groups using the students t-test and chicken growth rates were compared using linear regression. Blood was collected from these chickens through venipuncture at the end of their IVM diet regimen and for two days following IVM diet withdrawal. Serum was then isolated from the blood samples and stored at -80°C until further analysis. Eurasian collared doves (Streptopelia decaocto) were captured by mist net in Wellington, CO and brought back to CSU. They were housed in groups of three and provided ad libitum clean water and either control diet or powder IVM formulation diet of 200 mg IVM/kg of diet for 10 days. Three doves were fed each control and powder IVM formulation diet and then used for mosquito bioassays. Mosquito bioassays following blood feeding on birds were conducted on the last day of the IVM diet regimen for each group and for two days following IVM diet removal. For direct blood feeding on birds, the downy breast feathers were trimmed, and the exposed bird breast was placed on top of the mosquito cage. The birds were gently restrained for 30 minutes while the mosquitoes blood fed through the mosquito cage organdy. Given the difficulties of direct mosquito blood feeding on live chickens, supplemental serum-replacement membrane blood feeds were also performed, where frozen chicken serum was used in reconstituted blood meals using red blood cells from defibrinated calf blood [54,55]. All research with animals was reviewed and conducted under authorization by the Colorado State University Institutional Animal Care and Use Committee, protocol 16-6552A. Colorado State University Animal Care and Use is Public Health Service (PHS) and Office for Laboratory Animal Welfare (OLAW) assured (#A3572-01), United States Department of Agriculture (USDA) registered (#84-R-0003), and Association for Assessment and Accreditation of Laboratory Animal Care (AAALAC) accredited (#000834). All chemicals used in derivatization were HPLC grade and purchased from Sigma-Aldrich. IVM was extracted from serum following methanol precipitation [56]. 400 μL of methanol was added to 100 μL serum and vortexed for 1.5 min. Methanol precipitation was carried out at -80˚C overnight. Samples were centrifuged for 30 min at 16,000 x g. Supernatants were transferred and evaporated to dryness using a Speedvac concentrator (Savant). The dry residue was dissolved in 20 μL acetonitrile. Samples were derivatized according to previously published literature [57]. A Waters 700 autosampler system was used to quantify IVM by high-performance liquid chromatography (HPLC)-fluorescence. A mobile phase of acetonitrile/water (3:1, v/v) was pumped through a C8 column (Waters, XBridge BEH C8 XP, 130 Å, 2.5 μm, 2.1x100 mm) at a rate of 0.45 mL/min. Excitation and emission spectra were 365 and 470 nm, respectively. 10 μL of derivatized sample was injected by the autosampler. Precision was quantified as coefficient of variation (%CV). This was calculated interday and intraday, evaluating drug-free chicken serum samples (n = 5) spiked with IVM at 25, 50, and 100 ng/mL. Instrument CV was 6.11%. Intraday CV ranged between 4.36 and 9.77%. Interday reproducibility was 15.39%. Retention time CV was 1.77%. The method was linear across the range of concentrations tested in the standard curve (3.125–100 ng/mL). Linear regression curves containing fortified IVM serum samples with concentrations of 3.125, 6.25, 12.5, 25, 50, and 100 ng/mL had a R-square value of 0.9974. Limits of detection and quantification were 1.56 ng/mL and 3.125 ng/mL, respectively. For the 2017 pilot trial, field sites were located in urban and suburban areas in the City of Fort Collins (mainly in city open space areas and near water sources) that were weekly mosquito trapping sites used by the city for WNV surveillance efforts and have been maintained since 2006 (S1 Fig). Six field sites were chosen based on historical WNV surveillance data from all city trapping sites as those having the highest number of WNV-positive Cx. tarsalis pools since 2006, while excluding trap sites in neighborhoods that are regularly treated with adulticides or used as sentinel sites for the Colorado WNV surveillance system by the state department of health. The 6 chosen sites were all in east Fort Collins and were randomly placed into the treatment group (3 sites; mosquito traps surrounded by IVM-treated bird feed stations) or the control group (3 sites; mosquito traps surrounded by control un-treated bird feed stations). At each field site, an array of three bird feed stations was placed in an approximate triangular perimeter around the mosquito trap at a distance of 50 m (S1 Fig). IVM-treated bird feed was used at a concentration of 200 mg/kg of diet, and the diet was a mixture of white proso millet, cracked corn, and flour (47.5:47.5:5, v/v/v). IVM-treated bird feed was changed daily to account for any effects of IVM degradation due to exposure, which also allowed for daily monitoring of any obvious adverse effects of IVM in local fauna. Motion-activated trail cameras were used to document bird visits to feeders, with each field site having a motion-activated trail camera placed at one of the three feeders. Photos were screened using the Sibley Guide to Birds [58]. Due to an overabundance of pictures, only a random sampling of 6 days from the 2017 season were counted. Bird trapping and sampling of their blood was performed at two IVM sites. Birds were caught using mist nets placed approximately 10 m from an IVM-treated bird feeder. Blood was collected from netted birds using jugular venipuncture and placed into serum separator tubes. Bird sera were analyzed using HPLC-fluorescence and a subset of samples was analyzed using LC-MS. Because 200 μL of blood could not be drawn from the sparrows as needed for HPLC-fluorescence quantification, IVM analysis for sparrows was only documented as presence or absence. Control sera from house sparrows caught in spring 2014 were used as negative controls. Serum from one IVM-positive grackle was also used in a serum-replacement blood feed with colony Cx. tarsalis for a mosquito survival bioassay. Mosquitoes were processed as part of the Fort Collins WNV surveillance program according to established protocols [59]. Briefly, mosquitoes were collected weekly by Vector Disease Control International using miniature CDC light traps baited with CO2. Mosquitoes were sorted to species and pooled into groups of typically no more than 50. Mosquito pools were screened at CSU using qRT-PCR using the following primer sequences: forward 5’ 1160-TCAGCGATCTCTCCACCAAAG 3’, reverse 5’ 1209-GGGTCAGCACGTTTGTCATTG 3’, probe 5’ FAM-1186-TGCCCGACCATGGGAGAAGCTC 3’ [59]. Bird sampling was done under Colorado Parks and Wildlife Scientific Collection License #17TRb2104 and Fort Collins Natural Areas Permit #914–2017. Chicken feed consumption was compared between groups using a t-test. Linear regression was done on chicken weights and the rate of weight gain was compared using Analysis of Covariance. For mosquito bioassays, survival was analyzed using Kaplan-Meier survival curves and compared using Mantel-Cox (log-rank) test. LC50 values were calculated using a nonlinear mixed model with probit analysis [53]. IVM sera concentrations from chickens were compared using ANOVA. IVM sera concentrations from individual chickens were correlated to cumulative mosquito morality from bioassays conducted on the respective chickens using Spearman correlation. The field trial utilized control and treatment sites located in the City of Fort Collins; however, it was an exploratory trial to test a new trial design and sites, and so was not powered for detecting differences in Cx. tarsalis abundance and WNV infection. Cx. tarsalis abundance from control and treatment sites were compared against each other using a generalized linear mixed model with negative binomial distribution that included site, week of trapping, and treatment. Cx. tarsalis abundance was also shown in comparison to historical data from 2006–2016 (which lacked any bird feed stations surrounding the traps). WNV infection rate was calculated as maximum likelihood estimate (MLE) using the Excel PooledInfRate Add In [60], but Fisher’s exact test was again used to compare the total number of WNV-positive and WNV-negative pools between control and treatment sites. Statistical analyses were done in GraphPad Prism (Version 7) and R (Version 3.3.1). Mosquitocidal concentrations of IVM, selamectin, and eprinomectin were determined with mosquito bioassays following blood feeds with serially diluted drug (S2 Fig). IVM had the lowest LC50 concentration at 49.94 ng/mL (Table 1) as compared to eprinomectin with a LC50 of 101.59 ng/mL and selamectin with a LC50 of 151.46 ng/mL. With the lowest effective concentrations, ivermectin was chosen for further characterization in birds. Potential interactions of IVM and WNV on Cx. tarsalis mortality were assessed in a simultaneous blood meal containing IVM (LC75) and WNV. Feeding with IVM only resulted in significantly increased mortality compared to DMSO controls; however, the observed 41% and 83% mortality for IVM control groups (Fig 1A and 1B) reflect the variability of mosquito bioassays, especially for intermediate ranges of lethal concentrations. WNV (both low and high titer) exposure in the absence of IVM did not affect Cx. tarsalis mortality over 5 days immediately after the blood meal (Fig 1A and Fig 1B), or following a second untreated blood meal 10 days later (Fig 1C). On the other hand, Cx. tarsalis given a concurrent blood meal containing low-titer WNV and IVM exhibited significantly increased mortality at 51% compared to the control IVM group not fed WNV with 41% morality (p = 0.0268, χ2 = 4.904) (Fig 1A). However, there was no significant difference (p = 0.2529, χ2 = 1.307) in mortality between Cx. tarsalis fed a concurrent blood meal containing high titer WNV and IVM compared to the control (Fig 1B). Similarly, Cx. tarsalis given a first blood meal of either DMEM control or high titer WNV, and then a second blood meal containing IVM 10 days later, showed no significant differences in mortality (p = 0.1637, χ2 = 1.940) (Fig 1C). Over 7 days of observation, there were no observable clinical signs of IVM neurotoxicity—diarrhea, mydriasis, ptosis, stupor, and ataxia–in groups that consumed either liquid Ivomec or powder formulations of IVM of 200 mg IVM/kg of diet. For the Ivomec formulation diet, the chickens consumed an average 59.3 g of feed per chicken daily. This was significantly less than the corresponding control group which averaged 121.6 g of feed per chicken per day (p = 0.0045, t = 3.490). Consequently, there was also a significant difference (p <0.0001, F = 19.45) in the rate of weight gain between Ivomec and control groups (S3A Fig). For the powder IVM formulation diet, the IVM group consumed 60.97 g of feed per chicken each day, which was not significantly different from daily control group consumption of 55.2 g of feed per chicken (p = 0.2928, t = 1.100). This was also reflected in similar rates of weight gain between powder IVM and control groups (p = 0.0680, F = 4.022) (S3B Fig). Cx. tarsalis mortality following blood feeding on IVM-treated chickens increased as IVM concentration within the diet increased (S4 Fig). There were significant differences in mosquito mortality following blood feeding on chickens given 50 mg IVM/kg of diet (p = 0.0132, χ2 = 6.146) and 100 mg IVM/kg of diet (<0.0001, χ2 = 86.48). However, the largest increase in mortality (p<0.0001, χ2 = 461.1) following blood feeding was at 200 mg IVM/kg of diet with 95.2% mortality in mosquitoes fed on IVM-treated chickens and 2.7% mortality in mosquitoes fed on control chickens. All subsequent experiments used IVM-treated feed at 200 mg IVM/kg of diet. For the Ivomec formulation at 200 mg IVM/kg of diet, there was a significantly increased mortality in mosquitoes blood fed on chickens consuming Ivomec-diet for either 3 or 7 days as compared to mosquitoes blood fed on control chickens (Fig 2; left and right panels, respectively). On the last day of Ivomec feed administered, for both 3 or 7 days, there was a significant increase (p<0.0001, χ2 = 80.22 and χ2 = 76.41, respectively) in mortality between mosquitoes blood fed on chickens consuming an Ivomec diet with upwards of 80% mortality as compared to mosquitoes blood fed on control chickens with less than 40% mortality (Fig 2A and 2B). This difference in mosquito mortality between treatment and controls decreased when the blood feed occurred 1 day following the withdrawal of the Ivomec diet in the treatment group (Fig 2C and 2D). After 2 days following Ivomec diet withdrawal, there was no significant difference in mosquito mortality between those blood fed on Ivomec-consuming chickens as compared to mosquitoes blood fed on control chickens in the 3 day group, but there was a significant difference in the 7 day IVM group (p = 0.0117, χ2 = 6.354) which is likely due to the variability in mosquito bioassays (Fig 2E and 2F). In addition, the time administered Ivomec-treated diets (3 vs. 7 days) did not affect mosquito survival curves following direct blood feeding on chickens, regardless if the mosquitoes were blood fed on the last day of chicken time on the diets, or if the chickens were 1 or 2 days post withdrawal of the diets (Fig 2, left vs. right panels). There was also significantly increased mosquito mortality in mosquitoes blood fed on chickens consuming the powder formulation of IVM (200 mg IVM/kg of diet) compared to mosquitoes fed on control chickens (S5 Fig). Because bioassays from the Ivomec formulation and a preliminary powder formulation indicated no differences between mosquitocidal effects for groups given IVM for 3 or 7 days, these and subsequent experiments focused on the 7 day time point. A direct blood feed of mosquitoes on chickens given a powder IVM diet for 7 days resulted in 92.3% mosquito mortality as compared to 25.7% mosquito mortality from those blood fed on control chickens (p<0.0001, χ2 = 41.23) (S5A Fig), while an indirect, serum-replacement blood feed using sera from chickens given a powder IVM diet for 7 days resulted in 79.0% mosquito mortality as compared to 16.7% mortality from those blood fed on control chicken serum (p<0.0001, χ2 = 42.83) (S5B Fig). Furthermore, the mosquito survival curves between those blood fed directly on IVM-treated chickens as compared to sera from IVM-treated chickens were significantly different (red lines in S5A Fig vs. S5B; p<0.0001; hazard ratio 2.007). At 1 day post-powder IVM diet withdrawal, there was still a significant difference (p = 0.001, χ2 = 10.86) in mosquito mortality between those directly blood fed on IVM-diet vs control-diet chickens (S5C Fig; 90.9% vs. 0% mortality). However, this mosquitocidal effect was not apparent in a serum-replacement blood feed derived from chicken blood taken 1 day after IVM diet withdrawal (p = 0.7445, χ2 = 0.1062) (S5D Fig). As above, the mosquito survival curves between those blood fed directly vs. indirectly on treated chickens 1 day post-diet withdrawal were also significantly different (red lines in S5C Fig vs. S5D; p<0.0001; hazard ratio 6.742). At 2 days post-IVM diet withdrawal, blood/serum from treated chickens was no longer mosquitocidal in either direct blood feeding (p = 0.8402, χ2 = 0.04065) or serum-replacement (p = 0.1792, χ2 = 1.804) assays (S5E and S5F Fig). Direct blood feeds of Cx. tarsalis were also conducted on six wild caught Eurasian Collared Doves fed either a powder IVM formulation diet of 200 mg IVM/kg or control diet in the laboratory (Fig 3). There was a significant difference in mosquito mortality (p<0.0001, χ2 = 60.34) with 88.5% mortality in Cx. tarsalis fed on IVM-treated doves as compared to 14.3% mortality from mosquitoes blood fed on control doves. Additionally, there were no clinical signs of IVM toxicity observed in this treated bird species. Neither the IVM formulation nor the time for which the chickens consumed IVM-treated diet resulted in significant differences in average IVM serum concentrations (p = 0.2715, F = 1.472) (Fig 4A, blue vs. green bars). On the last day of IVM diet, the average IVM serum concentrations (with SD) were 88.575 (±43.613) ng/mL for 3-day Ivomec, 45.255 (±70.051) ng/mL for 3-day powder IVM, 21.910 (±20.914) ng/mL for 7-day Ivomec, 45.745 (±33.852) ng/mL for 7-day powder IVM. Chicken IVM serum concentrations decreased following withdrawal of the IVM diet and were nearly undetectable at 2 days post-withdrawal, which corresponded with mosquito bioassay results showing decreases in mosquitocidal activity following IVM-diet removal. Additionally, IVM serum concentrations were correlated to resulting mosquito mortality from blood feeding on these corresponding IVM-powder fed chickens (Fig 4B). There was a higher correlation between IVM serum concentrations and mortality from serum-replacement feeds with a Spearman r of 0.8629 (P = 0.0007), while the correlation between IVM serum concentrations and mortality from direct blood feeds was 0.4153 (p = 0.3062). For a pilot trial testing IVM feed in a natural transmission cycle, feeder stations were placed in urban and suburban areas within the City of Fort Collins (S1 Fig) and randomized to treatment or control sites. Bird visits to IVM feeders at all sites were dominated by grackles with infrequent visits by house (Passer domesticus) and sagebrush sparrows (Artemisiospiza nevadensis) and black-capped chickadees (Poecile atricapillus) (Table 2). There were also two visits by blue jays (Cyanocitta cristata), and a few other birds which could not be identified from the photographs. A more homogenous mix of grackles, house and brewers (Spizella breweri) sparrows, blue jays, black-capped chickadees, bushtits, and squirrels visited control feeders (Table 2). Birds were also caught by mist net and their sera assayed for IVM at the end of the field season. Ten grackles and 5 sparrows were caught over 4 mornings of sampling on August 30th and September 2nd, 3rd, and 7th. Most birds had been observed feeding from the IVM-treated feeder immediately preceding mist net capture. Nine grackles and 4 sparrows (87% of tested sera) had detectable levels of IVM within their serum, and the negative control sparrow serum from 2014 had no detectable IVM (Table 3). Serum from grackle #5 (Table 3) was plentiful and thus further used in a LC-MS assay to confirm the presence of IVM, and also tested in a serum-replacement bioassay. Interestingly, even though the IVM serum concentration in grackle #5 was measured as 5.7 ng/mL, there was strong mosquitocidal effect from this serum (100% mortality within 2 days; p<0.0001, χ2 = 54.15) compared to control mosquitoes fed on control calf serum (Fig 5). Cx. tarsalis abundance over time in 2017 at the urban and suburban field sites was similar to historical data collected from the same traps for 10 years prior (Fig 6A). A generalized linear mixed model with negative binomial distribution did not find a significant difference between Cx. tarsalis abundance at IVM sites compared to control sites (p = 0.161, z = 1.401) (Fig 6B). The low number of WNV infections did not allow for robust statistical analysis, although MLE was calculated (Fig 6C). A combined Fisher’s Exact Test of all 6 field sites showed a non-significant decrease in the proportion of WNV-positive pools to WNV-negative pools among control and treatment traps (p = 0.2081) (Fig 6D). This study presents a novel characterization of IVM-treated bird feed as a systemic endectocide to control WNV transmission. Lab studies characterized the effects of IVM-treated bird feed in both domestic and wild birds, especially mosquitocidal effects in Cx. tarsalis blood fed on birds consuming this IVM-containing diet. In addition, a pilot field trial was performed over a WNV season to gather preliminary efficacy data on the effects of IVM-treated bird feed within a natural WNV transmission cycle between wild birds and mosquitoes. IVM was determined to be the most effective endectocide tested with the lowest lethal concentrations for Cx. tarsalis. In addition, there did not appear to be a synergistic effect of IVM and WNV on Cx. tarsalis mortality in either a simultaneous blood feed of IVM and high titer WNV or sequential blood feeds, the first containing WNV and the second containing IVM. There was a statistical difference between survival curves of Cx. tarsalis fed a concurrent blood meal of a low WNV titer IVM compared to Cx. tarsalis fed only IVM. However, this increased mortality was likely due to the variable survival response of mosquitoes to IVM particularly at intermediate lethal concentrations, rather than a biologically significant interaction between WNV and IVM as there was no mortality difference between mosquitoes fed a concurrent higher titer WNV+IVM blood meal compared to mosquitoes fed DMEM+IVM. There was also no difference between mosquitoes previously exposed to WNV and then fed IVM as compared to mosquitoes unexposed to WNV and then fed IVM. While there is a study suggesting that IVM can inhibit WNV replication by targeting NS3 helicase activity, this was an in vitro cell-culture study using mammalian cells, and the concentration of IVM needed to inhibit 50% of the RNA synthesis in the Vero cells infected with WNV was considerably higher than what was achieved in our chickens following IVM feed consumption [61]. No clinical signs of toxicity were observed in any of the birds consuming either formulation of IVM feed. This was not surprising as IVM is given therapeutically in bird species in a wide range of doses (0.2 mg/kg to 2 mg/kg), depending on route of administration. However, more detailed studies of IVM toxicity should be conducted in multiple bird species in future controlled experiments. Previous studies have identified neurotoxic effects in pigeons following long-term consumption of a diet containing avermectin [62,63], of which IVM is a safer derivative [64]. Specifically, Chen et al. observed clinical signs of neurotoxicity, ranging from reduced activity and food intake following avermectin consumption for 60 days on a 20 mg/kg diet, to ataxia and spasms following avermectin consumption for 30 days on a 60 mg/kg diet [63]. On the other hand, a characterization of IVM pharmacokinetics, metabolism, and tissue distribution in laying hens treated intravenously (400 μg/kg) or consuming IVM-treated water (400 μg/kg/day) for 5 days did not report any ill effects in the birds [51]. Following the intravenous injection of the hens, the highest IVM plasma concentrations (739.6 ± 50.2 ng/mL) were 30 minutes after administration and plasma concentrations remained below 10 ng/mL after 24 hours [51]. Mean IVM concentrations in our chickens fed exclusively on an IVM-containing diet for 3 and 7 days were approximately 45 ng/mL, and similarly we did not observe any neurotoxicity. It remains to be determined if these results vary among different bird species or longer times on the diet. However, in the field studies, it is unlikely that the IVM-treated bird feed was the sole or even primary source of food for the wild birds visiting the feeders given the abundance of alternative food sources during summer. While chickens on the powder IVM and control diets consumed equivalent quantities of food, there was a significant difference in feed consumption among chicken fed the Ivomec diet and their controls. This may be a result of the glycerol formal and propylene glycol carriers in Ivomec that could give an unpleasant taste, as propylene glycol has been identified as a unpleasant and unpalatable feed additive in cattle [65]. Consequently, the decreased Ivomec feed consumption relative to control feed consumption is likely responsible for the significantly reduced rate of weight gain in the Ivomec group as compared to controls. Chickens that consumed either a powder IVM or Ivomec diet reached mosquitocidal levels of IVM in their blood within 3 days, as demonstrated by both the IVM serum concentrations in the chickens as well as the significant difference in survival curves of mosquitoes blood fed on IVM-treated chickens compared to controls. There were no notable differences between either IVM diet formulations in mosquitocidal efficacy when considering either time to achieve a mosquitocidal effect and IVM persistence in chicken serum following IVM withdrawal. Furthermore, the time the chickens were placed on the two IVM diets (3 and 7 days) did not significantly affect mosquito mortality, serum concentrations, or the elimination time of IVM from serum following feed withdrawal. This is corroborated by the similar IVM serum concentrations at all time points among the different IVM administration times and formulations. A mosquitocidal effect, but no observable bird toxicity, was demonstrated for wild-caught Eurasian collared doves following consumption of the 200 mg IVM/kg diet, indicating similar mosquitocidal efficacy of the approach in one other bird species and thus potential application to other wild bird species in field settings. The mosquito mortality in control groups had a greater variation for direct blood feeds (17.75% CV) relative to control groups for serum-replacement blood feeds (3.57% CV), indicating that direct blood feeds results in more inherent variability in mosquito mortality. This increased variability could be a result of increased mosquito handling and rougher conditions during direct blood feeding on birds. It is also possible this higher variability is partly due to smaller sample sizes from the direct blood feeds due to the low success of our colony mosquitoes imbibing full blood meals from live chickens. Regardless, the higher variability among direct blood feed data led to a weaker correlation between IVM serum concentrations and mosquito mortality compared to that from serum-replacement blood feed data. However, despite this higher variability, cumulative mosquito mortality from these direct blood feeds was higher (consistently above 75%) compared to that from the serum-replacement feeds, and mostly independent of measured IVM concentration in the chickens’ sera. One likely possibility for this discrepancy is that the IVM concentration within serum extracted from venous blood may not always be an accurate representation of the IVM concentration in subdermal capillary blood on which mosquitoes blood feed. It has been previously proposed that because IVM is extremely lipophilic and sequestered in fatty tissues, there may exist a concentration gradient of higher IVM or IVM metabolite concentrations in adipose tissue and blood of the surrounding capillaries compared with venous blood [66]. This is also one explanation for the observation that the IVM serum concentrations in chickens correlated with higher cumulative mosquito mortality than would be predicted from the LCx values calculated using artificial membrane feeds. A useful future analysis would be to compare mosquito mortality results from direct skin blood feeding on chickens, membrane blood feeds using venous blood drawn from the chickens, and serum replacement blood feeds using unfrozen serum from the same chickens. The mosquitocidal effect from chickens on an IVM-containing diet did not extend past one day after IVM-feed withdrawal, and this corresponded with the IVM serum concentrations that were generally below detectable limits by two days post-IVM feed withdrawal. This could potentially be a concern for applying this strategy in the field as it would suggest that frequent bird visits would be necessary to maintain their mosquitocidal blood concentrations of IVM. However, our field data indicated that wild birds were visiting the bird feeders and did have detectable levels of IVM within their sera during multiple days throughout the trial. In addition, one grackle from our 2017 field trial had strongly mosquitocidal serum as assessed in a bioassay, even though the IVM concentration in that serum was surprisingly low. It is promising that a majority of the birds tested had detectable levels of IVM within their sera, indicating that there was an unexpectedly high coverage of IVM in captured birds. However, the placement of mist nets at roughly a 10 m distance from an IVM feeder may have biased the sampling towards birds that visited the feeder, so future studies should more intensively sample birds at wider radii from the feeders. Understanding IVM coverage and persistence within wild birds is an important component of determining the efficacy of this strategy and should be supplemented with detection of IVM in wild-caught blood fed Cx. tarsalis in future field seasons. This could also be coupled with mosquito survival bioassays using wild bird sera to assess mosquitocidal activity as we performed here. This use of IVM-treated feed as a systemic endectocide to control WNV transmission is based on targeting Cx. tarsalis by medicating its preferred host species. Previous studies in California implicate Cx. tarsalis as a regionally adaptive, opportunistic blood feeder with a preference for avian hosts, and the diversity of available blood meal sources is reflected in the composition of its blood meals [67–71]. Important avian hosts for Cx. tarsalis in small rural towns within Weld County, which is adjacent to our Fort Collins field site area, include American Robins, doves, and other Passeriformes [33]. American Robins are an important Cx. tarsalis blood meal source and WNV amplification host that does not frequent bird feeders and would not be targeted by this current strategy [33,72,73]. However, doves and passerines are preferred blood meal sources of Cx. tarsalis and contribute to the cumulative number of WNV-positive Cx. tarsalis at estimated rates of approximately 30% in June, 60% in July, and 85% in August [33]. This represents a large proportion of Cx. tarsalis blood meal sources and WNV-positive contributions from birds that consume grain and seed that could be targeted throughout the summer season. However, our trail camera data did not show a large proportion of visits from these species identified as regionally important. For example, grackles were predominantly visiting our IVM-treated feeders, while control feeders were visited mostly by grackles, blue jays, brewer’s sparrows, and squirrels. However, the single trail camera we employed per site may not have fully documented bird visits to other feeders at the field site. Camera placement was limited to tree-filled areas where a feeder could be placed with a camera locked to a tree across from the feeder, and this may have biased the camera data against bird species that feed in open space or brush rather than among trees. This limitation of the field camera data is illustrated by our detection of IVM in house sparrows caught by mist net, but we had no documentation of sparrow visits on the trail camera for this specific field site. An important future direction will also be to gather a more updated understanding of the Cx. tarsalis blood meal sources within urban and suburban area of the City of Fort Collins, which might allow for specific targeting of these bird species with attractive bird feed compositions and an optimized bird feeder design. In addition to a better characterization of avian blood meal sources for Cx. tarsalis, a more complete understanding of bird and Cx. tarsalis spatial dynamics is also important for determining the best placement for the IVM-treated feeders. Because our field sites were chosen based on historical mosquito and WNV surveillance, we did not account for crucial bird parameters that may have influenced mosquito sampling. For example, birds may have fed at the IVM-treated feeders and returned to their communal roosts where they would have been blood fed on by Cx. tarsalis [33,70,74], representing a treatment effect in a different population of Cx. tarsalis than sampled at our traps. Accounting for these bird-mosquito spatial dynamics by placing IVM-treated feeders near communal roosts of granivorous birds and sampling mosquitoes within close range may show the greatest entomological treatment effect, especially as Kent et al. gives an example of a house sparrow roost serving as both a major blood meal and amplification source of WNV-positive Cx. tarsalis [33]. While communal bird roosts could present a critical target, this strategy should continue to be tested in areas of increased human use such as parks and backyards. This highlights that future studies should also consider the best placement of bird feeders in the context of both human land use, and bird and mosquito interactions. Our pilot field trial was ultimately inconclusive and did not find a significant difference in Cx. tarsalis abundance or WNV infection due to IVM treatment. This is likely due to three field sites for each trial arm being underpowered to observe a significant effect. However, these preliminary field data will serve as important effect size variables with which to properly power future field trials. In addition, this strategy of controlling vector pathogen transmission with an endectocide like IVM is based on shifting the mosquito population age structure in a treatment area from older, infectious mosquitoes to younger, non-infectious mosquitoes, and is less dependent on reducing total mosquito abundance. This has been modeled, as well as observed with empirical data, in trials testing IVM for malaria transmission control [75,76]. We would also expect to see a shift in the age structure of the population to fewer older, infectious Cx. tarsalis and more uninfected, younger mosquitoes. However, our preliminary results from ovary dissections and parity scoring according to Detinova [77] showed consistently high parous rates within the field-caught Cx. tarsalis. This suggested that autogeny, or the ability to develop a batch of eggs without imbibing a blood meal, could be present among the Cx. tarsalis in our study area and confounded our data, and we chose to not conduct further parity scoring during our pilot field trial. As determining age structure of the wild Cx. tarsalis population would be additional way to evaluate this control strategy, future studies should integrate other age-grading techniques such as near infrared spectroscopy (NIRS) [78,79]. Our characterization of IVM as a systemic endectocide in birds demonstrates its feasibility to be developed into a novel WNV transmission control tool. We have demonstrated that birds readily consume IVM-treated feed in the lab and field with our formulation and concentration, while not displaying any observable clinical signs of toxicity following consumption. Furthermore, Cx. tarsalis mosquitoes blood feed on these IVM-treated birds and often die as a result. Our pilot field trial testing IVM-treated feed in natural transmission cycles within wild birds and mosquitoes was ultimately inconclusive, but did provide critical effect size variables to inform future trial design. Important future directions will be to optimize treated bird feed formulations for the field and better characterize the pharmacokinetics and pharmacodynamics of this diet within multiple bird species, especially in relation to mosquitocidal activity and physiological/clinical signs of toxicity. In addition, a more-updated, regionally-specific understanding of the blood meal host preferences of Cx. tarsalis across urban, suburban and rural habitats would allow for better targeting of these preferred host species through the design of an attractive bird feed composition, discriminating bird feeders, and optimized bird feeder location for application to different geographic areas. Finally, our field study provides an important template for future field studies across multiple WNV seasons that will be adequately-powered for measuring effect sizes in entomological and other outcomes.

10.1371/journal.pntd.0005622

Rapid deployment of a mobile biosafety level-3 laboratory in Sierra Leone during the 2014 Ebola virus epidemic

Ebola virus emerged in West Africa in December 2013. The high population mobility and poor public health infrastructure in this region led to the development of the largest Ebola virus disease (EVD) outbreak to date. On September 26, 2014, China dispatched a Mobile Biosafety Level-3 Laboratory (MBSL-3 Lab) and a well-trained diagnostic team to Sierra Leone to assist in EVD diagnosis using quantitative real-time PCR, which allowed the diagnosis of suspected EVD cases in less than 4 hours from the time of sample receiving. This laboratory was composed of three container vehicles equipped with advanced ventilation system, communication system, electricity and gas supply system. We strictly applied multiple safety precautions to reduce exposure risks. Personnel, materials, water and air flow management were the key elements of the biosafety measures in the MBSL-3 Lab. Air samples were regularly collected from the MBSL-3 Lab, but no evidence of Ebola virus infectious aerosols was detected. Potentially contaminated objects were also tested by collecting swabs. On one occasion, a pipette tested positive for EVD. A total of 1,635 suspected EVD cases (824 positive [50.4%]) were tested from September 28 to November 11, 2014, and no member of the diagnostic team was infected with Ebola virus or other pathogens, including Lassa fever. The specimens tested included blood (69.2%) and oral swabs (30.8%) with positivity rates of 54.2% and 41.9%, respectively. The China mobile laboratory was thus instrumental in the EVD outbreak response by providing timely and reliable diagnostics. The MBSL-3 Lab significantly contributed to establishing a suitable laboratory response capacity during the emergence of EVD in Sierra Leone.

A Mobile Biosafety Level-3 Laboratory (MBSL-3 Lab) and a well-trained diagnostic team were dispatched to Sierra Leone to assist in Ebola virus disease (EVD) diagnosis when the largest outbreak of EVD to date emerged in West Africa in 2014. This setup allowed for the diagnosis of suspected EVD cases in less than 4 hours from the time of sample receiving. The laboratory was composed of three container vehicles and was equipped with advanced ventilation system, communication system, electricity and gas supply system. Multiple safety precautions were strictly applied to reduce exposure risks. A total of 1,635 suspected EVD cases were evaluated from September 28 to November 11, 2014, and none of the staff members was infected with Ebola virus or other pathogens. The China mobile laboratory was thus instrumental in the EVD outbreak response by providing timely and accurate diagnostics. Therefore, the MBSL-3 Lab played a significant role in establishing a suitable laboratory response capacity during the emergence of EVD in Sierra Leone.

Ebola virus belongs to the Filoviridae family of enveloped viruses and contains a non-segmented negative-strand RNA genome [1,2]. Infection in humans can cause Ebola hemorrhagic fever, with exceptionally high case-fatality rates of more than 50% [3,4]. The incubation period of Ebola virus disease (EVD) is 2 to 21 days [5]. The clinical signs and symptoms are extremely similar to those of the Marburg virus and include fever, body aches, vomiting, diarrhea, rash and, in some cases, both internal and external bleeding [5]. Patients usually die of multiple-organ failure or hypovolemic shock. No licensed therapeutic or prophylactic treatments are currently available. The largest outbreak of EVD has been ongoing in West Africa since December 2013. As of April 15, 2015, 25,826 cases (10,704 deaths [41.4%]) had been reported by the World Health Organization (WHO) [6]. Although direct contact is the main route of transmission [7–10], EVD is still easily contagious, and healthcare workers have constituted a considerable proportion of all cases. In particular, by April 11, 2015, 864 healthcare workers (503 deaths [58.2%]) had been infected [6]. Ebola virus is classified as a biosafety level-4 agent. Clinical specimen inactivation should be performed in a biosafety level-3 laboratory, and subsequent to this step, routine testing can be performed in a biosafety level-2 laboratory. However, at the time of the outbreak, West Africa had few high-level biosafety facilities, so scientists had to work under difficult and dangerous conditions associated with potential exposure risks [11]. It would take a fairly long time, a large staff and many resources to construct a new fixed biosafety facility, thus delaying prevention and control of the epidemic. Therefore, a mobile unit [12,13] with both biosafety and flexibility was urgently needed to manage epidemics and emergent public health incidents such as the EVD outbreak. In September 2014, China responded to the appeal made by the United Nations and WHO and offered assistance to the government of Sierra Leone. A truck-based mobile biosafety level-3 laboratory (MBSL-3 Lab) and a well-trained diagnostic team were then dispatched and deployed to the Sierra Leone-China Friendship Hospital, in one of the hardest-hit areas, near Freetown, to assist in EVD diagnosis. The team members and aid supplies arrived on September 17, 2014. It took approximately one week to rebuild part of the hospital into multiple functional regions to meet the specimen testing requirements, including a specimen-receiving region, a supply-storage region, a waste-incineration region, a nucleic-acid-detection region, and a staff-rest area, among others. The MBSL-3 Lab was transported by an airlift jet aircraft (Antonov An-124 Ruslan, Russia) from Beijing Capital International Airport on September 24, 2014, at 03:00 (Beijing time) to Freetown International Airport on September 25, 2014, at 14:00 (Freetown time), with a flight duration of 43 h. It took another three and a half hours to drive the MBSL-3 Lab to the Sierra Leone-China Friendship Hospital. With strict training and standard operating procedures (SOPs), clinical specimen testing began within 60 h after the arrival of the MBSL-3 Lab, enabling the diagnosis of suspected EVD cases in less than 4 hours from the time of sample receiving. In total, 1,635 suspected EVD cases (824 positive [50.4%]) were tested from September 28 to November 11, 2014, and none of the staff members was infected with Ebola virus or other pathogens. Here, we provide a brief overview of the MBSL-3 Lab and the biosafety precautions applied to manage the EVD outbreak. This Ebola outbreak response was a humanitarian aid mission. The SOPs used were approved by the WHO and the Sierra Leone Ministry of Health and Sanitation (MoHS). The diagnostic results were released immediately after the specimen analyses were completed. Specimens were delivered to our worksite daily from two sources: the emergency operations center jointly established by the Sierra Leone MoHS and the China medical aid team who accompanied us and was also deployed to the Sierra Leone-China Friendship Hospital. When picking up the specimens, the staff wore one layer of personal protective equipment (PPE), including a protective suit (Lakeland INC or DuPont, USA), an N95 mask (3M, USA), an anti-impact goggle (3M, USA), two pairs of latex gloves with the inner pair taped to the protective suit and a pair of dedicated shoes and waterproof shoe covers (S1 Fig). The surface of the specimen bucket and the packing bag were disinfected by spraying with 0.25% chlorine-containing disinfectants. The staff extracted RNA in the BSL-3 Lab wearing two layers of PPE. The inner PPE included a protective suit, an N95 mask, a pair of inner gloves and a pair of dedicated shoes and waterproof shoe covers (S1 Fig). The external PPE included a HEPA filter-equipped powered air purifying respirator (3M, USA), a disposable sterilized surgical gown, a pair of external gloves and waterproof shoe covers (S1 Fig). The specimen bucket was opened within the biosafety cabinet. As Buffer AVL in the QIAamp Viral RNA Mini Kit (Qiagen, Germantown, MD, USA) was insufficient to inactivate samples [14], a combination of physical and chemical inactivation was performed to enhance the inactivation efficiency. The specimens were first inactivated by incubation in a water bath at 62°C for 1h before opening the tube cap to pipette the samples and were then further inactivated by the addition of Buffer AVL to the samples. RNA was extracted using the QIAamp Viral RNA Mini Kit (Qiagen, Germantown, MD, USA) according to the manufacturer’s protocol. All waste was first chemically inactivated (with 0.25% chlorine-containing disinfectant), then sterilized using a double-leaf autoclave and finally incinerated. Quantitative real-time PCR (Q-RT-PCR) assays were performed using a set of published primers and probes [15], targeting regions of the glycoprotein gene (F: 5’-TGGGCTGAAAAYTGCTACAATC-3’; R: 5’-CTTTGTGMACATASCGGCAC-3’; Probe: FAM-5′-CTACCAGCAGCGCCAGACGG-3′-TAMRA). RNA was amplified using the One Step PrimeScript RT-PCR Kit (TaKaRa, Japan), and 40-cycle Q-RT-PCR assays were run on the LightCycle 96 System (Roche, Switzerland). Melt curve analysis was performed to confirm the identity of the amplification products. The specimens were considered positive if there was an apparent logarithmic phase in the amplification curve, with melting point confirmed amplification products and the Ct value≤36 (Ct value<26, intense positive; 26≤Ct value≤ 36, weak positive). In contrast, the specimens were considered negative if there was no apparent logarithmic phase, with the Ct value undetermined, and they were considered suspect when 36<Ct value≤40. The MBSL-3 Lab was equipped with a -20°C freezer and a -80°C freezer, and there was another -80°C freezer outside the MBSL-3 Lab. As a result, we could store a total of 1500–2000 specimens. For short-term storage, namely, within 1 day, we stored the specimens at -20°C. For long-term storage, we stored the specimens at -80°C. The specimens were well packed and surface disinfected with 0.25% chlorine-containing disinfectant before storage. The Sierra Leone-China Friendship Hospital was guarded by the military guard of Sierra Leone, and the freezers were well locked. Every patient was assigned a unique Outbreak Case ID by the emergency operations center jointly established by the MoHS. Each time a sample was collected, the patient was asked to complete a “VIRAL HEMORRHAGIC FEVER CASE INVESTIGATION FORM”. The sample tube and the investigation form were marked with the Outbreak Case ID and patient name and were then delivered to us. Therefore, the Outbreak Case ID provided a unique number for tracking the patient, the specimen and the test result. The information in our testing report included the Outbreak Case ID, the Ct value yielded by Q-RT-PCR and the confirmed result (Yes/No/Suspect). According to an agreement with the MoHS, we usually did not contact hospitals directly. Instead, we submitted the testing report to the WHO and the MoHS, which was in charge of delivering the results to hospitals. In particular, the China medical aid team who came with us and was also deployed to the Sierra Leone-China Friendship Hospital could get testing results from us directly. The China MBSL-3 Lab arrived in Sierra Leone on September 25, 2014, and specimen tests were carried out within 60 h of its arrival. The worksite layout was shown in Fig 1. After receiving specimens, scientists sent them to the MBSL-3 Lab, where RNA was extracted. One room in the hospital was rebuilt and used for subsequent Q-RT-PCR analysis. The MBSL-3 Lab was powered by alternate use of 200kW diesel generators. Lab and household trash was incinerated away from the lab or structures in a pit. There were surveillance cameras all around the worksite and inside every experimental room, and scientists could watch real-time surveillance video and communicate with the experimenters in the laboratory. An overview of the composition of the China mobile laboratory diagnostic team and the team members’ tasks was shown in Table 1. One scientist was in charge of contacting the MoHS to coordinate issues such as sending specimens and releasing analysis results. In addition, eight scientists engaged in virus detection. Technical support personnel were in charge of the operation of the MBSL-3 Lab, including overseeing the water and electricity supply, maintenance and repair of equipment, sterilization and incineration of lab trash as well as watching and recording the daily experimental process. Two medical doctors monitored the health conditions of every staff member. The MBSL-3 Lab was composed of three container vehicles. The container encompassing the BSL-3 laboratory was called the main container (L×W×H: 9125×2438×2896mm); the second container, of the same size, was used for personnel cleaning and technology support and was called the auxiliary container; and the third container was the command container (L×W×H: 6300×2460×2100mm). As shown in Fig 2, the main and auxiliary containers were connected by an airtight soft connection and together formed a complete BSL-3 Lab. From the entrance to the inside, in order, there was the outside locker room (0-5Pa), the inside locker room (Buffer room-2, -10Pa), the semi-contaminated channel (-20±5Pa), the air lock room (Buffer room-1, -45±5Pa) and the BSL-3 laboratory (-70±10Pa). The doors were interlocking. The checklist for the different workplaces and instruments in the MBSL-3 Lab was listed in S1 Table. The MBSL-3 Lab provided triple protection for humans, specimens and the environment. The main performance of the MBSL-3 Lab was detailed as follows. “Four Flows” management were the key elements of biosafety measures in the MBSL-3 Lab (Fig 4). To assess the aerosol exposure risk when working in or around the MBSL-3 Lab, air samples were collected from the BSL-3 lab, locker rooms, water treatment room, equipment room, exhaust outlet and command container and were concentrated for EVD detection every 15 days (S1 Fig). Fortunately, all results were negative. We also collected swabs from the surfaces of potentially contaminated objects to determine whether there was an existing exposure risk (S2 Table). On one occasion, the pipette used to pipette samples from the blood-collection tubes tested positive for EVD, with a Ct value of 27.75. The diagnostic algorithm for laboratory testing and the rationale for positive/negative/suspect test results were presented in Fig 5. We repeated the testing of the suspect and negative cases and strongly recommended collecting specimens again if collection was performed <3 days post onset of symptoms. We found no evidence of RNA contamination during the entire operation. We added positive and negative controls to every experiment, and all controls produced the expected results. Overall, 1,635 suspected EVD specimens were tested from September 28 to November 11, 2014, primarily blood/serum samples (69.2%) and oral swabs (30.8%). The sample sources and test results were presented in Table 2. In total, 824 cases (50.4%) were identified as positive, and the positive rate of the swab samples (41.9%) was slightly lower than that of the blood samples (54.2%). The number of various paroxysmal public health events has been growing, and most have occurred in poverty-stricken areas. However, the resources for medical treatment, outbreak management and laboratory research are concentrated in developed regions, and substantial expenditure would be required to build new medical systems in these areas. Because epidemic situations are always urgent, scientists thus work under inadequate conditions and face exposure risks. Therefore, rapid, safe and flexible outbreak response capacity is urgently needed [17]. A mobile laboratory unit can easily be promptly deployed when needed and can provide a safe working environment, which will be a vital part of the outbreak response to emerging public health events or bioterrorism acts and will make great contributions to lessening and controlling epidemics. Several mobile units have previously been used in natural disaster scenarios [18,19], in health surveys [20,21], during the outbreak of severe infectious diseases [22–24] and in military campaigns [25]. Our MBSL-3 Lab meets the requirements of on-site collection, isolation, cultivation and detection of emergent infectious pathogens. This laboratory also protects humans as well as the environment and specimens, and it was designed to be functional in a field setting, even without logistical support. The major challenges in a remote location may be power supply and water supply, but there are ways to overcome them. There was an 80kVA (≈70kW) diesel generating set in the auxiliary container of the MBSL-3 Lab. Full fuel in the oil box can power the MBSL-3 Lab in continuous operation for 24h. We can bring as much fuel with us as possible using oil tanks, and wherever the MBSL-3 Lab can arrive, a refueling truck could also arrive. The MBSL-3 Lab is also equipped with a water storage tank and a water softener, and water can be re-supplied with water from a well or clear stream. If the experimenters could do not take a shower in the MBSL-3 Lab, the water requirement is not large, approximately 200L per day. In addition, the MBSL-3 Lab is equipped with a leveling system, but it still needs a 20m×8m level ground. This was the first time that we executed a mission in Africa. In total, 1,635 specimens were tested from September 28 to November 11, 2014, accounting for more than one quarter of the nation’s specimen volume during the same period. In all, 824 (50.4%) specimens were EVD-positive, representing 33.3% of the total number of confirmed cases reported in Sierra Leone during the same period. The maximum number of specimens that we could reasonably process in one day is approximately 120–150. We developed strict SOPs, adopted comprehensive protective measures and used comprehensive medical and logistical support systems to ensure safe and orderly performance of the virus diagnosis task. In particular, the “Four Flows” biosafety protocol was strictly followed. We monitored the exposure risk during clinical specimen testing. Air samples were collected from every workspace, and the test results were all negative, indicating that the working environment was relatively safe. The surfaces of potentially contaminated objects were also swabbed. On one occasion, the pipette used to pipette samples from blood-collection tubes tested positive. Given that a portion of the specimens contained only a small sample volume, the pipette had to be placed deep into the tubes and was easily contaminated by touching the inner wall. Therefore, it was suggested that the barrel of the pipette should be disinfected with disinfectant-containing gauze after pipetting each sample to avoid personnel infection and cross-contamination of samples. The test results played an important role in the disposal of symptomatic individuals and might, in a sense, determine their fates. For positive cases, the patients would be properly isolated and treated without visiting family members, and traditional religious funerals for the dead were forbidden. For negative cases, the patients would be separated from the positive cases and kept in an observation ward for follow-up testing or discharge to relieve the limited wards. Hence, the accuracy of the test results was crucial. False-positive results might lead to the individual being infected by positive patients, whereas false-negative results might lead to the spread of EVD to families and even the community. Our diagnostic algorithm suggested a suspect conclusion when 36<Ct value≤40 and strongly recommended resampling and considering clinical information and epidemiological links. Q-RT-PCR is now a preferred method for pathogen diagnosis due to its rapid and sensitive features [26], but it is prone to contamination and may result in false-positive results. Therefore, we conducted every experiment in the biosafety cabinet. The cabinet and PCR room were exposed periodically to ultraviolet radiation to eliminate nucleic acid contamination. Additionally, PCR tubes were never opened. Every control included in the PCR assays produced the expected result, indicating high experimental accuracy. Moreover, the MoHS was in charge of retrospective look at the disease progresses of the patients, and to date, we have not received any feedback regarding a false diagnostic case from the MoHS. We have shown that the positive rate of oral swabs was lower than that of blood samples. The technique and efficiency of swabbing might be one of the most important factors. Swab samples should be obtained by vigorous sampling to acquire sufficient biologic material for testing [27]. A quality-control PCR target (housekeeping gene target), such as Beta 2 Microglobulin (B2M), should be added for sample integrity assessment in the future. Our MBSL-3 Lab continuously worked for six months and managed 4,867 specimens for EVD diagnostics. During that time, the China CDC established a fixed BSL-3 Lab near the Sierra Leone-China Friendship Hospital for long-term surveillance and to serve as the public health system for future outbreaks and epidemics. Currently, the EVD epidemic situation is under effective control, and our MBSL-3 Lab has been proven to be an important force for disease control and emergency disposal.

10.1371/journal.ppat.0030058

Identification of a CCR5-Expressing T Cell Subset That Is Resistant to R5-Tropic HIV Infection

Infection with HIV-1 perturbs homeostasis of human T cell subsets, leading to accelerated immunologic deterioration. While studying changes in CD4+ memory and naïve T cells during HIV-1 infection, we found that a subset of CD4+ effector memory T cells that are CCR7−CD45RO−CD45RA+ (referred to as TEMRA cells), was significantly increased in some HIV-infected individuals. This T cell subset displayed a differentiated phenotype and skewed Th1-type cytokine production. Despite expressing high levels of CCR5, TEMRA cells were strikingly resistant to infection with CCR5 (R5)–tropic HIV-1, but remained highly susceptible to CXCR4 (X4)–tropic HIV-1. The resistance of TEMRA cells to R5-tropic viruses was determined to be post-entry of the virus and prior to early viral reverse transcription, suggesting a block at the uncoating stage. Remarkably, in a subset of the HIV-infected individuals, the relatively high proportion of TEMRA cells within effector T cells strongly correlated with higher CD4+ T cell numbers. These data provide compelling evidence for selection of an HIV-1–resistant CD4+ T cell population during the course of HIV-1 infection. Determining the host factors within TEMRA cells that restrict R5-tropic viruses and endow HIV-1–specific CD4+ T cells with this ability may result in novel therapeutic strategies against HIV-1 infection.

HIV-1 infection profoundly perturbs the immune system and is characterized by depletion of CD4+ T cells and chronic immune activation, which lead to AIDS. Although HIV-1 targets CD4+ T cells, it also requires a second receptor in order to infect the target cells. The majority of HIV-1 strains that are transmitted use a cell surface molecule called CCR5, which is expressed on a portion of T cells. In this manuscript we identify a subset of human CD4+ T cells, which we termed TEMRA cells, that express CCR5 but still remain resistant to infection. We show that HIV-1 infection is blocked in TEMRA cells after entry of the virus, but before it has a chance to integrate into the cellular genome. TEMRA cells are present at low frequency in HIV-1 uninfected individuals but greatly increase in some HIV-infected individuals, which correlates with higher CD4+ T cell numbers. These findings provide the basis for future studies to understand the role of TEMRA cells during HIV-1 infection and identify the host factors that could restrict the virus. This knowledge may be used to endow susceptible T cells with the ability to resist infection and result in novel vaccine or therapeutic strategies against HIV-1 infection.

Chronic immune activation and homeostatic disturbance of T cell subsets that accompany viral replication are hallmarks of HIV-1 infection [1–4]. The cause and implications of these profound quantitative and qualitative changes in CD4+ memory T cell subsets during HIV-1 infection are still not well understood [2]. Elucidating the causal relationships between perturbed naïve and memory T cell compartments during the course of HIV-1 infection could be critical in understanding its pathogenesis. Human T cells are categorized as naïve (TN) and memory (TM) subsets based on expression of CD45RA and CD45RO isoforms, respectively [5–8]. It is now known that memory T cells are comprised of distinct subsets that can be identified based on other surface markers and effector functions [9]. Sallusto and colleagues defined two CD4+ memory T cell subsets, termed central memory (TCM) and effector memory (TEM) cells [8]. TEM cells have low expression levels of the chemokine receptor CCR7 and lymph node homing receptor CD62L, express receptors for migration to inflamed tissues, and display immediate effector functions [8,10]. In contrast, TCM cells express high levels of CCR7 and lack potent effector functions. It has been proposed that TCM cells are responsible for maintaining long-term memory, and upon re-exposure to antigens, differentiate into TEM cells with effector functions. Prior studies indicated that HIV-1 preferentially infects memory, rather than naïve CD4+ T cells [11–16], possibly because of exclusive expression of the HIV-1 coreceptor CCR5 on memory T cells. Within the memory population, TEM cells are enriched for expression of CCR5 relative to other CD4 memory cells [17,18], suggesting that they may be primary targets for CCR5-tropic (R5-tropic) viruses that predominate in most infected persons. Because chronic HIV-1 infection disrupts the balance between naïve and memory T cell subsets [19], we characterized the distribution of these cells during HIV-1 infection. We found that a small subset of CD4+ TEM cells, which we called CD4+ TEMRA cells, were greatly increased in some HIV-infected individuals relative to uninfected individuals. Remarkably, CD4+ TEMRA cells displayed a specific post-entry block to R5-tropic HIV-1, despite expressing high levels of CCR5. Accumulation of this effector memory CD4+ T cell subset during chronic HIV infection could have important implications in understanding intrinsic resistance to the virus and perturbation of T cell compartments in infected individuals. The dynamics of T cell changes were studied in HIV-infected and HIV-uninfected individuals by staining their peripheral blood mononuclear cells (PBMCs) with monoclonal antibodies against CD3, CD4, CCR7, and CD45RO cell surface molecules. In most uninfected individuals, this analysis divides CD4+ T cells into three subsets that can be readily quantified: naïve T cells (TN; CD45RO−CCR7+), central memory T cells (TCM; CD45RO+CCR7+), and effector memory T cells (TEM; CCR7−) (Figure 1A, left panel). However, in HIV-uninfected individuals, a fourth subset (CD45RO−/dullCCR7−) was also observed (Figure 1A), albeit with a low frequency (0.5%–3%). This subset was greatly increased in some of the HIV-infected individuals (Figure 1A, right panel). Because these cells resembled a previously defined CD8+ T cell subset (called CD8+ TEMRA cells) with effector functions that expressed CD45RA with effector functions [20], we tentatively termed them CD4+ TEMRA cells (referred to as TEMRA cells hereafter). Conversely, we denoted the CD45RO+CD45RA−CCR7− effector memory CD4+ T cell subset as TEMRO cells. The relationship between TEMRA cells and HIV-1 infection was studied in 33 HIV-infected and 30 HIV-uninfected individuals (Figure 1B). The proportion of TEMRO and TEMRA subsets was significantly increased in HIV-infected individuals (Figure 1B). On the other hand, the proportion of TN cells was significantly decreased in HIV-infected individuals, while the proportion of TCM cells remained similar in both groups (Figure 1B). The high proportion of TEMRA cells found in HIV-infected individuals prompted further analysis of this subset. We hypothesized that TEMRA cells are a subset of effector memory CD4+ T cells, analogous to a subset recently described for CD8+ T cells with the same surface marker phenotype [20]. The four subsets of CD4+ T cells (TN, TCM, TEMRO, and TEMRA) obtained from HIV-infected and HIV-uninfected individuals were analyzed for expression of cell surface molecules known to be expressed differentially in naïve, memory, and effector T cells. All TN cells expressed CD28, CD27, CD7, and CD62L, with progressively less expression on CD4+ TCM, TEMRO, and TEMRA cells (Figure 2). In contrast, expression of CD11b, CD57, and HLA-DR were increased on TEMRO and TEMRA cells compared to CD4+ TN and TCM cells (Figure 2). In contrast to TCM and TEMRO cells, TEMRA cells also expressed high levels of CD45RA, similar to TN cells (Figure 2, top panel). This profile suggested that TEMRA cells are a subset of CD4+ effector memory T cells with a peculiar CD45RA+CD45RO−/dull phenotype. Differentiated effector memory T cells have a reduced proliferative capacity [8,20]. To assess the relative proliferative capacity of the different CD4+ T cell subsets, each subset was purified from an HIV-uninfected individual according to CCR7 and CD45RO expression as shown in Figure 1A, and stimulated with dendritic cells (DCs) pulsed with superantigen (staphylococcal enterotoxin B [SEB]). Activated T cells were counted at day 12 (Figure 3A). DC-mediated activation caused robust cell division of TN and TCM cells (Figure 3A), whereas TEMRO cells and TEMRA cells divided fewer times (Figure 3A). The reduced proliferative capacity of effector T cells correlates with a decrease in telomere length and with an increased propensity to undergo apoptosis [9]. To determine whether TEMRA cells undergo apoptosis similar to effector T cells, all T cell subsets were stained with a marker of apoptosis (Annexin V) before and after cells were stimulated through the T cell receptor (TCR) by anti-CD3 plus anti-CD28 antibodies for 18 h. A higher proportion of effector T cells underwent apoptosis compared to TN and TCM cells (Figure 3B). Levels of apoptosis were comparable between TEMRO and TEMRA cells before and after TCR stimulation (Figure 3B). A hallmark of TEM cells is secretion of greater quantities of cytokines when stimulated through the TCR, as compared to TN and TCM cells [8,10]. We therefore explored cytokine profiles of TEMRO and TEMRA subsets. As expected, [8,10] TEMRO cells secreted greater amounts of most cytokines assayed (IL-4, IL-5, IL-10, TNF-α, and IFN-γ) compared to TN and TCM cells (Figure 3C). TEMRA cells secreted high levels of IFN-γ, but much lower levels of IL-4, IL-5, or IL-10 compared to TEMRO cells (Figure 3C). This cytokine profile suggested that the TEMRA subset is skewed towards a Th1 phenotype. Recently, a cell surface molecule called CRTH2 was shown to be highly expressed on Th2 but not on Th1 cells [21]. To confirm Th1 skewing of TEMRA cells, we analyzed the surface expression of CRTH2. In agreement with the cytokine profile, significantly fewer TEMRA cells expressed CRTH2 compared to TEMRO or TCM subsets (Figure S1). Taken together, we conclude that TEMRA cells are differentiated effector memory T cells that are skewed toward a Th1 phenotype. Because TEMRA cells were proportionately increased in some HIV-infected individuals, we next investigated the susceptibility of these cells to HIV-1 infection. For these experiments, TN, TCM, TEMRO, and TEMRA cells purified from PBMCs of HIV-infected and HIV-uninfected individuals were activated through the TCR to render them susceptible to infection. The activated T cells were then infected with either R5-tropic replication-competent HIV (R5.HIV), CXCR4 (X4)–tropic replication-competent HIV (X4.HIV), or replication-defective viruses that only undergo a single round of replication and are pseudotyped with vesicular stomatitis virus glycoprotein G (VSV-G.HIV). Each virus used here encoded green fluorescent protein (GFP) that was used to quantify infection by flow cytometry at specific time points after inoculation [22]. Prior to the infectivity assay, we analyzed the expression of HIV-1 co-receptors CCR5 and CXCR4 on TN, TCM, TEMRO, and TEMRA cells isolated from an HIV-uninfected individual (Figure 4A). TEMRA and TEMRO cells expressed the highest levels of CCR5, while all four subsets expressed high levels of CXCR4 (Figure 4A). In addition, the median CCR5 expression was quantitated from 20 HIV-infected individuals, and the similar subset expression trends were confirmed (Figure S2). When each T cell subset isolated from an HIV-uninfected individual was challenged with R5.HIV, CD4 TCM and TEMRO cells were more susceptible to infection than TN cells (Figure 4B), most likely reflecting high CCR5 surface expression levels on these memory T cells (Figure 4A). In contrast, TEMRA cells were resistant to a high multiplicity challenge with R5.HIV (Figure 4B, top panel). This was an unexpected finding given the high cell surface CCR5 levels on TEMRA cells (Figure 4A). At day 12 post-infection, R5.HIV spread through the cultures, producing more infected TN, TCM, and TEMRO cells as compared to 5 d post-infection. Even at this late time point, TEMRA cells remained almost completely refractory to infection (Figure 4B, second panel). In contrast, TEMRA cells were similarly susceptible to infection with X4.HIV, as well as other T cell subsets (Figure 4B, third panel). Surprisingly, TEMRA cells were also 5- to 10-fold less susceptible to VSV-G.HIV infection than other T cell subsets (Figure 4B, bottom panel). We then sought to determine whether over time the TEMRA subset would progressively become more susceptible to infection post-activation, or whether these cells were being killed in culture by rapidly replicating virus. For this experiment, T cells were infected with R5.HIV or X4.HIV for 2 d at different multiplicities of infection (MOIs) and then washed to remove input virus. Infection was quantified based on GFP expression at different time points after inoculation (Figure 5A), and viral replication was assessed by quantifying HIV p24 protein in culture supernatants. R5.HIV replicated efficiently in TN, TCM, and TEMRO cells, but there was little or no replication in TEMRA cells (Figure 5B). In contrast, X4.HIV infected and replicated efficiently in all four subsets and rapidly killed most of the T cells (Figure 5A and 5B, right panels; unpublished data). Similar results were observed when primary HIV-1 isolates, utilizing different R5-tropic, X4-tropic, and R5X4-dual tropic HIV-1 envelopes that also express nef, were used (Figure 5C). The infectivity of TEMRA cells activated with SEB-pulsed DCs also remained identical (unpublished data). The surface marker CD57 identifies terminally differentiated cells [23], and expansion of CD57+ cells occurs in HIV-infected individuals [24]. Because TEMRA cells were enriched in CD57+ cells (Figure 2), we asked whether CD57+ T cells were differentially susceptible to R5-tropic or X4-tropic viruses. For this experiment, TEMRA and TEMRO cells were further subdivided into CD57+ and CD57− subsets by flow cytometry cell sorting. Both CD57+ and CD57− subsets of TEMRA cells were resistant to infection by R5-tropic virus, whereas both CD57+ and CD57− subsets of TEMRO cells remained susceptible to R5-tropic virus infection (Figure 6, top panel). However, the CD57+ and CD57− subsets of both TEMRO and TEMRA cells were similarly susceptible to X4-tropic viruses (Figure 6, bottom panel). Thus, the relative resistance of TEMRA cells to R5-tropic HIV was not attributable to enrichment with the CD57+ subset. We next investigated where in the HIV-1 life cycle R5-tropic infection of TEMRA cells was blocked. Because large numbers of cells were required for these experiments, we expanded TN, TCM, TEMRO, and TEMRA cells purified from PBMCs of HIV-uninfected individuals using SEB-pulsed DCs for 12 d in IL-2–containing medium. In order to verify that CCR5 expression levels were maintained on expanded T cell subsets and that TEMRA cells remained resistant to R5-tropic infection, CCR5 expression was determined post-activation and expansion (Figure 7A, top panel). The expanded subsets were then reactivated with SEB-pulsed DCs and subsequently infected with R5.HIV. Although the TEMRA cells maintained very high CCR5 expression, they remained resistant to R5-tropic infection (Figure 7A, bottom panel). We first asked whether the block of R5-tropic infection was at the level of fusion. For this experiment, we employed a recently developed reporter assay to quantify HIV particle entry [25]. Expanded TCM, TEMRO, and TEMRA cells were infected with either R5.HIV, X4.HIV, or VSV-G.HIV. Fusion of these three viruses with TEMRA cells was similarly efficient, whereas fusion was inhibited in both TN and Jurkat cells, which do not express CCR5, or when cells were pre-treated with T20, a fusion inhibitor (Figure 7B). Collectively, these data indicate that the R5.HIV infection block in TEMRA cells is post-fusion. We next conducted analysis of the stage in the HIV-1 life cycle at which R5-tropic and VSV-G pseudotyped virus replication was blocked in the TEMRA subset. Late reverse transcripts in cells infected with VSV-G.HIV, R5.HIV, and X4.HIV were analyzed. Infection was blocked at the level of reverse transcription in TEMRA cells infected with R5.HIV and VSV-G.HIV, suggesting an early block to infection in these cells that did not affect X4.HIV (Figure 7C). Because we did not see the accumulation of late reverse transcription products, we wanted to understand whether earlier steps in reverse transcription were impaired. Therefore, we investigated the initiation of reverse transcription of R5-tropic and VSV-G pseudotyped virus in TEMRA cells. Early reverse transcription was assessed by the presence of strong-stop, minus-strand viral DNA (R/U5 DNA) by quantitative real-time PCR. Early transcripts were not formed in TEMRA cells infected with R5.HIV or VSV-G.HIV (Figure 7D). These data suggest that the block in the HIV-1 life cycle occurs at or prior to the initiation of reverse transcription. Our findings that TEMRA cells are expanded in a portion of HIV-infected individuals and are highly resistant to R5-tropic infection prompted us to examine relationships between high TEMRA cells and CD4 numbers. Among HIV-infected individuals, the TN cell percentage correlated positively with absolute CD4+ T cell numbers (Figure 1B). Conversely, the total TEM cell (TEMRO + TEMRA) percentage correlated negatively with CD4+ T cell numbers (Figure 1B; unpublished data). To further delineate the association between TEMRO and TEMRA cell proportions and CD4 numbers, we subdivided infected individuals into three groups based on their total TEM cells (Figure 8). Infected individuals in whom the TEM percentage of their CD4+ T cells was similar to healthy individuals (bottom group) had high CD4+ cell numbers (Figure 8; unpublished data). In contrast, the group with a very high TEM cell percentage had low CD4+ T cell numbers, and all of these individuals had high levels of TEMRO cells (Figure 8, top group). Importantly, however, when we subdivided the infected individuals with median levels of TEM cells (Figure 8, middle group), a highly significant association between higher TEMRA cell percentage and higher CD4+ T cell numbers and higher TN cells was established (Figure 8). These results imply that a greater proportion of TEMRA cells within the effector T cell subset may identify individuals with better preservation of CD4+ cell numbers, and possibly slow HIV-1 disease progression. Our investigation of memory T cell subsets during HIV-1 infection led to the discovery of a unique subset of CD4+ T cells called CD4+ TEMRA cells. We found that these cells are highly susceptible to infection by X4-tropic HIV-1 but are almost completely resistant to R5-tropic HIV-1 despite high levels of cell surface CCR5 expression. These cells are also relatively resistant to infection by VSV-G pseudotyped HIV-1. Our findings are consistent with a recent ex vivo analysis of T cell subsets from HIV-infected individuals, which demonstrated that CD4+CD57+ effector memory T cells were associated with approximately ten times fewer copies of viral DNA than TCM cells [23]. Although both CD57+ and CD57− subsets of TEMRA cells displayed the same R5-tropic HIV-1 infection (Figure 5C), overall, CD57+ cells are more enriched within TEMRA cells (Figure 2). Thus, TEMRA cells represent the first unique subpopulation of CD4+ T cells that are uniquely resistant to HIV-1 infection and may emerge as a consequence selection during infection. Further studies are required to elucidate how TEMRA cells can be resistant to R5-tropic infection despite high levels of CCR5 expression, yet remain susceptible to X4-tropic viruses. In order to exclude that this restriction was at the level of post-entry and not because of downregulation or block of CCR5 by beta-chemokines, we showed that 1) TEMRA cells permitted entry of R5-tropic HIV-1 as measured by the BlaM-Vpr virion fusion assay, 2) TEMRA cells continued to express high levels of CCR5 at the time of infection, 3) and TEMRA cells were partly less susceptible to VSV-G pseudotyped viruses that bypass the coreceptor requirement. Taken together, these results indicate that the post-entry pathway followed by R5-tropic HIV-1 may differ in TEMRA cells compared to other CD4+CCR5+ T cell subsets and to X4-tropic HIV-1–infecting TEMRA cells. It is conceivable that either signaling or the entry pathway through the CXCR4 receptor elicits intracellular events needed for HIV replication or bypasses mechanisms that otherwise restrict HIV-1 in TEMRA cells. Elucidating cellular mechanisms that determine why some, but not all, CCR5-expressing CD4+ T cells are permissive to R5-tropic HIV-1 infection could provide clues to identify natural cellular HIV-1 barriers. Our findings suggest that at least one subset of primary human T cells display intrinsic restriction that limits HIV-1 infection. The presence of differentiated TEMRA cells in HIV-1 infected individuals and in uninfected individuals, albeit at lower frequency, suggests that these cells expand and survive during the course of the normal immune response. These findings also pose several important questions: How do TEMRA cells arise? Are they repeatedly stimulated memory T cells? What aspect of the TEMRA cell differentiation program renders them resistant to HIV-1 infection? For example, TEMRA cells displayed a preferential Th1 phenotype and exhibited a reduced proliferative capacity as well as a cell surface marker and cytokine profile characteristic of highly differentiated T cells. A subset of CD8+ T cells that are CD45RA+CD27− (CD8+ TEMRA cells) has been shown to display similar phenotypic features to CD4+ TEMRA cells characterized here [8,20,26]. It is not yet clear whether CD4+ TEMRA cells are functionally similar to CD8+ TEMRA cells or what role these subsets play during chronic viral infections. The homeostatic mechanisms that induce and maintain CD4+ TEMRA cells also remains to be determined. Our finding that TEMRA cells correlate with higher CD4+ T cell numbers in a portion of HIV-infected individuals suggests that virus infection may positively drive selection for HIV-resistant cells in vivo, a phenomenon previously observed only in cell culture but usually involving loss of CD4 expression. Studies using animal models for HIV-1 infection may aid in determining whether there is a causal relationship between virus infection and selective enrichment of the TEMRA subset. Remarkably, HIV-infected individuals whose TEM cells were composed mostly of TEMRA cells were significantly associated with higher CD4+ T cell and TN cell levels. How TEMRA cells accumulate or expand in HIV patients, and whether they have a protective role against progression of disease, remains to be determined. Memory and effector T cells are enriched for CCR5 expression [17,18], suggesting that they are targets for HIV-1, especially T cells resident in the gut tissue [27–30]. It is conceivable that after continuous destruction of susceptible TEMRO cells, an HIV-resistant subset of TEMRA cells is selected. Alternatively, TEMRA cells may have a protective role against HIV-1 infection, perhaps because HIV-specific T cells are enriched in this subset. If TEMRA cells contain a high proportion of HIV-specific effector T cells, this would overcome a potential Achilles' heel of the immune response during HIV-1 infection; that is, CD4+ T cells that are activated by HIV-1 antigens themselves become highly susceptible targets for the virus [31]. Conferring an HIV-resistant ability to HIV-1–specific CD4+ T cells could lead to novel strategies aimed at potentiating a protective immune response against HIV-1 infection. During the primary and asymptomatic phases of HIV-1 infection, R5-tropic viruses predominate, whereas X4-tropic viruses are found in about 50% of infected individuals at late stages of HIV disease [32–34]. A more rapid decline in total CD4+ T cell counts is often associated with a switch from R5-tropic to X4-tropic HIV or R5/X4 HIV variants [35]. At present, it is unclear whether the switch to X4-tropic viruses is a cause or a consequence of the collapse of the immune system. Because TEMRA cells remain highly susceptible to X4-tropic viruses, it would be expected that these cells would also be rapidly depleted when an X4-tropic switch occurs. If TEMRA cells contain HIV-specific T cells or play some other protective role against infection, then elimination of these cells by X4-tropic viruses would further weaken the immune response against HIV-1 and facilitate immunological deterioration. In summary, our results demonstrate that CD4+ TEMRA cells are present at a higher frequency in HIV-infected than uninfected individuals and are resistant to R5-tropic HIV infection, but not to X4-tropic HIV-1 infection. Studies focused on emergence of these effector memory T cell subsets will contribute to a better understanding of HIV-1 pathogenesis and the role of these cells during normal immune responses. Decoding the precise molecular mechanism of the intrinsic resistance of TEMRA cells to R5-tropic infection may have significant implications for developing novel approaches to endow this unique phenotype on HIV-1–susceptible T cells. PBMCs were separated from blood of HIV-uninfected and HIV-infected individuals through Ficoll-Hypaque (Pharmacia, http://www.pfizer.com). Resting CD4+ T cells were purified as previously described [22] and were at least 99.5% pure as determined by post-purification FACS analysis. To purify naïve, central, and effector memory subsets, purified CD4+ cells were stained with CCR7 and CD45RO antibodies, and CD45RO−CCR7+ (TN), CD45RO+CCR7+ (TCM), CD45RO+CCR7− (TEMRO), and CD45RO−CCR7− (TEMRA) subsets were sorted using the flow cytometer (FACSAria; BD Biosciences, http://www.bdbiosciences.com). The culture medium used in all experiments was RPMI (Cellgro, http://www.cellgro.com) and prepared as described before [22]. All cytokines were purchased from R&D Systems (http://www.rndsystems.com). In some experiments, TEMRO and TEMRA subsets were further subdivided into CD57+ and CD57− subsets by flow sorting by staining purified CD4+ T cells with CCR7, CD45RO, and CD57 antibodies. Monocyte-derived DCs were generated as previously described [22]. The superantigen SEB (Sigma, http://www.sigmaaldrich.com) was used to stimulate resting T cells in the presence of DCs [36]. Uninfected individuals were adults (ages 21–64, mean age was 32) with no history of HIV infection. Whole blood samples from adult participants with HIV infection were obtained during routine primary care visits. Among the HIV-infected individuals, 76% were Caucasian, 82% were male, the median (range) age was 41 (28–59) years, and 79% were receiving potent antiretroviral therapy. Median (IQR) CD4+ T cell and log10 plasma HIV-1 RNA values were 380 (270–592) cells/mm3 and 2.7 (2.6–3.8) copies/ml plasma, respectively, and 50% had fewer than 400 HIV-1 RNA copies/ml in plasma. There were no selection criteria based on race or sex. All participants provided written informed consent that was approved by the Vanderbilt Institutional Review Board. VSV-G pseudotyped replication-incompetent HIV were generated as previously described [36]. R5-tropic and X4-tropic replication-competent viruses were prepared similarly by transfecting 293T cells with HIV that encodes either R5-tropic (BaL) or X4-tropic (NL4–3) envelope and EGFP (Clontech, http://www.clontech.com) in place of the nef gene as previously described [37]. Wild-type virus (NL4–3) with X4-tropic or with R5-tropic envelope (BaL) and virus (R8) encoding heat stable antigen (HSA) in place of vpr [38] with intact nef gene were obtained from Chris Aiken (Vanderbilt University). Additional viruses used in this study were as follows. NL4–3–based proviral constructs encoding Env genes from R5-tropic proviral 92MW965.26, NL JRFL, NL YU2, and dual-tropic NL89.6 were obtained from Paul Bieniasz (Aaron Diamond AIDS Research Center) and have been previously described [39]. R5-tropic virus JRCSF and X4-tropic virus R9 were obtained from Vineet KewalRamani (National Cancer Institute [NCI]). Typically, viral titers ranged from 1 × 106 to 5 × 106 IFU/ml for replication-competent viruses and 10 × 106 to 30 × 106 IFU/ml for VSV-G pseudotyped HIV, as titered on CCR5-expressing Hut78 T cell lines (gift of Vineet KewalRamani, NCI). Viral replication in T cell cultures was determined by measuring p24 levels within supernatants by an ELISA. To determine apoptosis, T cells were stimulated with α-CD3 (OKT-3, ATCC)–coated plates in the presence of soluble α-CD28 (1 μg/ml; Pharmingen, http://www.bdbiosciences.com) for 18 h. T cell apoptosis was measured by PE-conjugated Annexin V according to manufacturer's instructions (BD Biosciences). Cytokines (IL-2, IL-4, IL-5, IL-10, TNF-α, IFN- γ) in the supernatants were assayed using a commercially available cytometric bead assay (CBA) (BD Biosciences) [40], and analyzed using CBA 6-bead analysis software (BD Biosciences). T cells were stained with the relevant antibody on ice for 30 min (chemokine receptor staining performed at room temperature for 20 min) in PBS buffer containing 2% FCS and 0.1% sodium azide. Cells were then washed twice, fixed with 1% paraformaldehyde, and analyzed with a FACSCalibur or FACSAria flow cytometer. Live cells were gated based on forward and side scatter properties, and analysis was performed using FlowJo software (Tree Star, http://www.treestar.com). The following anti-human antibodies were used for staining: CD3, CD4, CD8, CD45RO, CD45RA, CD28, CD27, CD11b, CD57, CD7, CD62L, HLA-DR, CCR5 (all from BD Biosciences), CCR7, and CCR4 (R&D Systems). The CRTH2 antibody used for these experiments has been previously described [41]. Secondary goat-anti-mouse antibodies were conjugated with allophycocyanin or PE (BD Biosciences). For the intracellular p24 stain, fixation and permeabilization was performed using a commercial kit (BD Biosciences) according to the manufacturer's instructions. Subsequently, cells were stained with anti-p24 for 1 h, followed by goat-anti-mouse conjugated to allophycocyanin for 30 min. HIV fusion assays were performed essentially as previously described [25]. Briefly, viruses carrying a β-lactamase reporter protein fused to the amino terminus of the virion protein Vpr (BlaM-Vpr) were added to expanded T cell subsets at 37 °C for 2 h to allow virus–cell fusion. CCF2/AM (20 μM; Aurora Biosciences Corporation, http://www.vrtx.com) was added, and the cultures were incubated for 14 h at room temperature. Cells were pelleted and resuspended in PBS, and the fluorescence was measured at 447 and 520 nm with a microplate fluorometer after excitation at 409 nm. Uncleaved CCF2 fluoresces green, due to fluorescence resonance energy transfer between the coumarin and fluorescein groups; however, cleavage by BlaM results in the dissociation of these fluorophores, and the emission spectrum shifts to blue. Thus, the ratio of blue to green cellular fluorescence is proportional to the overall extent of virus–cell fusion. Fluorescence ratios were calculated after subtraction of the average background fluorescence of control cultures containing no virus (blue values) and wells containing PBS (green values). Viral DNA was quantified by real-time PCR using an ABI 7700 instrument (PE Biosystems, http://www.appliedbiosystems.com) with SYBR Green chemistry. The reaction mixtures (25 μl total volume) contained 2.5 μl of infected lysate, 12.5 μl of 2x SYBR Green PCR Master Mix (PE Biosystems), and 50 nM of each primer. A standard curve was prepared from serial dilutions of HIV plasmid DNA. The reactions were amplified and analyzed as previously described [42]. The sequences of primers (R and U5) specific for early products were 5′-GGCTAACTAGGGAACCCACTGCTT (forward) and 5′-CTGCTAGAGATTTTCCACACTGAC (reverse). The late-product primer sequences (R and 5NC) were 5′-TGTGTGCCCGTCTGTTGTGT (forward) and 5′-GAGTCCTGCGTCGAGAGAGC (reverse), as previously described. Statistical analyses were performed using Stata version 9.0 (http://www.stata.com). T cell subset and clinical data are presented as means (standard deviation). Statistical significance between groups was determined by Wilcoxon rank sum test. Differences were considered significant at p < 0.05.

10.1371/journal.ppat.1007876

Structural mechanism for guanylate-binding proteins (GBPs) targeting by the Shigella E3 ligase IpaH9.8

The guanylate-binding proteins (GBPs) belong to the dynamin superfamily of GTPases and function in cell-autonomous defense against intracellular pathogens. IpaH9.8, an E3 ligase from the pathogenic bacterium Shigella flexneri, ubiquitinates a subset of GBPs and leads to their proteasomal degradation. Here we report the structure of a C-terminally truncated GBP1 in complex with the IpaH9.8 Leucine-rich repeat (LRR) domain. IpaH9.8LRR engages the GTPase domain of GBP1, and differences in the Switch II and α3 helix regions render some GBPs such as GBP3 and GBP7 resistant to IpaH9.8. Comparisons with other IpaH structures uncover interaction hot spots in their LRR domains. The C-terminal region of GBP1 undergoes a large rotation compared to previously determined structures. We further show that the C-terminal farnesylation modification also plays a role in regulating GBP1 conformation. Our results suggest a general mechanism by which the IpaH proteins target their cellular substrates and shed light on the structural dynamics of the GBPs.

Shigella flexneri is a Gram-negative bacteria that causes diarrhea in humans and leads to a million deaths every year. Once inside the cell, S. flexneri injects the host cell cytoplasm with effector proteins to suppress host defense. The guanylate-binding proteins (GBPs) have potent antimicrobial functions against a number of pathogens including S. flexneri. For successful infection, S. flexneri relies on an effector protein known as IpaH9.8, a unique ubiquitin E3 ligase to target a subset of GBPs for proteasomal degradation. How these GBPs are specifically recognized by IpaH9.8 was unclear. Here, using a combination of structural and biochemical approaches, we reveal the molecular basis of GBP-IpaH9.8 interaction, and show that subtle differences in the seven human GBPs can significantly impact the targeting specificity of IpaH9.8. We also show that the GBPs have considerable structural flexibility, which is likely important for their function. Our results provide further insights into S. flexneri pathogenesis, and laid the groundwork for future biophysical and biochemical studies to investigate the functional mechanism of GBPs.

The guanylate-binding proteins (GBPs) play critical roles in cell-autonomous immunity against a diverse range of bacterial, viral, and protozoan pathogens. The charter member of this family is GBP1, which was identified as a protein that is strongly induced by the interferons and can specifically bind to the guanylate affinity column [1, 2]. There are seven GBPs in human (GBP1-7), which share 52%-88% sequence identity between each other [3]. GBP1, GBP2, and GBP5 contain C-terminal CaaX box sequences that allow them to be prenylated in cells. GBP1 is farnesylated, which is important for its localization to membrane structures such as the Golgi apparatus [4, 5]. The farnesylation modification, together with a nearby triple-arginine motif, is also required for the localization of GBP1 to cytosolic bacteria [6, 7]. Once on the bacterial surface, GBP1 is able to recruit other GBPs via heterodimerization and oligomerization [7, 8]. A unique property of GBP1 is its ability to hydrolyze GTP first to GDP and then to GMP in a processive manner [9, 10]. In contrast, GBP2 only converts ~10% GTP to GMP, whereas GBP5 hydrolyzes GTP only to GDP [11, 12]. The physiological significance of the unusual enzyme activity of GBP1, as well as the biochemical differences between different GBPs, remains unclear. Mechanistically, the GBPs belong to the dynamin superfamily of GTPases, which often mediate membrane fission or fusion [13, 14]. By analogy, the GBPs could also function in the membrane remodeling processes. For example, they may contribute to the lysis of pathogen-containing vacuoles. Other reported functions of the GBPs include promoting autophagy, initiating inflammasome assembly, and inhibiting bacterial motility (for recent reviews, see [15–20]). However, our understanding towards the functions of these important proteins is still in its infancy. The GBPs have complex structural dynamics. Crystal structures have been determined for the full-length GBP1 in the monomer state and the isolated GTPase domain of GBP1 in the dimer state [10, 21, 22]. GBP1 contains an N-terminal large GTPase (LG) domain and a C-terminal helical region, which can be further divided into a middle domain (MD) that contains the α7-α11 helices and a GTPase effector domain (GED) that consists of the α12-α13 helices. The GED folds back and interacts with LG and MD, which is important to maintain GBP1 at the resting state [21, 23]. Binding of GTP induces the release of GED from the rest of the protein, resulting in an extended conformation that was previously interpreted as a “dimer” based on size-exclusion chromatography analyses [24]. Unlike the isolated LG domain that readily dimerizes under several guanine nucleotide conditions, full-length GBP1 only forms a stable dimer in the presence of GDP-AlFx that mimics the catalytic transition state [10, 24]. Due to the extended conformation of the GED domain, the dimer of the full-length protein has a large hydrodynamic radius and was long regarded as a “tetramer”. Dimerized full-length GBP1 can cause the tethering of unilamellar vesicles in vitro, and this activity depends on the C-terminal farnesylation modification [25]. Furthermore, the farnesylated GBP1 can form a transient ring-like oligomer that is reminiscent of dynamin and related proteins such as the Mx (Myxovirus resistance) proteins [25]. Whether these properties are related to the cellular functions of GBP1 remains to be investigated. Pathogens often antagonize key cellular proteins to evade host defense. Due to the important functions of the GBPs in innate immunity, it is not a surprise that some pathogens have evolved strategies to counter their activity. The IpaH family of proteins are unique E3 ubiquitin ligases that are only found in bacteria, especially pathogenic bacteria such as Shigella and Salmonella [26]. They all contain an N-terminal Leucine-rich repeat (LRR) domain and a C-terminal novel E3 ligase (NEL) domain. Although the NEL domain is structurally unrelated to the HECT family of E3 ligases, it also catalyzes the ubiquitination reaction by forming a ubiquitin-thioester intermediate via an invariant Cys in the CxD motif [27, 28]. IpaH9.8 from Shigella flexneri, an intracellular bacterium that causes bacillary dysentery, is one of the most extensively studied member of the IpaH family. In fact, it is one of the first IpaH proteins that is demonstrated to be an E3 ligase [26]. Recent studies have discovered that IpaH9.8 ubiquitinates and degrades a subset of GBPs, which is important for S. flexneri to suppress host defense and replicate in the cells [6–8]. To delineate how the GBPs are targeted by IpaH9.8 and gain further insights into GBP-mediated immunity, we have first determined the crystal structure of IpaH9.8LRR in complex with GBP1LG-MD, which explains the specific recognition of select GBPs by IpaH9.8. Mutating the GBP1-binding residues in IpaH9.8 diminish its ability to degrade the GBPs. By comparing with other IpaH protein structures, we have identified interaction hot spots in the LRR domains of this unique family of bacterial ubiquitin ligases. A large rotation of GBP1MD is observed in our structure, revealing that the elastic α7 helix plays an important role in regulating the structural dynamics of GBP1. Finally, we determined the structure of farnesylated full-length GBP1 and show that the farnesylation modification is involved in restraining GBP1 conformation. The IpaH proteins are modular enzymes that all contain a LRR domain and a NEL domain. The NEL domains are highly conserved, and therefore the substrate specificity is largely dictated by the variable LRR domains. Indeed, IpaH9.8LRR binds to GBP1 [6]. Swapping the LRR domains of IpaH4 and IpaH7.8 to IpaH9.8LRR enables the chimera IpaHs to degrade the GBPs (Fig 1a). To elucidate the molecular basis of how IpaH9.8LRR recognizes GBP1, we sought to determine their complex structure. We first crystallized full-length GBP1 in complex with IpaH9.8LRR. However, the crystal diffracted to only ~10 Å and could not be improved despite extensive attempts. We subsequently crystallized the LG-MD region of GBP1 (GBP1LG-MD) in complex with IpaH9.8LRR and determined the structure at 3.6 Å (Table 1, Fig 1b). The moderate resolution is likely caused by a high solvent content of the crystal (73.4%). Nevertheless, the electron density map generated from the molecular replacement solution is of high quality and allows unambiguous model building (S1 Fig). The LG domain of GBP1 features a canonical globular GTPase fold that highly resembles GBP1LG in the full-length GBP1 structure [21, 22]. Superimposing it to the full-length structure generates a root mean square deviation (rmsd) of 1.0 Å for 257 Cα atoms. The MD domain features two three-helix bundles that spiral around the common α9 helix and also resembles the corresponding region in the full-length structure. Superimposing the MD domain from our structure to the corresponding region in full-length GBP1 yields a rmsd of 1.9 Å for 169 Cα atoms. However, the arrangement of the LG and MD in our structure is different from that in the full-length structure, and a large swing of the MD is observed (S2a Fig). IpaH9.8LRR is very similar to the previously determined IpaH9.8LRR alone structure [29], and contains eight LRR motifs (LRR1-LRR8) that are organized into a slightly curved solenoid. In the complex structure, it engages GBP1LG using the concave surface of the solenoid (Fig 1b). Three regions in GBP1LG are involved in interacting with IpaH9.8LRR: the P-loop, the switch II region, and the α3 helix (Fig 1b). These regions are located on the opposite side of the GED domain in the full-length GBP1 structure, so the GED domain, which is not present in our structure, would not interfere with the binding (S2a Fig). On the other hand, these regions are involved in forming the dimer interface in the LG dimer structure [10], and therefore binding of IpaH9.8 would lead to the disruption of the GBP1LG dimer (S2b Fig). This is consistent with our previous observation that IpaH9.8 disrupts the GBP1 “tetramer” in the presence of GDP-AlFx [6]. In the structure, seven out of the eight LRR modules in IpaH9.8LRR contribute residues to interact with GBP1 (S1b Fig, Fig 2). In LRR1, Arg629.8 (superscripts 9.8 and G indicate residues in IpaH9.8 and GBP1, respectively) forms bidentate interactions with Glu105G in the Switch II region of GBP1. Asp649.8 forms a hydrogen bond with Tyr47G, and Arg659.8 interacts with Tyr47G via a cation-π interaction. Asn679.8 forms a hydrogen bond with Gln137G. In LRR2, Asn839.8 forms a hydrogen bond with Glu105G.Tyr869.8 forms a hydrogen bond with the main chain carbonyl group of Gly102G, at the same time forms van der waals interactions with Tyr47G. Gln889.8 appears to stabilize the position of Lys1089.8 in LRR3, which in turn forms a salt bridge with Asp140G. Other residues in LRR3 that interact with GBP1 include Tyr1039.8, which packs against the aliphatic region of Glu105G. His1269.8 from LRR4 interacts with Tyr143G via cation-π and van der waals interactions. In LRR5, Asn1439.8 forms a hydrogen bond with Asn109G, and Tyr1469.8 hydrogen bonds with Glu147G. Arg1669.8 from LRR6 forms a salt bridge with Glu147G. Arg1909.8 from LRR7 may form a hydrogen bond with His150G. The residues involved in binding GBP1 are unique to IpaH9.8 (S3 Fig), explaining the fact that only IpaH9.8, but not other IpaH proteins, specifically degrades the GBPs [6, 8]. The seven human GBPs are highly homologous to each other. However, only a subset of GBPs such as GBP1, GBP2, GBP4, and GBP6 are efficiently targeted and degraded by IpaH9.8 [6, 8]. GBP3 and GBP7 are particularly resistant (Fig 3a). Careful examination reveals subtle differences in their Switch II and α3 helix regions. For example, GBP3 contains a Lys (Lys105) in its Switch II that aligns with Glu105 in GBP1 (S4 Fig), which lies at the center of GBP1LG-MD/IpaH9.8LRR interface and makes critical interactions with several IpaH9.8 residues (Fig 2). Mutation of this residue to Glu allows the GBP3 mutant (GBP3-M) to be efficiently degraded by IpaH9.8 (Fig 3a). GBP3-M also binds strongly to IpaH9.8-C337A, an IpaH9.8 mutant that has abolished E3 ligase activity (Fig 3b). The α3 helix of GBP5 is slightly different when compared with GBP1 (S4 Fig). Gly137, Leu141, and His143 replace GBP1 residues Gln137, Gln141, and Tyr143, respectively. These differences likely reduce the interaction between GBP5 and IpaH9.8, and make GBP5 a suboptimal substrate that requires higher amounts of IpaH9.8 for degradation (Fig 3a). A double mutant of GBP5, G137Q/L141Q (GBP5-M), is degraded more efficiently by IpaH9.8 (Fig 3a). Several residues in the Switch II and α3 helix region of GBP7 are different compared to GBP1, including Met104 that replaces Val104 in GBP1 and His143 like in GBP5 (S4 Fig). The bulkier Met104 may hinder the binding of IpaH9.8. Furthermore, molecular dynamics simulation study suggests that the α3 helix region of GBP7 prefers to adopt a loop rather than a helical conformation (S5 Fig), caused partly by the presence of Ser111, instead of an Asn in other GBPs, at the end of its Switch II (S4 Fig). Ser111 appears to stabilize a hydrogen bond interaction between Ser113 and Glu147, which causes the α3 helix to unfold. Swapping the GBP7 Switch II-α3 region (residues 104–151) to the corresponding segment in GBP1 renders the GBP7 mutant (GBP7-M) susceptible to IpaH9.8-mediated degradation (Fig 3a). GBP7-M also shows a stronger interaction with IpaH9.8-C337A (Fig 3b). To further verify our structure, we mutated several IpaH residues that are involved in binding to GBP1, including Tyr86, Gln88, His126, Tyr146, and Arg190. When these mutations are generated in combination with C337A, the resulting mutants IpaH9.8-Y86A/Q88A/C337A, IpaH9.8-H126A/R190A/C337A, and IpaH9.8-Y146A/R190A/C337A all display greatly reduced interaction with GBP1, as shown by the co-immunoprecipitation experiments (Fig 4a). Mutating Tyr86 and Gln88 together generates the strongest effect. Similarly, IpaH9.8-Y86A/Q88A/C337A also failed to interact with other GBPs, including GBP2, GBP4, and GBP6 (Fig 4a). To validate the physiological relevance of these GBP-binding residues, we performed cell imaging experiments as we previously described [6]. We made mutations to IpaH9.8 that are fused with 10 tandem repeats of the SUperNova tags (SunTags) [30]. We then expressed these IpaH9.8 mutants in the S. flexneri ΔipaH9.8 strain and used these bacteria to infect HeLa cells stably expressing RFP-GBP1 and scFv-GCN4-GFP. GCN4 is a single chain antibody that specifically recognizes the SunTag. In uninfected cells, GCN4-GFP display a dispersed pattern in the cell (Fig 4b). When infected with S. flexneri expressing wild-type IpaH9.8-10xSunTag, the GFP signals are enriched in the cytoplasm due to the delivery of IpaH9.8 by the bacteria, and the RFP signal is largely diminished due to the degradation of GBP1 (Fig 4b). In contrast, RFP-GBP1 is not efficiently degraded by the bacterial strains expressing IpaH9.8-Y86A/Q88A, IpaH9.8-H126A/Y146A, or IpaH9.8-Y146A/R190A. In these cells, the RFP signal is most bright around the bacteria, due to the localization of GBP1 to the bacterial surface (Fig 4b). Together, these results demonstrate that an intact GBP-binding surface in IpaH9.8LRR is critical for the function of IpaH9.8 in vivo. The IpaH proteins have diverse substrates in the host [31]. In particular, two IpaH proteins from Salmonella, SspH1 and Slrp, use their LRR domains to target the host PKN1 kinase and Trx1 thioredoxin, respectively [32, 33]. Crystal structures have been determined for SspH1LRR in complex with a coiled-coil region of the PKN1 kinase [34], and Slrp in complex with Trx1 [35]. Comparing these structures with the GBP1LG-MD/IpaH9.8LRR complex reveals both differences and common features (Fig 5). Like IpaH9.8, SspH1 binds its target using the concave surface of its LRR domain. While the N-terminal region of IpaH9.8LRR mediates the majority of the interactions with GBP1, the contact site for PKN1 is more focused on the C-terminal half of SspH1LRR (Fig 5a and 5b). Nonetheless, the edge of the concave surface that are pointed by the LRR strands are important for the binding in both structures. In IpaH9.8LRR, Asn67 from LRR1, Gln88 from LRR2, Lys108 from LRR3, His126 from LRR4, Tyr146 from LRR5, Arg166 from LRR6, and Arg190 from LRR7 form a continuous surface patch that are critical for GBP1 binding (Fig 5a). In SspH1LRR, a similar edge is formed by Leu247 from LRR3, Asn266 from LRR4, Asn286 from LRR5, Asn326 from LRR7, His346 from LRR8, Asp368 from LRR9, and His392 from LRR10 (Fig 5b). When SspH1LRR is compared with IpaH9.8LRR, SspH1 residues Leu247, Asn266, Asn286, and Asn326 align exactly with IpaH9.8 residues Lys108, His126, Tyr146, and Arg190, respectively (S3 Fig). In the crystal structure of Slrp/Trx1, Slrp interacts with Trx1 using two interfaces [35]. The so-called type I binding site highly resembles the GBP1 binding site in IpaH9.8LRR (Fig 5a and 5c). This site is formed by the first six LRR modules of SlrpLRR, and also involves the concave surface. Trx1 binding residues Arg184, Lys186, Ile187, Ile205, Asn208, Tyr226, Gln231, Ile250, and His271 all align with IpaH9.8 residues Arg62, Asp64, Arg65, Asn83, Tyr86, Tyr103, Lys108, His126, and Tyr146 (S3 Fig). Although the physiological significance of the type I binding site in Slrp remains to be explored, these analyses suggest that the IpaH family proteins could generally bind their target proteins using the LRR concave surfaces. In particular, residues located at positions corresponding to Lys108 in IpaH9.8-LRR3, His126 in IpaH9.8-LRR4, and Tyr146 in IpaH9.8-LRR5 are important for binding in all three complexes (Fig 5, S3 Fig), suggesting that these three positions could function as “hot spots” to mediate the interaction between the IpaH proteins and their cellular targets. The dynamin superfamily proteins are considered mechanochemical enzymes that convert the energy from GTP binding and hydrolysis to mechanical force. The conformational dynamics of GBP1 is likely at the heart of its function but remains poorly understood. In the previously determined structures, the GED folds back and locks the conformation of GBP1 (Fig 6a). However, biophysical studies suggest that the GED domain is unleashed during the GTPase reaction cycle and the C-terminal region of GBP1 undergoes large degree of conformational change. In our structure, since the GED domain is not present, the MD domain is free to adopt a relaxed conformation. The α7 helix, which is forced to bend in the apo structure due to the interaction between the GED and the LG-MD domains, springs back to the straightened state (Fig 6b). Starting from a highly conserved Gln321 (S4 Fig), the C-terminal half of the α7 helix rotates ~13°, and this conformational change is transmitted toward the rest of the protein, causing an ~20° en bloc rotation of the α8-α11 helices (Fig 6c, S2a Fig). Due to the unfavorable geometry of the α7 helix in the “GED on” state, this conformational change likely also occurs in the full-length protein when the GED domain is set free during GBP1 function. The conformational change seen above prompted us to further investigate the conformation dynamics of GBP1. GBP1 is farnesylated at Cys589, and this modification is important for its localization to the Golgi apparatus and recruitment by various pathogens [4–7]. Despite this modification, GBP1 is primarily a cytosolic protein until the cells are infected by pathogens [4, 5], suggesting that the farnesyl group is probably not exposed at the resting state. The farnesylation modification changes the behavior of GBP1 on hydrophobic chromatography column and reduced its ability to hydrolyze GTP to GMP, suggesting that it impacts the conformation of GBP1 [36]. To assess how the farnesyl group affects GBP1 structure, we followed a previously described protocol [36] and prepared farnesylated GBP1 (GBP1F) by co-expressing GBP1 with the farnesyltransferase in E. coli. Successful modification is confirmed by mass spectrometry analyses of the purified protein (S6a Fig). We subsequently determined the crystal structure of GBP1F (Table 1). Interpretable electron density is present for the farnesyl group, as well as the entire C-terminal tail of GBP1 (S6b Fig). The farnesyl group is accommodated in a pocket formed by His378, Gln381, Lys382, Ala385 from the α9 helix and Tyr524, His527, Leu528, Leu531 from the α12 helix (Fig 7a). These interactions pull the α12 helix towards the α9 helix, and cause the GED domain to become more tightly fastened to the rest of the protein. In this conformation, the α7 helix remains bent; while the N-terminal half the α12 helix, as well as the majority of the MD domain, undergoes a ~10° rotation when compared to the previously determined full-length GBP1 structure (Fig 7b). Despite the fact that GBP1 was identified more than 30 years ago as one of the most prominent proteins that are induced by the interferons, its precise function remains elusive. Recent studies suggest that GBP1 inhibits intracellular bacterial replication by translocating to the bacterial surface, hindering their actin-dependent motility, and blocking their cell-to-cell spread [6–8]. Clearly, GBP1 plays an important role in cell-autonomous immunity, and poses a major threat to cytosolic bacteria such as S. flexneri. In the arms race between the bacteria and the host, S. flexneri has acquired the ability to eliminate a subgroup of GBPs through the action of its virulence E3 ligase IpaH9.8. To provide insight into the interaction between IpaH9.8 and the GBPs, we have solved the crystal structure of the LRR domain of IpaH9.8 in complex with a major fragment of GBP1. Our results show that the residues involved in interacting with GBP1 are unique to IpaH9.8, elucidating how IpaH9.8, but not other IpaH family proteins, can specifically target the GBPs. Due to the differences in the Switch II and α3 helix regions, GBP3, GBP5, and GBP7 are not efficiently degraded by IpaH9.8. Mutating relevant residues in these GBPs makes the mutant proteins more susceptible to IapH9.8-mediated degradation. By comparing our structure with other IpaH proteins in complex with their target molecules, we further reveal interaction hot spots in the LRR domain of this unique family of bacterial effectors. These results provide a deeper understanding on the pathogenesis of S. flexneri, and may facilitate the investigation of other IpaH proteins in the future. Our results also shed light on the structural dynamics of GBP1. Previously, GBP1 without the farnesyl moiety has been crystallized in the apo state and in complex with GMP-PNP, a nonhydrolyzable analog of GTP [21, 22]. However, the two structures are largely similar and have not provided sufficient insights into the conformational change of GBP1. Through the examination of the GBP1LG-MD and GBPF structures determined in this study, we uncovered two new conformations of GBP1. In a way, the GBP1F structure likely reflects GBP1 at its most tense state. By creating additional interactions between the GED domain and the MD domain, the farnesyl group appears to function as the second tier of bolt to lock the GED domain to the rest of the protein. A bending of the α7 helix is forced in this conformation. In contrast, the GBP1LG-MD structure likely reflects GBP1 at its most relaxed state. We envision that when the structural restraints imposed by the GED domain and the farnesyl group are relieved upon GBP1 activation, the α7 helix would become straight, and this would cause the C-terminal region to rotate like seen here in the GBP1LG-MD structure. How the GED domain and the farnesyl moiety are arranged in the active state, and how these conformational changes are translated to the function of GBP1, remain important questions to be addressed. In this regard, it is worth noting that, GBP5ta, a splicing variant of GBP5 that is associated with the T-cell lymphoma tissues, naturally lacks the GED domain [37]. GBP3ΔC, a splicing variant of GBP3 that does not have the α13 helix, has also been reported [38]. The functional significance of these GBP variants are unclear, but they would be more prone to adopt a relaxed conformation compared to full-length GBP5 and GBP3. Primers used in this study are listed in Supplementary Table 1. IpaH9.8LRR (residues 22–252) [6] and GBP1LG-MD (residues 1–479) were expressed as His6-SUMO fusion proteins in E. coli BL21(DE3). The bacterial cultures were grown at 37 °C in the Luria-Bertani (LB) medium to an OD 600 of 0.6–0.8 before induced with 0.5 mM isopropyl β-D-1-thiogalactopyranoside (IPTG) at 18 °C for overnight. The cells were collected by centrifugation and were resuspended in a lysis buffer containing 50 mM Tris-HCl, pH 8.0, 500 mM NaCl, 10 mM imidazole, 5 mM β-mercaptoethanol, and 1 mM phenylmethylsulfonyl fluoride (PMSF). The cells were then disrupted by sonication, and the insoluble debris was removed by centrifugation. The supernatant was applied to a Ni-NTA column (GE Healthcare). The column was then washed extensively with a wash buffer containing 50 mM Tris-HCl, pH 8.0, 500 mM NaCl, 30 mM imidazole, and 5 mM β-mercaptoethanol, and eluted with an elution buffer containing 50 mM Tris-HCl, pH 8.0, 150 mM NaCl, 250 mM imidazole, and 5 mM β-mercaptoethanol. Next, the eluted proteins were digested with the ULP1 protease to cleave the N-terminal His6-SUMO fusion tag. The protein samples were then passed through another Ni-NTA column to remove the His6-SUMO fusion tag and the ULP1 protease. Untagged IpaH9.8LRR and GBP1LG-MD were further purified by gel filtration chromatography using a Superdex 200 column (GE Healthcare), and eluted in the final buffer containing 25 mM Tris-HCl, pH 8.0, 20 mM NaCl, and 2 mM Dithiothreitol (DTT). To obtain the farnesylated GBP1 (GBP1F), full-length GBP1 was cloned into a vector that is kanamycin resistant and expresses GBP1 as a His6-SUMO fusion protein. The two subunits of the farnesyltransferase (FTase α and β, respectively) were cloned into the pACYCDuet-1 (Novagen) vector that is chloramphenicol resistant. His6-SUMO-GBP1 was then co-expressed with the FTase α/β in E. coli BL21(DE3). The bacterial cultures were supplemented with both kanamycin (50 μg/ml) and chloramphenicol (25 μg/ml), and were induced with 0.5 mM IPTG at an OD 600 of 0.8. The cells were then cultured at 20°C for 18h and were collected by centrifugation. The GBP1F was then purified similarly as described above for the GBP1LG-MD protein. To obtain the IpaH9.8LRR/GBP1LG-MD complex, purified IpaH9.8LRR and GBP1LG-MD were incubated overnight on ice using a 1.5:1 molar ratio. The mixtures were then passed through a Superdex 200 column and eluted using the final buffer described above. The protein complex was concentrated to 18 mg/ml for crystallization. Crystals were grown at 20°C using the sitting drop vapor diffusion method. The crystallization solution contains 1.6 M sodium/potassium phosphate, pH 6.5. Crystals grew to full size in several days and were transferred to a cryo solution containing 1.6 M sodium/potassium phosphate, pH 6.5, and 38% sucrose before flash-cooled in liquid nitrogen. GBP1F was crystallized using the sitting drop vapor diffusion method at a concentration of 15 mg/ml. Crystals appeared overnight in 20 mM citric acid, 80 mM Bis-Tris propane, pH 8.8, and 16% (w/v) Polyethylene glycol 3,350. For data collection, the crystals were transferred to a solution containing 20 mM citric acid, 80 mM Bis-Tris propane, pH 8.8, 16% Polyethylene glycol 3,350, and 20% ethylene glycol before flash-cooled in liquid nitrogen. The diffraction data were collected at Shanghai Synchrotron Radiation Facility (SSRF) beamline BL17U. The diffraction data were indexed, integrated, and scaled using HKL2000 (HKL Research). The structure was determined by the molecular replacement method using the published structure of IpaH9.8LRR (PDB ID:5B0N) and GBP1 (PDB ID:1DG3) as the search models. The structure modeling was performed in Coot [39] and refined using Phenix [40]. Structural validation was performed with MolProbity [41]. Composite omit map was generated with Phenix [42]. The structure models of GBP6 and GBP7 were obtained by homology modeling using MODELLER [43] with GBP1 structure as the template. The molecular dynamics simulations were carried out using the GROMACS 5.1.2 package (http://www.gromacs.org) [44]. HEK293T and HeLa cells, originally obtained from ATCC, were grown in a humidified incubator with 5% CO2 at 37 °C in Dulbecco's modified Eagle's medium (DMEM) supplemented with 10% fetal bovine serum (FBS) and 100 μg/ml penicillin/streptomycin (GIBCO). All cell lines were tested to be free of mycoplasma by the standard PCR method. The mammalian expression plasmids have been previously described [6]. Mutations were introduced into plasmids by a PCR-based method. For the immunoprecipitation experiments, a catalytically dead mutant of IpaH9.8 (IpaH9.8-C337A) was used, since wild-type IpaH9.8 would lead to quick degradation of co-expressed GBPs. HEK293T cells were grown in 10 cm dishes to 70%-80% confluency. They were then co-transfected with 5 μg IpaH9.8-C337A and 10 μg indicated GBP plasmids using Polyethylenimine (PEI). The cells were harvested 18–24 hours later, washed with the phosphate-buffered saline (PBS) buffer, and lysed in a buffer containing 25 mM Tris-HCl, pH 8.0, 2 mM MgCl2, 1 mM GTP, 1 mM PMSF, and 0.5% Triton X-100. The cell lysates were cleared by centrifugation, and then incubated with the Flag M2 beads (Sigma, A2220) for 2 hours. The beads were spun down and then washed three times with the wash buffer (25 mM Tris-HCl, pH 8.0, 2 mM MgCl2, 1 mM GTP, and 0.2% Triton X-100). The immunoprecipitated proteins were eluted from the beads using the 3x Flag peptides (NJPeptide, NJP50002) and analyzed by SDS-PAGE and western blotting. Purified GBP1 protein interacts strongly with purified IpaH9.8 under all nucleotide conditions (apo, GMP, GDP, GppNHp, and GDP-AlFx) [6]. Also, no nucleotide is required for the formation of the IpaH9.8LRR/GBP1LG-MD complex. However, we observed more consistent binding between GBP1 and IpaH9.8 co-expressed in cells when we included GTP in the lysis buffer. The reason for this is not entirely clear. We noticed that GBP1 tends to form puncta/aggregates when overexpressed in HEK293T cells, and we hypothesized that GTP may help to solubilize these aggregates. For the degradation experiments, HEK293T cells were grown to 70%-80% confluency in 6-well plates, and were transfected with indicated plasmids using PEI. 18–24 h after transfection, the cells were harvested, washed, and then lysed in a lysis buffer containing 25 mM Tris-HCl, pH 8.0, and 0.5% Triton X-100. The cell lysates were cleared by centrifugation and then analyzed by western blot using antibodies for HA (Sigma, H3663), c-Myc (HuaxingBio, HX1802), Flag (Sigma, F3165), and β-tubulin (TransGen, HC101). The IpaH9.8 gene with indicated mutations were cloned into the pME6032-10x SunTag plasmid as previously described [6]. S. flexneri ΔipaH9.8 2a strains were then transformed with these plasmids, and single colonies were picked up for each individual plasmid. The bacterial strains were cultured overnight at 37°C in the LB broth, before diluted 1:100 in fresh LB broth, and grown to an OD 600 of 0.8 in the presence of IPTG. The HeLa cell line stably expressing RFP-GBP1 and scFv-GCN4-GFP was described previously [6]. The cells were seeded onto glass coverslips in 24-well plates and cultured for 16 h before infection. The infection (MOI, 50) was facilitated by centrifugation at 800 g for 5 min at room temperature, and cultured for another hour at 37°C in a 5% CO2 incubator. Cells were washed three times with PBS. Fresh DMEM containing 100 μg/ml gentamycin was then added to kill the extracellular bacteria. Two hours later, infected cells were washed three times with PBS, fixed with 4% paraformaldehyde for 30 min at room temperature, and then place in the mounting medium (ZSGB-BIO, ZLI-9556) for imaging. Cell images were recorded using the Zeiss LSM 510 Meta confocal microscope and processed with the LSM software package.

10.1371/journal.pgen.1003185

Genetic Disruption of the Copulatory Plug in Mice Leads to Severely Reduced Fertility

Seminal fluid proteins affect fertility at multiple stages in reproduction. In many species, a male's ejaculate coagulates to form a copulatory plug. Although taxonomically widespread, the molecular details of plug formation remain poorly understood, limiting our ability to manipulate the structure and understand its role in reproduction. Here I show that male mice knockouts for transglutaminase IV (Tgm4) fail to form a copulatory plug, demonstrating that this gene is necessary for plug formation and lending a powerful new genetic tool to begin characterizing plug function. Tgm4 knockout males show normal sperm count, sperm motility, and reproductive morphology. However, very little of their ejaculate migrates into the female's reproductive tract, suggesting the plug prevents ejaculate leakage. Poor ejaculate migration leads to a reduction in the proportion of oocytes fertilized. However, Tgm4 knockout males fertilized between 3–11 oocytes, which should be adequate for a normal litter. Nevertheless, females mated to Tgm4 knockout males for approximately 14 days were significantly less likely to give birth to a litter compared to females mated to wild-type males. Therefore, it appears that the plug also affects post-fertilization events such as implantation and/or gestation. This study shows that a gene influencing the viscosity of seminal fluid has a major influence on male fertility.

Male reproductive fitness is strongly affected by seminal fluid. In many animals, the male's ejaculate coagulates in the female's reproductive tract to form a structure known as the copulatory plug. Here, I show that male mice without a functional copy of the gene transglutaminase IV cannot form a plug and suffer severe fertility defects. In spite of normal reproductive morphology, less of the ejaculate migrates through the female's reproductive tract and Tgm4 knockout males sire significantly fewer litters than wild type. This study demonstrates that the copulatory plug and/or Tgm4 itself is necessary for normal fertility.

The non-sperm component of an ejaculate can have large effects on male reproductive fitness. In internally fertilizing species, seminal proteins can modify female receptivity [1]–[3], egg laying behavior [4]–[6], implantation and gestation [7], and the female's immune response to sperm and embryo [7]–[11]. Seminal proteins can also interact with the ejaculates of competitor males to influence the outcomes of fertilization [12]–[14]. In many internally fertilizing taxa, ejaculated proteins coagulate to form a hardened copulatory plug in the vaginal-cervical region of the female [15]–[22]. In spite of its wide taxonomic distribution, the molecular details that underlie its formation remain poorly understood, which limits investigations into its function. After reviewing previous biochemical insights, I present a new genetic model that offers unprecedented power to being dissecting the function of the plug. Since the first published observation of a copulatory plug in a rodent nearly 165 years ago [19], several groups have attempted to characterize its molecular basis. Camus and Gley [23] showed that fluids extracted from the seminal vesicles coagulated in vitro upon contact with extract from the anterior lobe of the prostate (also referred to as the coagulating gland) [24], [25]. Building from the Camus & Gley experiment, Williams-Ashman and colleagues showed that the rate of coagulation depended on the concentrations of seminal vesicle and/or prostate protein extracts in vitro [26]. Because these early experiments were based on crude extracts, the general term “vesiculase” was coined to describe the unknown prostate-derived protein(s) responsible for inducing the coagulation of seminal vesicle proteins. More detailed biochemical investigations suggested the unknown vesiculase(s) was a transglutaminase [27], [28], a protein that crosslinks glutamines and lysines via γ-glutamyl-ε-lysine dipeptide bonds and causes the bound proteins to become insoluble and coagulate. A prostate-specific transglutaminase, transglutaminase IV (Tgm4), was later characterized from humans [29]–[31], and its protein is found in human ejaculates [32]. The ortholog in mouse is also ejaculated [33], and functionally analogous transglutaminases have been found in mosquito ejaculates [16]. In spite of these early advances, it remains unknown whether Tgm4 is necessary for the formation of the copulatory plug. It has been suggested that some seminal vesicle proteins self-coagulate in the absence of Tgm4 [34], that proteins other than Tgm4 induce the coagulation [35], and that female-derived proteins may be necessary for coagulation [36]. Furthermore, there is evidence that more than one transglutaminase exists in the male reproductive tract of rodents [33], [37], [38], though this could also be due to post-translational modifications [39]. Interestingly, human ejaculates do not coagulate strongly even though they have large amounts of Tgm4 [32], calling into question its role in seminal fluid coagulation. More fully characterizing the biochemical basis of seminal fluid coagulation is critical for understanding the function of the copulatory plug. Early attempts to study copulatory plug function necessarily relied on surgical removal of male accessory glands [40]–[47]. Although copulatory plugs were abnormal and male fertility compromised in some cases, inferences were limited by the invasiveness of the procedures, the confounding effects associated with the potential alteration of hundreds of ejaculated proteins, and the failure to fully prevent a copulatory plug-like structure from forming. Other experiments showed that manual removal of the plug soon after copulation did not prevent pregnancy or parturition [48], [49]. However, the copulatory plug may have affected fertility prior to experimental removal. To address these early experimental limitations requires a method to fully prevent the formation of a copulatory plug with minimal invasiveness. Here, I use Tgm4 knockout mice to better understand the molecular basis and functional importance of the copulatory plug, and report two main findings. First, Tgm4 knockout males failed to produce a copulatory plug after mating, demonstrating for the first time that this gene is necessary for the coagulation of seminal fluid in mice. Tgm4 knockout males therefore provide a powerful model to investigate the function of the copulatory plug. Second, in spite of normal sperm count, sperm motility, and reproductive morphology, Tgm4 knockout males sired significantly fewer litters than their wild type brothers. Analyses presented below suggest Tgm4 knockout males suffer fertility defects at two important stages: 1) less of their ejaculate migrates into the female's reproductive tract, and 2) females mated to Tgm4 knockout males produce significantly fewer litters even though a “normal” absolute number of oocytes were fertilized, suggesting additional defects in implantation and/or gestation. This study demonstrates that a gene influencing the viscosity of semen has major affects on male reproductive success. Heterozygous “knockout first” mice were acquired from the Knockout Mouse Project (see [50],[51], and Materials and Methods). Heterozygotes were crossed in the laboratory to generate homozygous and heterozygous knockout males, as well as homozygous wild type males that were used as controls in all experiments. All females used throughout the manuscript were homozygous wild type. All mice were essentially genetically identical except for the ∼7 kb “knockout first” cassette that spans exons 2–3 of Tgm4. Tgm4 knockout males (homozygous for the “knockout first” allele) did not form a copulatory plug (Table 1), demonstrating for the first time that this gene is necessary for seminal fluid coagulation. From 13 successful 3-hour pairings to Tgm4 knockout males (“success” being defined as the presence of sperm somewhere in the female's reproductive tract after three hours of pairing), complete dissection of each female's reproductive tract failed to yield a copulatory plug or plug-like structure (Table 1). In contrast, 14 of 16 successful 3-hour pairings to wild type males resulted in a copulatory plug, which normally occupies most of the vaginal canal and extends into the cervix, appearing “glued” to the epithelium. Herein, “wild type” includes males that were either heterozygous or homozygous for the wild type allele, as they were phenotypically indistinguishable from each other. I obtained similar results from 20-hour long male-female pairings: 0 of 8 females successfully paired with Tgm4 knockout males, and 11 of 15 paired to wild type males, yielded a plug. Because they cannot form a plug, Tgm4 knockout males represent a powerful genetic tool to investigate its role in reproduction. In the absence of a plug, the ejaculates of Tgm4 knockout males did not traverse the female reproductive tract properly. After mating to wild type males, female uterine horns appeared swollen, full of sperm and seminal fluid (Figure 1A). In contrast, after mating to Tgm4 knockout males, female uterine horns did not swell and sperm were difficult to locate upon dissection (Figure 1B). The difference in uterine horn width was statistically significant between females mated to wild type (N = 6) vs. Tgm4 knockout males (N = 15) (wild type: 2.64 mm, SD = 0.30; Tgm4 knockout: 2.14 mm, SD = 0.43; t = 2.98, df = 19, P = 0.01). The defect in ejaculate migration cannot be explained by defects in reproductive morphology of Tgm4 knockout males. Sperm count was not statistically different between wild type (N = 19) vs. Tgm4 knockout (N = 10) males (mean = 133,900 sperm/µl, SD = 55,000 vs. mean = 106,000 sperm/µl, SD = 43,000, respectively: t = 1.39, df = 27, P = 0.18), nor was sperm motility (mean = 0.96 sperm/sec, SD = 0.28 vs. mean = 0.87 sperm/sec, SD = 0.27, respectively: t = 0.78, df = 27, P = 0.44). From these same males, testis and seminal vesicle weight were analyzed in a full factorial ANCOVA to account for the potential covariation with body weight. There was not a significant difference in testis weight between Tgm4 knockout vs. wild type males (F1,25 = 0.02, P = 0.88), nor was there a genotype×body weight interaction effect on testis weight (F1, 25 = 0.95, P = 0.34). Similarly, there was no difference in seminal vesicle weight between genotypes (F1, 25<0.01, P = 0.97), nor was there a genotype×body weight interaction effect on seminal vesicle weight (F1, 25 = 0.03, P = 0.86). Furthermore, Tgm4 knockout males successfully copulated at a rate similar to wild type; from the 3-hour pairings, 16/26 to wild type, and 13/28 pairings to Tgm4 knockout males succeeded (Table 1, χ2 = 0.7, P = 0.4). Therefore, both genotypes display normal copulatory behavior. As might be expected from the reduced number of sperm that make it into the female's uterus, Tgm4 knockout males fertilized significantly fewer oocytes in 20-hour assays. Among successful 20-hour pairings, Tgm4 knockout males fertilized 45 of 122 oocytes dissected from the female's oviducts (36.9%), compared to wild type males, which fertilized 153 of 231 oocytes (66.2%) (Table 1). Fertilized oocytes from all successful pairings appeared healthy, with almost no signs of fragmentation. Oocytes originating from the same female are not independent observations, so I compared the proportion of fertilized oocytes on a per-female basis. Seven females successfully mated to Tgm4 knockout males yielded a mean 39.4% fertilized oocytes (range 21.1%–72.7%), significantly lower than 12 females successfully mated to wild type (mean = 67.9%, range 12.5%–93.3%) (t = 2.78, df = 17, P = 0.01). The number of females analyzed (7 mated to Tgm4 knockout and 12 mated to wild type) does not add up to the numbers in Table 1 (8 mated to Tgm4 knockout and 15 mated to wild type) because scorable oocytes were not always recovered from oviduct dissections. It should be noted that even though Tgm4 knockout males fertilized a lower proportion of oocytes compared to wild type, they always fertilized at least 3 oocytes (mean = 6.4, range 3–11), suggesting they should be able to impregnate females without difficulty. In contrast to this prediction, after being paired with females for 10–14 days, Tgm4 knockout males sired significantly fewer litters than wild type (Table 2). Of 30 pairings with Tgm4 knockout males, only 17 produced litters, significantly fewer than wild type males, which produced litters in 135 of 165 pairings (χ2 = 7.93, P = 0.005) (Table 2). Among all litters born to wild type fathers, 42/106 (39.6%) yielded 6 or fewer pups (the mean number of oocytes fertilized by Tgm4 knockout males, see previous paragraph), and 14/106 (13.2%) resulted in 3 or fewer pups (the minimum number of oocytes fertilized by Tgm4 knockout males, see previous paragraph). In other words, Tgm4 knockout males sired significantly fewer litters in spite of the fact that they appeared to fertilize enough oocytes for a healthy litter. Although Tgm4 knockout males sired significantly fewer litters (Table 2), there were no signs of maternal neglect, as judged by the likelihood a litter reached weaning age, the litter size, and the size of offspring at weaning. Specifically, 11 of 17 (65%) litters sired by Tgm4 knockout males reached weaning age (21–28 days old), compared to 106 of 135 (79%) litters born to a wild type male (Table 2; χ2 = 0.94, P = 0.3). Sometimes litters do not reach weaning age because of maternal neglect. Furthermore, the number of offspring weaned per litter was not significantly different among the two male genotypes (mean = 6.0 vs. 6.4 pups weaned per litter, SD = 3.5 vs. 2.4, range 1–12 for both, from N = 11 vs. 106 weaned litters sired by Tgm4 knockout or wild type males, respectively: Welch's t = 0.59, df = 10.97, P = 0.57), nor was weanling weight (mean weight = 11.61 g vs. 12.14 g, SD = 3.0 vs. 3.8 from N = 66 pups weighed from 11 litters vs. 77 pups weighed from 13 litters sired by Tgm4 knockout or wild type males, respectively: Welch's t = 0.93, df = 140.0, P = 0.35). The lack of statistical significance may be due to small sample sizes, but suggests that once litters are born, the pups have an equal chance of reaching healthy weaning age regardless of sire genotype. Tgm4 knockout males failed to produce a copulatory plug, demonstrating for the first time that this gene is necessary for plug formation. In spite of normal sperm count, motility and reproductive morphology, Tgm4 knockout males suffered reduced fertility, most importantly in the significant reduction of litters born compared to wild type. Taking all the data into consideration, a model of the copulatory plug acting at two important stages of reproduction seems to explain the fertility defects of Tgm4 knockout males. First, the plug may facilitate passage of the ejaculate through the cervix and into the uterine horns and oviducts (Fig. 1, Table 1), perhaps by sealing off the vagina and preventing backflow of the ejaculate [16], [41], [42], [52]–[54]. Second, the plug may enhance the embryos' ability to implant in the female's uterus, and/or reduce the chances of abortion after implantation (Table 2). For example, the plug may contribute to the physical stimulation necessary to shift the female's physiology towards “pseudopregnancy” [53], [55]–[57], a state where the uterus becomes primed for implantation in mice. This second aspect of the model is supported by the reduced number of litters born to Tgm4 knockout males in spite of the fact that they fertilized between 3–11 oocytes in 20-hour assays. There does not appear to be any fertility defects that arise from differential maternal investment post-parturition. Four observations suggest that the fertility defects observed in the current study arose from the absence of the copulatory plug rather than from additional pleiotropic functions of Tgm4. First, Tgm4 expression has so far only been detected in the prostate [58]–[60], and never in any other tissues of a male or a female [61], [62], thus it should only affect ejaculate composition. Second, the only annotated domains in the Tgm4 protein are related to the formation of γ-glutamyl-ε-lysine bridges in its target proteins (www.ensembl.org), suggesting that it has a limited biological role. Third, although transglutaminases may alter the sperm surface in vitro [63], [64], Tgm4 has never been detected on the sperm surface [65], [66], suggesting it does not directly affect the gamete. Fourth, Tgm4 has accumulated multiple loss-of-function mutations in some species that do not form a plug [67], which is not predicted if Tgm4 functions outside the context of plug formation. Although the present study demonstrates the importance of the copulatory plug in non-competitive matings, it does not reject the hypothesis that the copulatory plug evolved in response to sperm competition [20], which occurs when a female mates with more than one male during a single fertile period [68]. Copulatory plugs are larger and show stronger coagulation intensity in species with high levels of inferred sperm competition [21], [22], [69], and have been lost in some species that experience low levels of sperm competition [67], [70], [71]. Some copulatory plug proteins evolve rapidly in species with high levels of inferred sperm competition, which is predicted if the plug inhibits female remating [67], [70]–[73]. In mice, the copulatory plug forms immediately upon ejaculation and remains intact for approximately 24 hours [20 and unpublished data], which is longer than the 4–12 hours that a female is able to be fertilized during her estrus cycle. Males contribute protease inhibitors in their ejaculates, which may function to preserve their copulatory plugs from female degradation [33]. Interestingly, males missing one of these protease inhibitors make a plug that degrades more quickly than wild type, which is associated with fertility defects [74]. Although the above patterns suggest the plug inhibits female remating, over 20% of wild caught pregnant females carry a litter sired by more than one male [75], [76], suggesting the plug is an imperfect barrier, and females or competitor males sometimes remove the plug [77]–[81]. Interestingly, copulatory plugs do not always bias fertilizations towards the first male to mate in one-female-two-male mating experiments [82]–[85], and some evolutionary patterns do not fit the sperm competition hypothesis. For example, the socially and genetically monogamous rodent Peromyscus polionotus forms a plug [86], [87]. By showing that the copulatory plug is correlated with normal fertility in one-male-one-female matings, the current study offers an explanation for the evolutionary maintenance of the copulatory plug in the absence of intense sperm competition. For example, the copulatory plug may prevent loss of semen [52], promote transport of semen through the female's reproductive tract [16], [41], [42], [53], [54], contribute to the threshold stimulation females require for proper implantation and pregnancy [55], and/or serve as a reservoir for the slow release of sperm in the female reproductive tract [88]. In reality, the copulatory plug may have multiple functions and the genetic model presented here enables unprecedented power to begin dissecting these hypotheses. Many human seminal fluid proteins have orthologs in mouse ejaculates, including Tgm4 [33]. Even though human ejaculates do not form copulatory plugs, human seminal fluid enters a phase of coagulation and liquefaction [89], and defects in these transitions have been associated with subfertility [90]. There are 250 known nucleotide polymorphisms in human Tgm4 mRNA, including 120 missense mutations (www.ensembl.org version 69), and Tgm4 was not detected in all ejaculates of five humans [91]. Future studies may reveal genetic and proteomic variation in Tgm4 associated with differences in human male fertility. All mouse husbandry techniques, experimental methods, and personnel involved were approved by the University of Southern California's Institute for Animal Care and Use Committee, protocols #11394 and #11777. The Tgm4 knockout mouse model was constructed by the multi-institutional Knockout Mouse Project [50], [51]. A ∼7 kb “knockout first” cassette was inserted into the C57BL/6N (6N) genetic background (project #CSD30105). Alternative crossing to Cre and/or FLP mice allows for further genetic modification of the knockout allele, but was unnecessary in the present study. All experimental males used in this study had 6N parents that were heterozygous for the knockout (KO) and wild type (+) allele. When possible, all three genotypes were taken from the same litter to control for simple maternal effects. Sires and dams were paired for one to two weeks, then separated so the dam gave birth in isolation. Between 21–28 days after birth, males were weaned in groups until genotyping, at which point they were separated into their own cages to avoid dominance interactions between brothers. Sexually mature males show reduced fertility when grouped together, presumably as a result of dominance interactions [92]. Females were weaned in groups of up to three individuals. All three possible male 6N genotypes - but only homozygous wild type 6N females - were used in various experiments described below. Shortly after weaning, ear snips were taken for PCR-based genotyping. DNA isolated from ear snips was genotyped with four PCR reactions. Two PCR reactions specifically amplified the wild type allele: Reaction 1 primers (5′-AGGTGAAAAACCAAGAAATACCATC-3′ and 5′-CTATTCCAAAACCACCAGACAGTAC-3′) amplified a 704 bp fragment and Reaction 2 primers (GTGGACAGATATTCACTCTGAAGGT and GGAAACACCAATAGAAAAGTGAGTC) amplified a 1,170 bp fragment. Two PCR reactions specifically amplified the knockout first allele: Reaction 3 primers (GCTTTACATGTGTTTAGTCGAGGTT and GTTAAAGTTGTTCTGCTTCATCAGC) amplified a 1,244 bp fragment and Reaction 4 primers (GATTAAATATGATGAAAACGGCAAC and ATTATTTTTGACACCAGACCAACTG) amplified a 1,349 bp fragment. DNA was amplified using 35 cycles of denaturation (94 C, 20 seconds), annealing (58 C, 20 seconds) and extension (70 C and 40 seconds for Reaction 1, 70 C and 80 seconds for the other three reactions). All PCR reactions used Fermentas 2× PCR premix. Presence/absence of bands was scored on agarose gels. Only genotypes consistent across all four reactions were included in experiments. All experimental males were individually paired with homozygous wild type 6N females. Males were between 60 and 90 days old. For the 3-hour and 20-hour assays (see below), ∼28 day-old females were induced to ovulate using standard techniques [93], [94]. Briefly, females were administered 5U Pregnant Mare's Serum Gonadotropin (PMSG) followed 48 hours later by 5U Human Chorionic Gonadotropin (hCG). For the 10–14 day assays (see below), females between 2 and 10 months old were used and ovulation was not artificially induced. Between 2–6 months of age, a subset of experimental males were sacrificed, standard measurements taken, and testes and seminal vesicles dissected and weighed. Unless otherwise stated, Student's t-tests were used to compare phenotypes among groups. In all t-tests, assumptions of normality and equal variances were confirmed using Shapiro-Wilk tests and F-tests, respectively. In a few comparisons indicated above, the two groups being compared had significantly different variances; in these cases Welch's t-test [97] was used. Importantly, no conclusions changed if Student's t-tests, Welch's t-test, or non-parametric Mann-Whitney U tests were used in any comparisons. Testis and seminal vesicle weight were each analyzed in a full factorial ANCOVA using male genotype (knockout vs. wild type) and body weight as factors. An ANCOVA was employed to account for the potential covariation of testis or seminal vesicle weight with body weight. To test for differences in the number of litters born to Tgm4 knockout vs. wild type males, a 2×2 contingency table was tested against a χ2 distribution. All statistical analyses were performed in R (www.r-project.org) or customized Python scripts (www.python.org).

10.1371/journal.pbio.2006601

Heterogeneity and longevity of antibody memory to viruses and vaccines

Determining the duration of protective immunity requires quantifying the magnitude and rate of loss of antibodies to different virus and vaccine antigens. A key complication is heterogeneity in both the magnitude and decay rate of responses of different individuals to a given vaccine, as well as of a given individual to different vaccines. We analyzed longitudinal data on antibody titers in 45 individuals to characterize the extent of this heterogeneity and used models to determine how it affected the longevity of protective immunity to measles, rubella, vaccinia, tetanus, and diphtheria. Our analysis showed that the magnitude of responses in different individuals varied between 12- and 200-fold (95% coverage) depending on the antigen. Heterogeneity in the magnitude and decay rate contribute comparably to variation in the longevity of protective immunity between different individuals. We found that some individuals have, on average, slightly longer-lasting memory than others—on average, they have higher antibody levels with slower decay rates. We identified different patterns for the loss of protective levels of antibodies to different vaccine and virus antigens. Specifically, we found that for the first 25 to 50 years, virtually all individuals have protective antibody titers against diphtheria and tetanus, respectively, but about 10% of the population subsequently lose protective immunity per decade. In contrast, at the outset, not all individuals had protective titers against measles, rubella, and vaccinia. However, these antibody titers wane much more slowly, with a loss of protective immunity in only 1% to 3% of the population per decade. Our results highlight the importance of long-term longitudinal studies for estimating the duration of protective immunity and suggest both how vaccines might be improved and how boosting schedules might be reevaluated.

Immunological memory, mediated by antibodies, is a hallmark of immunity. A key problem for determining the longevity of protective immunity is heterogeneity in the responses of different individuals. We characterize the extent of this heterogeneity and determine how it affects the longevity of protection. We found that some individuals have higher antibody titers and these same individuals tend to have slower decay rates than others. We also found substantial heterogeneity in both the magnitude and decay rate of responses. Furthermore, differences in these two factors contribute comparably to the variation in antibody titers between different individuals over their lifetime. We then use statistical models to determine how variation in the magnitude and decay rate affect how protective immunity is lost at the population level to different virus and vaccine antigens. We identified different patterns for the loss of protective immunity elicited by protein immunization (tetanus and diphtheria) versus replicating viruses (measles, rubella, and vaccinia). While our results agree with the conventional view that antibodies elicited by protein immunization decay faster than those elicited by replicating viruses, we found that this is compensated for by the higher magnitude of responses (relative to the level for protection) for tetanus and diphtheria. Indeed, for the first 4 decades, a higher fraction of vaccinated individuals have protective immunity to tetanus and diphtheria than to measles, rubella, and vaccinia.

Immune memory is a cardinal feature of the adaptive immune response of vertebrates and is the principle that underlies vaccination [1–3]. Immunological memory arises as a consequence of the increase in the magnitude of the antigen-specific response, augmented by increases in the quality of the response [1,4]. A major problem in quantifying the duration of immune memory in humans is the long timescale involved: immunity typically lasts for many decades [3,5–10]. One approach is to undertake cross-sectional studies that measure immunity in individuals at different times following immunization [5,7,8,11]. While cross-sectional studies provide an estimate of the average rate of loss of immunity, it is difficult to determine the variation in the rate of loss of immunity among different individuals. There are only a few longitudinal studies that follow the decline in immunity to a vaccine and/or virus antigen over time in different individuals [12–15]. Most of these studies focus on the responses to a single vaccine or virus. In this analysis, we used data from a longitudinal study that followed the antibody levels to a panel of 7 vaccine or virus antigens in serum samples drawn from 45 individuals over several decades, as described in detail in ref [12]. The first paper describing this dataset found that antibody responses to virus infections or vaccination with live-attenuated viruses (vaccinia, measles, mumps, rubella, varicella zoster virus [VZV]) were remarkably stable, with half-lives ranging from 50 to over 200 years. In contrast, the antibodies elicited by protein antigens (tetanus and diphtheria toxoids) waned more rapidly, with half-lives of 11 and 19 years, respectively. In the current study, we asked two sets of questions. The first was to quantify the heterogeneity in antibody responses to different vaccines and viruses at the population level. We wanted to determine the extent to which this heterogeneity depended on the vaccine or virus, the individual (e.g., strong versus weak responders), and the interactions between the vaccine or virus and the individual. A second goal was to use these data to quantify the time to loss of protective immunity to each vaccine or virus at the population level. In particular, we wanted to know both the average duration of protective immunity to each vaccine or virus antigen as well as the extent of the variation of the time to loss of protection between different individuals in the population. Our approach is illustrated in Fig 1. First, we rescaled the antibody titer so that the threshold of protection was equal to 1, and we set the mean time for the time series for each individual to t = 0 (panel A). We then used a mixed-effects modeling framework to characterize the heterogeneity in antibody responses. This framework allowed us to determine the extent to which heterogeneity in responses depends on the vaccine or virus that induced the response, the individual (i.e., whether there were "strong responders" who tend to make high responses to all vaccines or viruses or who lose immunity more slowly than others in the population), and the interaction between these two effects. It also allowed us to quantify the heterogeneity in magnitude and decay rate of responses to each vaccine or virus and whether there was a relationship between the magnitude and decay rate (panel B). We then used a simple exponential decay model to determine the longevity of immunity in different individuals to different vaccine or virus antigens and quantify the loss of immunity to each vaccine or virus in the population (panel C). We used data from a study that determined the antibody responses to a number of vaccine and viral antigens in a group of 45 individuals over a period of between 5 and 26 years, with a mean range of about 15.2 years (see [12] for details). These samples were taken as part of a center-wide, comprehensive program to permit serologic testing of people working in close proximity to nonhuman primates. We focused on the antibody responses to measles, rubella, vaccinia, mumps, and VZV antigens that were elicited either by a live-attenuated vaccine or natural infection, as well as antibodies to tetanus and diphtheria antigens that were elicited by immunization with inactivated protein toxin (i.e., toxoid) vaccines. Because we wanted to consider the decay of antibody in the absence of boosting, the antibody data were curated to remove spikes due to revaccination or infection as described in more detail previously [12]. This involved censoring to exclude the following: timepoints for 3 years after immunization, when there was a rapid change in antibody levels [12,16], seronegative or unvaccinated individuals, and individuals with fewer than 4 contiguous data points. We then kept the time series with the largest number of contiguous data points for the response of each individual to each vaccine or virus antigen. As previously described, antibody titers were measured using ELISA and calibrated when possible in terms of international units (IUs). This allowed us to rescale the antibody concentration by dividing it by the level at which protection is lost [8,17–21]. In our plots and analysis, the magnitude of responses is shown as the log of the scaled titer. We did not have a level at which protection is lost for mumps and VZV, and for these, we scaled against the threshold of detection for the ELISA assay for that antigen. The level of antibody required for protection for different infections was taken from the literature and was assumed to be the same for all individuals—we did not consider variation in the protective threshold between different individuals due to lack of relevant data (see Discussion). Note that we did not consider the absolute magnitude of the antibody response (e.g., moles/L or mg/mL) because the ELISA assays used do not measure this quantity. Estimating the variability in the magnitude of the responses of different individuals to a given vaccine required taking into account uncertainty in the time of vaccination or infection and the different ages covered by the time series for different individuals. Because we do not find a significant correlation of antibody titer with age (and gender), but do find a strong correlation with time (see analysis in [12]), we used time rather than age as the main factor governing antibody titer. Because we did not know the time of vaccination, we shifted the time axis so that time equal to 0 corresponded to the mean age of the time series for each individual, and we used the intercept as a summary measure of the average magnitude of the response of the individual. We emphasize that the magnitude is not the peak magnitude just after vaccination or infection but rather the magnitude at the mean timepoint for that time series, which is expected to be many years or decades after vaccination or infection. We used a mixed-effects model framework (Eq 1) to determine the contributions of different factors to the magnitude (scaled titer) and decay of the antibody response. The factors we considered are vaccine or virus antigen, the individual, and the interaction between these two. We assumed that the loss of antibody for each individual follows an exponential decay so that the log of the antibody titer decays linearly with time [5,7,8,12]. We used a linear mixed-effects model (implemented using the lme4 package [22] in R [23]) described below. log10(ScaledTiterij(time))=+a−b*time←fixedeffects+ui−vi*time←randomeffectsforvaccinei+uj−vj*time←randomeffectsforindividualj+uij−vij*time←randomeffectsforvaccinei×individualj+e←residualforthemodel (1) a and b are the fixed, and u's and v's are the random effects for the magnitude and rate of decay. ui and vi tell us how much the response differs for different vaccines, uj and vj tell us how much an individual's response differs from the average response across all vaccines, and the terms uij and vij describe the interaction between vaccine and individual. Statistical support for each random-effect parameter was determined by the extent to which its inclusion improved the Akaike Information Criterion (AIC) of the model [24]. Specifically, we determined ΔAIC = AIC of the model without that random effect minus AIC of the full model in Eq 1. The magnitude of ΔAIC indicates the measure of support for including the random effect, with 4<ΔAIC<10 indicating some support and ΔAIC>10 indicating strong support. We also checked the robustness of the results obtained using the mixed-effects model by repeating our analysis using a fixed-effects model and obtained similar results. We estimated the duration of immunity of an individual to each vaccine in the absence of boosting by calculating the time (relative to the mean of the time series for each individual) for immunity to fall to the threshold (taken from the literature) at which immunity is lost. Because we do not have data on variation in the protective level of immunity in different individuals, this threshold is assumed to be the same for all individuals. Prior studies suggest that, following an initial rapid decay following infection or vaccination, the magnitude of the antibody as well as T-cell responses wane exponentially with time [5,7,8,12]. As described earlier, the data were censored to exclude timepoints 3 years after immunization or infection when there is a rapid change in antibody levels. This allowed us to use an exponential decay model for the waning of antibodies. Let Mij equal the magnitude of the response relative to protective titer and Dij the rate of decay of the response for individual j to vaccine i; then the time, Tij, to loss of immunity is given by Tij=log10(Mij)Dij=a+ui+uj+uijb+vi+vj+vij (2) The curated data for the changes in antibody titer over time in different individuals to different vaccines are shown in S1 Fig, and the rescaled data are shown in Fig 2. The different panels show the responses to the different virus or vaccine antigens, and the different colors correspond to different individuals. The y-axis shows the logarithm of the antibody titer rescaled by dividing by the threshold for protection, and the x-axis shows time in years, with the mean of the time series for each individual assigned a value of time equal to 0. We used a mixed-effects model framework described by Eq 1 to determine the contributions of different factors on the magnitude and rate of decay of antibody responses. The factors we considered were vaccine (or virus) antigen, the individual, and the interaction between these two. The results are shown in Table 1. The standard deviation (SD) indicates the amount of variation in the random effects for magnitude and decay rate for the relevant factor. This can be interpreted as follows: the random-effects model fits a line to each time series, and the variability in the magnitude and decay rate between these time series is partitioned between the 3 levels. Hence, for example, a large SD for vaccine magnitude (first line of Table 1) indicates large differences in the magnitude of responses from different vaccines. We see that all 3 factors (vaccine, individual, and vaccine × individual) contribute to the magnitude and decay rate of responses. The small residual indicates that, together, these factors account for most of the variation in the responses that are observed. We compared the relative contributions of variation in the magnitude and variation in the decay rate to the variation observed. The SD of the decay rate in Table 1 corresponds to the variation arising per year. The ratio of the SD of the magnitude (units log10 [scaled titer]) to the SD of the decay rate (units log10 [scaled titer] per year) gives approximate time (in years) at which variation in the decay rate and magnitude of responses contribute comparably to the overall variation in the data (log10 [scaled titer]). This ratio is approximately 65, 52, and 38 for vaccine, individual, and vaccine × individual level effects, respectively. This indicates that, at times much less than 50 years, differences in magnitude contribute more than differences in decay rate of responses to variation in antibody titers. At times much greater than 50 years, differences in magnitude contribute less than differences in decay rate of responses to variation in antibody titers. The largest effect is due to the vaccine level and indicates large differences in both the magnitude and the rate of decay of responses to different vaccine or virus antigens. Notably, tetanus has the highest magnitude, on average 195 times the threshold for protection, while responses to vaccinia are on average only 1.7 times the threshold for protection. Tetanus also has the fastest decay rate at 6.2% per year, while responses to measles exhibit virtually no decay (decay of 0.2% per year). The correlation between magnitude and decay rate arises because responses to toxoid vaccines (tetanus and diphtheria) have higher magnitude (relative to the threshold for protection) and decay rates compared with responses to live or attenuated viruses, but this correlation only improves the AIC of the model marginally (ΔAIC = 2.4) because of the small number of vaccine or virus antigens being considered. The individual level also contributed significantly to the antibody titers in the mixed-effect model, though the effect size is modest. This showed that some individuals made higher responses on average to all vaccine or virus antigens than others, and the decay rate of antibody titers occurred slower in some than others (Table 1). With regard to the magnitude of responses, we estimate that an individual at the top 10th percentile would make about 3.2-fold higher response than an individual at the bottom 10th percentile. With regard to decay rates, we estimate that the response of an individual at the top 10th percentile would decay about 2.2% per year faster than an individual at the bottom 10th percentile. Furthermore, the negative correlation between the magnitude of the response of an individual and its rate of decay indicates that individuals who make higher responses on average have slower rates of loss of antibodies. This suggests that some individuals on average have larger, slower decaying antibody responses than others. The effect of variation in responses between individuals is also shown visually in S2 Fig, where we plot the relative magnitude and relative decay rates of the responses of different individuals. Finally, the interaction term (vaccine × individual) has a large effect. It indicates the amount of variation in the magnitude and rate of decay that is not attributed to the vaccine and individual-level effects. We note the small SD associated with the residual, which suggests that the model fits the data well. Fig 3 (and S2 Table) shows the average magnitude and decay rate of responses to different vaccine or virus antigens, as well as the extent of variation in these quantities. As has been previously noted [12], we found that antibody responses to diphtheria and tetanus (monovalent protein or toxoid antigen vaccines) decay faster than responses elicited to measles, rubella, and vaccinia (multivalent replicating virus vaccines). Our study suggests that this faster decay is compensated for by the responses to diphtheria and tetanus having a higher magnitude (measured relative to the threshold for protection) compared with responses to measles, rubella, and vaccinia. In addition to different vaccines or viruses eliciting responses of different magnitudes, there is considerable variation in the response of different individuals to a given vaccine or virus antigen (Fig 3). We looked for correlations between the magnitude of the response an individual makes and the decay rate of the response. We found that for a given vaccine, individuals who made large responses, for the most part, tended to have slower decay rates, though for any given vaccine this trend did not reach statistical significance (at p = 0.05 level) (S3 Fig and S2 Table). The large negative correlation between the magnitude and decay rate associated with the individual-level effect across all vaccines seen in Table 1 arises due to the following: correlations between the magnitude of the response to one vaccine or virus antigen and the decay rate of responses to other vaccine or virus antigens, as well as the aggregation of weak correlations across multiple vaccine and virus antigens. We calculate the time to loss of protective immunity, defined as the time at which the antibody titer reaches the level defined to be protective for that vaccine or virus antigen. We assume that the antibody titer in an individual continues to decline exponentially. This allows us to estimate the time for the antibody titer in each individual to reach the level defined as the threshold for protection for the given virus or vaccine using Eq 2, namely, T=log10(M)D, where M is the scaled titer at time equal to 0 and D is the decay rate for antibody concentration. We call the time until the antibody level reaches the defined threshold for protection for that pathogen the duration of immunity. Due to lack of data, we do not account for variation in the threshold of protection to a given pathogen in different individuals. In Fig 4, we plot the proportion of the cohort immune to each vaccine or virus antigen as a function of time. As described in Materials and methods, we set time to 0 at the mean of each individual's time series in order to make comparisons between people of different ages. Year 0 corresponds on average to about 45 years in age. We found that virtually all individuals in this small cohort have protective levels of antibodies to tetanus, diphtheria, measles, and rubella at year 0, but only about two-thirds of individuals have protective levels of antibodies to vaccinia at this time. We see different patterns for the loss of protective immunity elicited by protein or toxoid antigens (diphtheria and tetanus) versus live-attenuated vaccines or viruses (measles, rubella, and vaccinia). For both tetanus and diphtheria, it takes over 40 years for protective immunity to begin to be lost in our cohort. Thereafter, there is a relatively rapid loss of protective immunity, with about 7% and 12% of the population losing immunity to diphtheria and tetanus per decade. In contrast, immunity to measles and rubella begins to be lost sooner, but the decay rate of antibodies to these antigens is slower, with only 1% to 2% of the population losing protective immunity per decade. For vaccinia, just over 60% of the population have protective levels of antibodies at the outset (year 0), and this has dropped to 40% after 100 years—which corresponds to about 3% of the population losing protective immunity per decade. We analyzed the data from a unique longitudinal study that quantified the level of antibodies to a panel of vaccine and virus antigens over a period of several decades in 45 individuals [12]. The original paper describing this dataset quantified the average rate of decay of antibodies to the different vaccine or virus antigens. In this paper, we extended this analysis in the following ways. First, we explicitly describe the heterogeneity in both magnitude (scaled to the threshold of protection) and decay rates of responses in different individuals to different vaccine and virus antigens (see Fig 3). We used a mixed-effects model to determine the contributions of different factors—vaccine, individual, and vaccine × individual—on antibody responses to different virus and vaccine antigens. This directly allows us to compare the extent to which the antibody response depends on the particular vaccine, the particular individual, and the interaction between these two. Second, we used a simple model to estimate the rates of loss of protective titers of antibodies to different vaccine and virus antigens at the population level. Our results are summarized in Fig 5, where the y-axis represents the magnitude of the response and the x-axis the rate of decay of the antibody response. The responses of different individuals to a given vaccine is represented by a cloud of points of a given color. The plot allows visualization of the variation in both the magnitude and the decay rate of responses of different individuals to a given vaccine, and it indicates how this variation translates into different times to loss of immunity shown by the dashed lines. The variation due to the factor "vaccine" can be seen by the separation between the clouds of points of different colors—for example, there is virtually no overlap between the responses to tetanus and measles. The extent of variation in the responses of different individuals to a given vaccine or virus antigen is indicated by the size of the cloud of points of each color. The extent of variation we observe is similar to that observed in other studies. For example, studies of antibody and T-cell responses following smallpox and yellow fever vaccination show similar variation [8,25]. We now describe the insights obtained for vaccination and boosting. Fig 5 also allows us to visualize the effects of increasing the magnitude (moving points up on the figure) or decreasing the decay rate (moving points to the left) on the time to loss of protective immunity to these infections. Given the ranges of variation in magnitude and decay rate seen in Fig 5, the duration of protective immunity to live-attenuated vaccines (measles, rubella, and vaccinia) could be best improved by focusing on increasing the magnitude of the responses to these vaccines. We see, for example, that increasing the magnitude of the response to vaccinia by 10-fold would be much more effective to provide long-term protection than decreasing the rate of decay. In contrast, the duration of protective immunity to tetanus and diphtheria might be best improved by lowering the rate of waning of immunity to these antigens rather than increasing the magnitude of these responses. We observed that the fast rate of decay in responses to protein antigen vaccines (tetanus and diphtheria) was counteracted by the high magnitude of these responses relative to the levels needed for protection. This resulted in a similar fraction of the population having lost protective levels of antibodies after 50 years to tetanus, measles, diphtheria, and rubella (see Fig 4). This is a much longer timescale than the current boosting schedule for tetanus and supports the results of a much larger cross-sectional study of tetanus antibody titers [11]. The current study is an early step in the quantification of the longevity of memory to different vaccine and virus antigens in the human population. It has a number of limitations, and identifying these can aid in the design of future studies. The study is restricted to the antibody response. This is because it was possible to get longitudinal measurements of the magnitude of antibody responses from sera that had been stored for over several decades. Obtaining similar data for the dynamics of T-cell responses is possible but more complicated because it requires isolation and cryopreservation of peripheral blood mononuclear cells (PBMCs) and long-term storage in liquid nitrogen. Antibody levels to different vaccine and virus antigens were measured by their ELISA titer and normalized to the threshold of protection. The thresholds for protection were taken from the literature, and there are different standards for the protection to different infections [26,27]. We calculated the time for antibody titers to fall to the defined threshold for protection for the given vaccine or virus antigen. Due to a lack of data, we were not able to determine the consequences of variation in the threshold for protection between individuals. Different levels of immunity are required for different types of protection [27–30]. For example, higher levels of antibodies might be required to prevent infection, whereas lower levels of antibodies may not prevent infection per se but may still ameliorate disease or protect against lethal infection. Smallpox vaccination with vaccinia virus results in antibody titers that are predicted to provide full protective immunity against virulent poxvirus infection in about 40% of subjects at 100 years (Fig 4), and the remaining 60% of the population are likely to have partial protective immunity. This point was illustrated during the American monkeypox outbreak in 2003. Of 8 previously vaccinated individuals who were diagnosed with monkeypox, three-eighths had no clinical symptoms (monkeypox infection identified only through immunological assays), and the other five-eighths of vaccinated subjects had milder disease than that associated with unvaccinated monkeypox cases [31]. We did not have dates of prior vaccinations or infections, and some would occur prior to the period when the individuals were sampled. Because we were interested in antibody titers over the long term, we curated the data to eliminate a 3-year interval following a spike in antibody levels (see Materials and methods). Consequently, our results describe the duration of immunity in this population of adults rather than as a function of time since immunization. The lack of data on prior vaccinations also makes it difficult to consider how the number of prior exposures affects the level of boosting or changes in the decay rate of immunity. We note that antibody spikes were relatively infrequent (<0.3 events per 100 person-years) for vaccinia, mumps, and rubella, likely due to less frequent immunization as well as sterilizing immunity preventing boosting of immunity following vaccination with live-attenuated vaccine strains or natural infection. Antibody spikes following revaccination against tetanus and diphtheria (with protein toxoid antigens) were more common (4.9 events per 100 person-years). Finally, our extrapolation for the duration of immunity assumed that the rate of decay of antibody titers is constant and does not either increase (potentially due to negative effects of aging [32, 33]) or decrease (potentially due to "selection" for longer-lived plasma cells [15]) over time. We note that a cross-sectional study of antitoxin antibody responses suggests that antibody decay rates to tetanus and diphtheria are not significantly different in individuals under and over 50 years of age [11]. Overcoming the preceding limitations would require more detailed sampling just prior to and following immunizations, as well as extending the studies for a longer duration that incorporates older individuals. We examined the heterogeneity in both the magnitude and decay rate of antibody responses to different virus and vaccine antigens and used simple models to quantify how this heterogeneity affected the duration of protective immunity to a panel of vaccines and viruses. We found that variation in magnitude and decay rates of responses contribute comparably to the differences in antibody titers, that some individuals tend to make higher responses and these individuals also tend to have slower decay rates, and that different patterns of duration of protective levels of antibodies were elicited by replicating viruses and proteins. As noted in our previous study, we found that levels of antibodies to protein toxoid immunization to tetanus and diphtheria are higher relative to the protective threshold and decay faster than those elicited by replicating viruses. Integrating the effects of these two factors suggested that, in the absence of boosting, more adults will, for the first 4 decades, tend to have protective levels of antibodies to tetanus and diphtheria in comparison with measles, rubella, and vaccinia, suggesting the need for reevaluation of their boosting schedules. Our results emphasize the need for collecting longitudinal data that can allow quantification of the magnitude and decay of antibody titers in a large cohort of individuals to different vaccine and virus antigens, and we illustrate how integrating such measurements with models will allow us to generate a nuanced quantitative picture for the loss of protective immunity and help optimize boosting strategies.

10.1371/journal.ppat.1002062

Bacteria-Induced Dscam Isoforms of the Crustacean, Pacifastacus leniusculus

The Down syndrome cell adhesion molecule, also known as Dscam, is a member of the immunoglobulin super family. Dscam plays an essential function in neuronal wiring and appears to be involved in innate immune reactions in insects. The deduced amino acid sequence of Dscam in the crustacean Pacifastacus leniusculus (PlDscam), encodes 9(Ig)-4(FNIII)-(Ig)-2(FNIII)-TM and it has variable regions in the N-terminal half of Ig2 and Ig3 and the complete Ig7 and in the transmembrane domain. The cytoplasmic tail can generate multiple isoforms. PlDscam can generate more than 22,000 different unique isoforms. Bacteria and LPS injection enhanced the expression of PlDscam, but no response in expression occurred after a white spot syndrome virus (WSSV) infection or injection with peptidoglycans. Furthermore, PlDscam silencing did not have any effect on the replication of the WSSV. Bacterial specific isoforms of PlDscam were shown to have a specific binding property to each tested bacteria, E. coli or S. aureus. The bacteria specific isoforms of PlDscam were shown to be associated with bacterial clearance and phagocytosis in crayfish.

Invertebrate animals lack an adaptive immune system and have no antibodies. Vertebrate antibodies belong to the immunoglobulin super family of proteins, and one other member of this large family is the Down syndrome cell adhesion molecule or Dscam. Of specific interest is that Dscam proteins in invertebrates show a great diversity of isoforms, and its gene structure in Drosophila melanogaster and other insect species allow for more than 30,000 different isoforms. Dscam proteins are important for the interaction between neurons in insects, but recently a role for this hypervariable protein in immune defense has been shown. Here, we show that Dscam proteins with similar highly variable structures are present in a crustacean, the freshwater crayfish Pacifastacus leniusculus. We also found that specific isoforms could be induced in the animal after injection of different bacteria. The Dscam isoforms induced by Escherichia coli were found to cluster together in a phylogenetic analysis. Furthermore we produced recombinant proteins of the different isoforms that were induced by E. coli and Staphylococcus aureus and we could demonstrate that these proteins can bind specifically to their corresponding bacteria. The bacteria specific isoforms of Dscam were also shown to be associated with bacterial clearance and phagocytosis in crayfish. Our study therefore provides new insights into the function of invertebrate Dscams in immunity.

The immunoglobulin super family (IgSF) is composed of proteins that contain at least one immunoglobulin domain [1]. Several members of IgSF are expressed on the cell surface and there serve as receptors for diverse ligands, and contribute to a variety of cellular activities [2]. In vertebrates, many IgSF members play essential roles as immune molecules (also known as antibodies) by recognizing non-self entities and then promoting their elimination [3]. Although it appears as if all invertebrates lack true antibodies, diversified IgSF molecules have been shown to be involved in immune defense of several invertebrates [4], [5]. However, this does not imply that the diversification of IgSF in invertebrates have any relation to the antibody diversification in vertebrates [5]. Recently, one IgSF member, the Down syndrome cell adhesion molecule gene or Dscam, that can generate hypervariable isoforms through alternative splicing was shown to act as an opsonin and to enhance phagocytosis in insects [4], [6]. Dscam was first detected on the human chromosome 21q22, a region associated with Down Syndrome [7]. Then, orthologues of Dscam were identified in various species and the typical domain structure of the Dscam gene is highly conserved. The Dscam molecules are widely expressed in the nervous system and play an essential role in neural circuit formation [8], [9]. Moreover, transcripts of fly Dscam was detected in fat body cells and hemocytes, which both are important components of the insect immune system [4], [10]. The Dscam in hemocytes of Drosophila melanogaster and Anopheles gambiae can bind to Escherichia coli and potentially acts as both a phagocytic receptor and as an opsonin. In the mosquito, Dscam can generate pathogen-specific spliced forms upon immune challenge [6]. These findings provide some evidences that Dscam may have functions not only in neuronal wiring, but also in innate immunity in insects. In the present study, the full-length cDNA and variable regions of P. leniusculus Dscam (PlDscam) were identified and characterized. We also present results showing that different isoforms of PlDscam can be induced by immune challenge, that they can bind bacteria and that they are important in bacterial clearance. A large open reading frame of PlDscam (6,009 bp) was identified that encodes a polypeptide of 2,002 aa (Figure 1A). The closest sequence matching that of PlDscam was Dscam of L. vannamei (identity  = 85%). Domain homology analysis using SMART showed that the deduced amino acid sequence contains a signal peptide at amino acids 1–24, ten tandem repeated immunoglobulin domains (Ig), six fibronectin type III domains (FNIII) and a transmembrane domain (TM). The domain organization of PlDscam is 9(Ig)-4(FNIII)-(Ig)-2(FNIII)-TM (Figure 1B). It also has a conserved cell attachment sequence (Arg-Gly-Asp: RGD motif) between Ig6 and Ig7 (Figure 1A). The sequence in the 3′ UTR contains a polyadenylation signal (AATAA) and is located at 55 bp upstream of the poly A tail (Figure 1C). To identify the variable regions of PlDscam, several regions of the PlDscam were amplified using different primer pairs and the location of each pair of primers is shown in Figure S1. Fifty clones of each of the six amplified regions were selected and sequenced, translated and aligned using ClustalW. Alternatively spliced mRNA segments of the PlDscam were detected in the N-terminal of Ig2 and Ig3, in the entire length of the Ig7 domain, in the transmembrane domain and in the cytoplasmic tail (Figure 2B–E and Figure S1A). In total 12, 29, 32 and 2 alternative spliced forms of the exons encoding Ig2, Ig3, Ig7 and the transmembrane domains, respectively were detected and therefore at least 22,000 different unique isoforms could in theory be generated (Figure 2A–2E). The other domains were highly conserved, especially Ig8-FNIII4. The cytoplasmic tail of PlDscam contains some highly conserved motifs similar to the corresponding area in Dscams of other species (Figure S1B). In crayfish, PlDscam was abundantly expressed in the heart, moderately expressed in the testis, the hematopoietic tissue (HPT), brain and nerve, whereas low expression was detected in the stomach, gill, muscle and hemocytes. The PlDscam was not detected in hepatopancreas or in intestine (Figure 3A). A phylogenetic analysis clearly separated the Dscam proteins of vertebrates and invertebrates in two different groups and PlDscam clustered with other members of invertebrate Dscams (Figure 3B). To investigate whether immune challenge induces higher expression of PlDscam in crayfish hemocytes, LPS, peptidoglycan (PG), E. coli, S. aureus or WSSV was used as immune elicitors. The results of quantitative RT-PCR revealed that the PlDscam mRNA expression was significantly induced by LPS-, E. coli- and S. aureus at 6 h to 24 h post injection compared to controls (Figure 4A, 4C, 4D). In contrast, the expression profile of PlDscam in PG- and WSSV- injected animals was not changed (Figure 4B and 4E). In order to characterize the major isoforms of PlDscam in hemocytes of normal and bacteria injected crayfish, PlDscam cDNA fragments encompassing the signal peptide to Ig3 from each group were amplified and cloned (Figure 5A). A total of 50 clones of each cDNA fragment were sequenced. Ten isoforms were detected in all groups, i.e. normal, E. coli injected and S. aureus injected crayfish. These isoforms were named as Normal isoform (N) (Table S2 and Table S3). Furthermore ten abundant isoforms were found only in the E. coli or S. aureus injected group and were named E. coli induced isoform (E) or S. aureus induced isoform (S), respectively (Table S2 and Table S3). All isoforms were subjected to multiple sequence alignment using clustalW. The similarity of each PlDscam isoform was then clustered by the maximum likelihood (ML) and Bayesian inference (BI) methods. As shown in Figure 5B, the clustering tree contained two major branches. The E. coli induced isoforms were separated into one branch, whereas the other branch consisted of two sub-branches of normal isoform and S. aureus induced isoform (Figure 5B). The E isoforms of PlDscam isoforms had a VNKEYIIRGDSA(F/I)LKCSIPSFVA(D/N) motif and a EIGSPATFTCRAQAHPVPQY motif present at the N-terminal of the Ig2 and Ig3 domains, respectively. As shown in Figure 5B and 5C the isoforms N6, E2 and S9 are different, and therefore we used these isoforms as representative alternative spliced forms of normal, E. coli- and S. aureus-induced isoforms for a bacterial binding assay. Recombinant proteins covering the Ig1-Ig3 domains of the PlDscam isoforms of N6, E2 and S9 were produced and tested whether these domains have any putative function in binding to E. coli or S. aureus. These proteins were expressed in bacterial systems and the size of all soluble recombinant proteins was ∼67 kDa (Figure 6A). All recombinant proteins were fused with GST at the N-terminus and contained the Ig1-Ig3 domains. We used these rPlDscams in bacterial binding assays to reveal whether different isoforms are capable of direct binding to bacteria. The GST protein was used as a non-specific binding control in these experiments. The in vitro bacterial binding assays showed that all isoforms of PlDscam had different binding ability to the two tested bacteria. The rPlDscam of isoform E2 clearly bound to E. coli and had significantly higher binding than the N6 and S9 isoforms. In contrast, binding of rPlDscam S9 to S. aureus was significantly higher (P<0.05) than that of N6 and E2 (Figure 6B). Recombinant proteins of isoforms E2 and S9 that specifically interacted with E. coli and S. aureus, respectively, were found to interfere with bacterial clearance and phagocytosis in crayfish. Pre-incubation of E. coli with E2 followed by injection into the animals did increase the number of bacteria in circulation (Figure 6C). Similarly with S. aureus, the number of bacteria was the highest in the S9 isoform of rPlDscam pre-incubated group (Figure 6D). This result clearly indicates that the specific rPlDscam could interfere with bacterial binding to the hemocytes. These results were in agreement with the phagocytosis assay, since if E. coli and S. aureus were coated with the E2 and S9 isoforms respectively, this resulted in a significant decrease in the phagocytic activity (Figure 6F). The role of PlDscam during a WSSV infection was investigated using PlDscam RNAi to suppress the PlDscam expression. The PlDscam gene was completely knocked down in an HPTcell culture (Figure 7A) whereas the 40S ribosomal gene was unaffected. However, PlDscam silencing did not have any effect on WSSV replication as shown with no changes in transcription level of WSSV structural protein transcript VP28 between control and PlDscam silenced groups (Figure 7B and 7C). Moreover, this result also indicates that the expression of PlDscam was not affected by WSSV infection. This agrees with our previous experiment where we injected WSSV to live crayfish and the transcript level of Dscams was not affected (Figure 4E). The typical domain structure of Dscam with an extracellular domain, a single transmembrane domain and a C-terminal cytoplasmic tail, is highly conserved within arthropods and vertebrates [7], [11]. This domain architecture was also found in PlDscam. Diversity in Dscam is generated through alternative splicing and variable alternative exons were found in the N-terminal half of Ig2 and Ig3, in the entire Ig7 domain and in the complete transmembrane domain [12]. These four variable domains are highly conserved within arthropods, including the PlDscam. It is noticeable that the alternative splicing of these exons clearly contributes to separate Dscam of vertebrates from invertebrates in our phylogenetic analysis [11]. The Dscam was initially identified for its essential roles in neuronal wiring, so its transcript is present in high quantity in neuronal organs [13]. The PlDscam was also detected in the neural system of crayfish. However, recently a putative role of Dscam in host defense was shown in D. melanogaster and A. gambiae [4], [6]. Both D. melanogaster and A. gambiae Dscams were required for host resistance and phagocytosis of bacteria [14]. Dscams of the mosquito respond to pathogen infection by generating specific isoforms and these pathogen-specific isoforms of Dscam can bind directly to pathogens [15]. The response of IgSF molecules to pathogens does not only generate specific isoforms, but also an increase in the number of these isoforms, such as is the case with the fibrinogen-related proteins (FREPs) in snails [5]. FREP production is enhanced following parasitic invasion, and these proteins can bind to parasitic invaders or their products. In crayfish challenge with both Gram-negative and Gram-positive bacteria induced higher transcription of PlDscam. A high transcription of PlDscam was obtained after LPS injection, whereas PG or viral injection of WSSV had no such effect. Most Gram-positive bacterial cell walls or cell membranes contain several components, including PG, lipoteichoic acid (LTA), and lipoproteins [16]. The high transcription of PlDscam achieved as a response to S. aureus injection but no response to PG may indicate that the PlDscam respond to other components of the S. aureus cell wall, such as LTA which has similar physiochemical properties to LPS from Gram-negative bacteria [17]. Previous results from D. melanogaster and L. vannamei showed that the hemocytes of immune challenged animals exhibited higher variability of the Ig2 and Ig3 domains but only a few Ig7 variants compared to normal animal [4], [18]. This is the reason why we studied bacteria specific induced isoforms and produced recombinant proteins covering only the Ig1-Ig3 region. Interestingly, PlDscam isoforms of E. coli injected crayfish mainly encoded “VNKEYIIRGDSA(F/I)LKCSIPSFVA(D/N)” and “EIGSPATFTCRAQAHPVPQY” motifs at the N-terminal part of the Ig2 and Ig3 domains, respectively. This implies that these two motifs might be important parts of the specific PlDscam isoforms in E. coli-infected crayfish. This is consistent with results from our bacterial binding assay, which showed that the recombinant proteins containing these two motifs (E2), could bind to E. coli better than the other isoforms (N6 and S9). In addition, binding of S. aureus-induced isoforms (S9) to S. aureus was also higher than E2 and N6. These results indicate that the pathogen induced isoforms of PlDscam have a specific binding property to each type of challenged bacteria. This specific interaction may be associated with some immune defense reaction as shown with mosquito Dscam [6]. To address this question, recombinant proteins (N6, E2 and S9) were pre-incubated with bacteria or FITC conjugated heat killed bacteria to study bacteria clearance and phagocytosis in vivo. When, the bacteria-induced PlDscam isoforms were coated on E. coli or S. aureus and then injected into live animals, this resulted in lowered clearing rates of bacteria in the hemolymph. This implies that the specific PlDscam fragments covered the binding sites of the bacteria so they could not bind to the membrane bound PlDscam on the hemocytes and hence the bacterial number increases since the clearance of bacteria by phagocytes is inhibited. In the case of mosquito, AgDscam is not only a determinant of resistance to bacteria but also affects the resistance towards the malaria parasite Plasmodium [6]. This implies that Dscam might be involved in other host pathogens reactions in crayfish. The Dscam belongs to a subfamily of the Immunoglobulin super family (IgSF). Indeed, many members of IgSF proteins have been reported to interact with and promote entry of numerous virus, including for example the junctional adhesion molecule A (JAM A), that could bind with and facilitate entry of reovirus [19], [20]. So, it is possible that PlDscam could bind to virus like the other members of IgSF and maybe facilitate entry of virus into crayfish hemocytes. WSSV is a virulent pathogen that causes death in many species of crustaceans such as crayfish [21] and therefore we tested a possible relationship between PlDscam and this important arthropod virus. We performed experiments to reveal whether WSSV challenge could increase transcription of PlDscam and whether PlDscam RNAi had any effect on WSSV replication. However, we could not detect any increase in PlDscam mRNA expression after WSSV infection, and more important, if the PlDscam gene was completely silenced this could not affect WSSV infection or replication. Healthy intermolt freshwater crayfish (P. leniusculus) were obtained from Lake Hjälmaren, Sweden and maintained in aerated tap water at 10°C. Total RNA (at least 1 µg) was extracted from the heart and converted into cDNA using ThermoScript (Invitrogen). The degenerate primers were designed from the conserved region of insect Dscam including Drosophila, Apis, Aedes and Tribolium (DSCAM-e5 F: 5′-AARCAYMGIYTIACIGGIGARAC-3′; DSCAM-e7 R: 5′-GTI ARIACIGTYTCIACI SWYTC-3′). The resulting PCR product was purified and cloned into a TOPO vector (Invitrogen) and sequenced. A partial sequence was used for the further step. Gene specific primers for Rapid amplification of cDNA ends (RACE) technique (Table S1) were designed from the partial sequence and 5′ or 3′ RACE-PCR was performed with a SMART universal primer A mix (SMARTer RACE cDNA Amplification Kit user manual, Clontech). Thermal cycling was as follows: 25 cycles of 94°C 30 s, 68°C 30 s, and 72°C 3 min. The 5′ and 3′ RACE PCR products were cloned into TOP10 vector and sequenced. The nucleotide sequence of PlDscam was compared to others in Genbank using BlastX. Multiple sequence alignment was done by ClustalW (http://www.ebi.ac.uk/Tools/clustalw/index.html). The deduced amino acid domain was predicted with SMART (http://smart.embl-heidelberg.de/). A phylogenetic tree representing the relationship between PlDscam and other proteins was analyzed by the maximum likelihood (ML) and Bayesian inference (BI) methods. A PhyML program (under the Whelan and Goldman (WAG) and gamma model with four categories) was used in ML analysis [22]. For the BI method, we used MrBayes program [23] with CAT model (3,000 cycles, first 1,000 cycles removed as burn-in, and the analysis was repeated three times with identical results). Internal blanch support values was from analysis of 1,000 ML bootstrap replicates. This evolutionary phylogram was based on the conserved region of Dscam from Ig8 to FNIII4, whereas phylogram for clustering of all PlDscam isoforms was based on the similarity of their amino acid sequence and were used with the same methods as described above. RNA from various tissues, including hepatopancreas, stomach, intestine, heart, hematopoietic tissue (Hpt), muscle, brain, hemocytes and nerves, was extracted following the instruction of GenElute Mammalian Total RNA Miniprep kit (Sigma) followed by treatment with RNase-Free DNase I (Ambion, Austin, TX). Complementary DNA was synthesized using ThermoScript. PlDscam gene specific primers (GSP-PlDscam-F, 5′- TGGGAAGTGATGCCAGGTTAGA-3′; GSP-PlDscam-R, 5′-TTGAATCAGCAGACATAACCAAAGC-3′) were designed from full-length cDNA of PlDscam and its PCR product covered the conserved Dscam region (from Ig8 to FNIII3). A 40S ribosomal gene was used as internal control in all PCR experiments and its specific primers (40S-F, 5′-CCAGGACCCCCAAACTTCTTAG-3′; 40S-R, 5′-GAAAACTGCCACAGCCGTTG-3′) were designed from P. leniusculus Lamda Zap Express library Hpt cDNA (Genbank accession no. CF542417). PCR conditions were as follows: 94°C 2 min, followed by 30 cycles of 94°C 20 s, 58°C 20 s, and 72°C 1 min for the PlDscam gene and 25 cycles for 40S ribosomal gene. The PCR products were analyzed on 1.2% agarose gel stained with ethidium bromide. Due to the large size of full length of PlDscam, it was necessary to use several pairs of primers for identification of the different variable regions. PCR of each region was performed with gene specific primers (Table S1) and thermal cycling was as follows: 94°C 2 min, followed by 30 cycles of 94°C 20 s, 60°C 20 s, and 72°C 1.30 min. The PCR products from four different tissues, such as heart, brain, hemocytes and HPT, were cloned into TOP10 vector and 25 individual clones from each tissue were sequenced. E. coli and S. aureus were cultured in LB broth at 37°C until OD 600 was ca 0.5. The bacteria were washed three times with 0.85% NaCl by centrifugation at 900 g for 10 min at room temperature. The pellets were resuspended in sterile 0.85% NaCl and adjusted to an approximate concentration of 2×108 CFU/ml. Two groups of three crayfish were injected in the base of the forth walking leg with 100 µl of E. coli and S. aureus, respectively (approximately 2×107 CFU/crayfish). The control group was injected with 100 µl of 0.85% NaCl. WSSV was purified with a method described by Xie et al.[24]. The purified virus was resuspended in sterile crayfish saline buffer (CFS: 0.2 M NaCl, 5.4 mM KCl, 10 µM CaCl2, and 10 mM MgCl2, 2 mM NaHCO3, pH 6.8) at a concentration of 2×107 copies/ml. One hundred microliter of WSSV (equivalent to 2×106 copies) or CFS (as control group) was injected as previously described. The experimental setup was made in triplicates. Two groups of three crayfish received injections with 100 µl lipopolysaccharides (LPS: Sigma, from E. coli) or peptidoglycan (PG: Sigma, from S. aureus) in a sterile CFS with a concentration of 0.2 mg/ml. One hundred microliter of CFS was used for the controls. Hemolymph of crayfish from all experiments was collected at 0, 6, 12 and 24 h post injection and the hemocytes were separately isolated for RNA extraction. The transcript levels of PlDscam were detected by quantitative RT-PCR using the QuantiTect SYBR green PCR kit (QIAGEN). The expression of PlDscam was normalized to the expression of the mRNA encoding the crayfish ribosomal protein gene (R40s) for each sample. The primers used are shown in Table S1. The qPCR reactions contained 5 µl of 1∶10 diluted cDNA template, 1× QuantiTect SYBR Green PCR master mix (QIAGEN) and 5 µM forward and reverse primers in a 25-µl reaction volume. The following amplification profile was used: 95°C for 15 min, followed by 45 cycles of 94°C for 15 s, 58°C for 30 s, and 72°C for 30 s. All qPCR reactions were performed in duplicate. The hemocytes from a least three crayfish were used for each time point. The variable region, from the signal peptide to the Ig3 domain region, was amplified from hemocyte cDNA templates from groups of crayfish injected with E. coli or S. aureus, respectively, and the control group at 12 h post injection with F1 and R2 primers (Table S1). The PCR products from each of these groups were subcloned into TOP10 vector and fifty colonies of each group were sampled and subsequently sequenced. Multiple sequence alignment was done by ClustalW and clustering tree of the different isoforms was constructed as described above. The ORFs without the signal peptide encoding different PlDscam isoforms were amplified from the original templates N4, E1 and S7 with Dscam-expression-BamHI-Forward (5′-TTTTGGATCCAACCCGACAACCGTGTGGACTTCA-3′) and Dscam-expression-XhoI-Reverse (5′- TTTCTCGAGCCAGGTAACAGACTTGACGGGGTTG-3′) primers. The resulting insert was cloned into pGEX-4T-1(GE healthcare) at BamHI and XhoI and transformed into BL21 E. coli. Single colonies were grown in LB medium containing 100 µg/ml ampicillin to OD600 = 0.6 and induced with 1 µM IPTG for 5 h at 37°C. The protein was expressed as a fusion product with a glutathione S-transferase part at the N-terminus of rPlDscam. After purifying this GST-fusion protein on a GST-trap FF column (GE healthcare), the presence of the recombinant protein was confirmed by western blot. The protein samples were subjected to 12% SDS-PAGE and then transferred electrophoretically to PDVF membranes. The membrane was blocked by immersion in 10% skimmed milk in TBST for 1 h and washed three times in 1 x TBST (10 mM Tris-HCl, pH 7.5, containing 150 mM NaCl and 0.1% Tween 20). The membrane was then incubated with 1: 2,000 dilution of a primary antibody for GST (Sigma) in TBST for 1 h. Then, the previous washing procedure was repeated before the membrane was incubated with anti-mouse-IgG peroxidase-linked species-specific whole antibody from sheep (GE Healthcare) at 1: 3,000 in 1xTBST for 1 h and washed with TBST for 3×10 min. For detection, the ECL Western blotting reagent kit (Amersham Biosciences) was used according to the manufacturer's instructions. A bacterial binding assay was modified from the method described by Yu et al.[25] Briefly, E. coli or S. aureus was prepared as previously described. An aliquot 100 µl of the bacteria was immobilized on 96 well plates at approximately 108 CFU/well by incubating for 30 min at room temperature and then shifted to incubation at 4°C overnight. The bacteria were then blocked with 3% (w/v) BSA in TBS buffer at room temperature for 1 hour. To access the binding of the proteins to the bacteria, 20 µg of each PlDscam isoform (with GST fusion tag) or GST diluted in 1% BSA in TBS were added to the wells and incubated at room temperature for 2 h. Following three careful washes with TBS, the bound PlDscam protein was detected with the GST antibody (1∶2,000) followed by rabbit antimouse HRP-conjugated secondary antibody (1∶3000). After adding 100 µl tetramethylbenzidine or TMB substrate (Sigma) for 20 min and 100 µl of stop solution (0.5 M sulfuric acid), the absorbance of the resulting color was measured at 450 nm. All binding assays were performed in triplicates and experiment was repeated three times. E. coli and S. aureus were washed six times in 0.9% NaCl at 1,200× g for 10 min and then incubated with 10 µg recombinant proteins of each PlDscam isoforms for 1 h at 4°C, followed by washing six times with 0.9% NaCl. One hundred microliter of E. coli or S. aureus at concentrations of 6×108 and 3×109 CFU/ml, respectively, in CFS were injected into crayfish. The bacteria count was carried out in hemolymph collected 40 min and 3 h after the bacterial injection. The homonym was serial diluted serially and was then dotted onto LB agar (10 µl for each dot) and then incubated at 37°C overnight followed by counting the bacterial colony forming units (CFU). Both, heat killed E. coli and S. aureus were conjugated with fluoresce in isothiocyanate (FITC) using a method previously described by Hed [26]. Briefly, heat-killed E. coli and S. aureus were washed six times in 0.9% NaCl at 1,200× g for 10 min and then incubated at concentration of 109 particles/ml in 0.1 M Na2CO3 containing 0.1 mg/ml FITC (Sigma), pH 9.5 for 30 min at 37°C, followed by washing six times with 0.9% NaCl as above. The FITC-conjugated heat-killed bacteria were incubated with 10 µg recombinant proteins of each PlDscam isoforms for 1 h at 4°C. The bacteria were washed five times as described above and resuspended at a concentration of 108 particles/ml in CFS. Two hundred microliter of FITC-conjugated heat-killed bacteria coated with recombinant PlDscam isoform (4×106 particles/ml in CFS) was injected into the animals via the base of the fourth walking leg. Hemocytes were bled for phagocytosis detection at 1.5 h post FITC-conjugated heat-killed bacteria injection. The fluorescence of FITC-conjugated heat-killed bacteria particles was quenched by adding 20 µl of 0.04% trypan blue (Pfaltz & Bauer, Waterbury, CT). The ingested bacteria were easily detected under the UV light microscope due to their fluorescence. The percentage of phagocytosing cells was determined by counting 10 individual fields and dividing the number of cells with ingested fluorescent bacteria with the total number of counted cells. Each experiment was performed in triplicates. The hematopoietic tissue was dissected according to Söderhäll et al.[27]. The Hpt was washed with CPBS (crayfish phosphate buffered saline: 10 mM Na2HPO4, 10 mM KH2PO4, 150 mM NaCl, 10 µM CaCl2, and 10 µM MnCl2, pH 6.8) and incubated in 600 µl of 0.1% collagenase (type I and IV) (Sigma) in CPBS at room temperature for 45 min to separate the Hpt cells. The separated cells were washed twice with CPBS by spinning down at 800× g for 5 min at room temperature. The cell pellet was re-suspended in modified L-15 medium and subsequently cells were seeded at a density of 2.5×106 cells/150 µl in 96-well plates. The Hpt cells were supplemented with partially purified plasma as a source of astakine [28] after 1 h of attachment at room temperature and the culture plates were incubated at 16°C, and 1/3 of the medium was changed at 48 h intervals. dsRNA of PlDscam was designed from the conserved region during Ig8-FNIII3. Gene specific primers for PlDscam and GFP was incorporated with a T7 promoter sequence (italic letters) at 5′ ends (DscamRNAi-F, 5′- TAATACGACTCACTATAGGGGCATCAAGCTGAGTGGACAA-3′; DscamRNAi-R, 5′- TAATACGACTCACTATAGGG GAAGCCAGGTAGGGGAAATC -3′ and GFP 63+, TAATACGACTCACTATAGGGCGACGTAAACGGCCACAAGT; GFP 719-, TAATACGACTCACTATAGGGTTCTTGTACAGCTCGTCCATG) and used to amplify PCR products as template for dsRNA synthesis. A GFP transcript was amplified with the pd2EGFP-1 vector (Clontech) as template and used as control. The amplified products were then purified using GenElute Gel extraction kit (Sigma) followed by in vitro transcription using the MegaScript kit (Ambion). The dsRNA was purified with the Trizol LS reagent (Invitrogen). The Hpt cells were divided into three groups with four replicates in each group. The Hpt cells received different treatments as follows: group 1: GFP dsRNA plus UV-killed WSSV, group 2: GFP dsRNA plus WSSV and group 3: PlDscam dsRNA plus WSSV. The dsRNA transfection and WSSV infection into Hpt cell cultures were performed as described by Liu et al.[29]. Briefly, 4 µl of dsRNA (250 ng/µl) was mixed with 3 µl of histone H2A (1 mg/ml) and with 20 µl of modified L15 and added to one well of 1-day-old Hpt cell cultures. The cells were then incubated at 16°C. At day 3, one replicate of group 2 and 3 were subjected to RNA extraction to determine RNAi efficiencies. For remaining replicates, the medium was replaced with 150 µl of L15 medium together with 5 µl of WSSV stock suspension and 5 µl crude astakine preparation and were further incubated for 36 h at 20°C followed by isolation of RNA. Total HPT RNA was extracted to determine PlDscam and WSSV VP28 transcripts by semi-quantitative RT-PCR. PCR was performed with three oligonucleotide primers (WSSV VP28 (Genbank accession no. AF502435):-F: 5′-TCACTCTTTCGGTCGTGTCG-3′ and -R: 5′- CCACACACAAAGGTGCCAAC-3′; previous primer for 40s ribosomal and PlDscam gene). The PCR conditions were as follows: 94°C 2 min, followed by 30 cycles of 94°C 20 s, 60°C 20 s, and 72°C 30 s for PlDscam and VP28, while 25 cycles for 40S ribosomal gene. The relative expression levels of different time groups were examined by One-way ANOVA followed by Duncan's new multiple range test and Tukey test. Differences were considered statistically significant at P<0.05. Results are expressed as the mean ± SE.

10.1371/journal.ppat.1005075

Nanoformulations of Rilpivirine for Topical Pericoital and Systemic Coitus-Independent Administration Efficiently Prevent HIV Transmission

Vaginal HIV transmission accounts for the majority of new infections worldwide. Currently, multiple efforts to prevent HIV transmission are based on pre-exposure prophylaxis with various antiretroviral drugs. Here, we describe two novel nanoformulations of the reverse transcriptase inhibitor rilpivirine for pericoital and coitus-independent HIV prevention. Topically applied rilpivirine, encapsulated in PLGA nanoparticles, was delivered in a thermosensitive gel, which becomes solid at body temperature. PLGA nanoparticles with encapsulated rilpivirine coated the reproductive tract and offered significant protection to BLT humanized mice from a vaginal high-dose HIV-1 challenge. A different nanosuspension of crystalline rilpivirine (RPV LA), administered intramuscularly, protected BLT mice from a single vaginal high-dose HIV-1 challenge one week after drug administration. Using transmitted/founder viruses, which were previously shown to establish de novo infection in humans, we demonstrated that RPV LA offers significant protection from two consecutive high-dose HIV-1 challenges one and four weeks after drug administration. In this experiment, we also showed that, in certain cases, even in the presence of drug, HIV infection could occur without overt or detectable systemic replication until levels of drug were reduced. We also showed that infection in the presence of drug can result in acquisition of multiple viruses after subsequent exposures. These observations have important implications for the implementation of long-acting antiretroviral formulations for HIV prevention. They provide first evidence that occult infections can occur, despite the presence of sustained levels of antiretroviral drugs. Together, our results demonstrate that topically- or systemically administered rilpivirine offers significant coitus-dependent or coitus-independent protection from HIV infection.

When taken consistently, PrEP has been shown to reduce the risk of HIV infection by up to 92% in people who are at high risk. However, PrEP is much less effective if it is not taken consistently. To improve adherence to the drug regimen, several new drug delivery systems, that include novel gel formulations and long-acting delivery systems, are being evaluated. In this manuscript, we used BLT humanized mice, an in vivo model of vaginal HIV transmission, to evaluate two novel delivery systems for HIV prevention. In the first approach, we combined the highly efficient encapsulation of antiretroviral drugs into nanoparticles with a thermosensitive gel that remains liquid at room temperature and solidifies at body temperature. Our results showed that this delivery system provided significant protection from HIV vaginal infection. In a second approach, we evaluated a long-acting nanoparticle formulation for coitus-independent protection from HIV acquisition. Our results showed that a single injection of the long-acting antiviral drug also resulted in reduced HIV infection. However, protection was not complete and transmission was concealed by a significant delay in the onset of plasma viremia that could result in superinfection by two different viruses administered up to four weeks apart.

Although the annual number of new HIV infections continues to decline, the global HIV-1 pandemic remains an unprecedented public health problem, with 2.1 million new infections in 2013 and an estimated 35 million people already infected [1]. This highlights the urgent need for effective and safe prevention strategies for HIV infection. With the continued absence of an effective vaccine, the efficacy of various antiretrovirals (ARVs) has been evaluated as pre-exposure prophylaxis (PrEP). HIV PrEP refers to the strategy of using ARV drugs to decrease the risk of HIV infection in uninfected individuals who are at high risk of infection. Multiple clinical trials, including the CAPRISA 004, the Chemoprophylaxis for HIV Prevention in Men (iPrEx), the Partners PrEP, and the TDF2 studies have shown that topical or oral pre-exposure administration of ARVs reduces the risk of HIV-1 infection by 39 to 75% [2–5]. Overall efficacy, as well as low rates of protection in some trials, correlates with adherence to the dosing regimen [3]. To improve adherence and PrEP efficacy, several strategies are being considered. These include effective antiretrovirals, easily administered in single topical or systemic dose pericoitally, and long-acting ARV formulations that release drugs over many weeks systemically, requiring infrequent parenteral administration [6–8]. Rilpivirine (RPV, TMC278), a non-nucleoside reverse transcriptase inhibitor (NNRTI), is a diarylpyrimidine derivative that inhibits HIV reverse transcriptase by binding to a hydrophobic pocket near the active site of the enzyme, and consequently preventing transcription of viral RNA. RPV has activity against wild type and many NNRTI-resistant HIV-1 strains [9]. Although RPV has an excellent profile for HIV prevention, there is currently no information regarding the effectiveness of oral RPV for HIV prevention, and no RPV formulations for topical use have been described. Recently, a long-acting crystalline nanoparticle suspension of RPV (RPV LA) has been developed, with the objective of providing drug exposure over extended periods of time following intramuscular administration [10]. A single intramuscular injection of RPV LA provided sustained release of RPV into plasma over 3 months in dogs, 2 months in rats, and 3 weeks in mice [10, 11]. In humans, a single intramuscular administration of RPV LA leads to substantial levels of RPV in plasma, cervico-vaginal fluid and vaginal tissue for 84 days. RPV levels measured at multiple sites of HIV transmission suggest a potential role for RPV LA as coitus-independent PrEP in humans [10, 12]. Animal models are essential to the effective evaluation of new HIV prevention strategies. For example, rhesus macaques (Macaca mulatta) and pigtail macaques (Macaca nemestrina) were recently used to test whether a long acting formulation of an integrase inhibitor could prevent transmission of HIV via rectal or vaginal routes [13–15]. However, the species-specific tropism of HIV prevents the evaluation of relevant viruses, including transmitted/founder viruses, for in vivo challenges in these models [16, 17]. Instead, chimeric simian/human immunodeficiency viruses (SHIVs) must be used. Here we tested the efficacy of a topical pericoital and a long-acting systemic nanoformulation of RPV to prevent vaginal HIV-1 transmission using humanized bone marrow/liver/thymus mice (BLT). BLT mice are immunodeficient mice individually bioengineered to express a de novo-generated human immune system distributed throughout each animal [18–21], allowing infection with a variety of transmitted/founder HIV-1 isolates via relevant routes of transmission. The mouse female reproductive tract (FRT) has anatomic similarities to that of humans, despite its smaller size and presence of two uterine horns that merge to form the main body of uterus. The murine vagina and ectocervix are covered with stratified squamous epithelium, whereas the endocervix and uterus consist of a simple columnar epithelium. The physical barrier that HIV would encounter is, therefore, somewhat similar to that in humans [22]. We previously demonstrated the presence of human CD4+ T cells, macrophages and dendritic cells throughout the mouse female reproductive tract (FRT), that render BLT mice susceptible to vaginal HIV-1 transmission [23, 24]. Both topical and systemic HIV prevention interventions, which parallel human clinical trials, have been successfully performed in BLT mice [23, 25–28]. These studies validate BLT mice as a suitable model for the evaluation of novel or improved drug formulations for the prevention of HIV transmission. We used humanized BLT mice to test two different strategies to prevent vaginal HIV transmission by RPV. Each strategy used a distinct nanotechnology formulation of RPV. RPV gel, a potential pericoital microbicide (developed at Creighton University, Omaha, Nebraska USA), was applied to the vaginal mucosa in single doses 1.5h or 24h before HIV challenge (Fig 1A). In other studies, a coitus-independent, systemic long-acting formulation of rilpivirine RPV LA (developed by Janssen Research and Development, Beerse, Belgium), was administered intramuscularly, 1 week before HIV-1 challenge (Fig 1B). The presence of plasma viral RNA was monitored over time as an early indication of infection. At the end of these experiments, multiple tissues were analyzed for the presence of cell-associated viral DNA. Only animals treated with RPV nanoformulations, and negative for both viral RNA and DNA in peripheral blood and tissues, were considered protected from vaginal HIV-1 transmission. The majority of clinical trials evaluating pericoital HIV prevention approaches, including CAPRISA 004, used conventional gels for vaginal microbicide delivery. However, such gels have major disadvantages, including gel leakage, uneven distribution, and messiness, which can decrease adherence to the dosing regimen [29]. To overcome these potential drawbacks, we formulated poly(lactic-co-glycolic acid) (PLGA) nanoparticles loaded with RPV (PLGA/RPV NP), in a thermosensitive gel which is liquid at room temperature but highly viscous at body temperature. This property minimizes chances of gel leakage. Moreover, as reported previously, thermosensitive pluronic gels are insensitive to dilution by simulated vaginal fluid [30–33]. PLGA NPs are US FDA-approved biodegradable particles, which can encapsulate ARV and provide their sustained release [34–38]. PLGA/RPV NPs were prepared by emulsion-solvent evaporation. Average particle size was 66.0 ± 4.2nm (mean ± SEM, n = 3) measured by dynamic light scattering, which was further validated using Scanning Electron Microscopy (Fig 2A). The average polydispersity index was 0.14 ± 0.05, zeta potential of the NPs was -10.96 ± 1.4mV (mean ± SEM, n = 3). The RPV encapsulation efficiency in the polymeric nanoparticle, determined by an indirect method as described in the Materials and Methods section, was 98 ± 0.7% and the RPV loading in nanoparticles was ~ 5% w/w of polymer. As shown in Fig 2B, the intracellular uptake of RPV by HeLa cells cultured in presence of a 5μg/ml RPV solution or in the presence of 5μg/ml RPV in PLGA/RPV NPs, was comparable. In TZM-bl indicator cells, PLGA/RPV NPs showed similar in vitro inhibition of HIV-1 infection as RPV in solution (Fig 2C). Together, these data confirmed the potential of PLGA/RPV NPs to effectively deliver ARV to target cells. For in vivo evaluation in BLT humanized mice, PLGA/RPV NPs were formulated into a previously characterized thermosensitive gel containing Pluronic F127 and Pluronic F68 [32]. The concentration of RPV in thermosensitive gel was 0.876mg/ml. In humans, cervicovaginal mucus is a significant barrier and clearance mechanism that limits vaginal drug delivery and retention. To achieve sustained drug release and maintain protective drug concentrations during pre-exposure prophylaxis, drug-loaded nanoparticles need to penetrate cervicovaginal mucus and be efficiently distributed across the female reproductive tract [39]. The ability of PLGA/RPV NPs in thermosensitive gel to distribute in the FRT of humanize mice was evaluated using rhodamine-labeled PLGA NPs. Rhodamine-labeled PLGA NPs in thermosensitive gel were instilled into the mouse vagina and their presence and localization was assessed by confocal microscopy. Ninety minutes after vaginal administration, the fluorescence signal was seen as a continuous layer at the luminal site of the vaginal epithelium. Interestingly, some fluorescence, although with much lower intensity, was still found on the vaginal epithelium 24h after administration. Fluorescence signals were also observed deeper in the tissue in close proximity of hCD45, hCD3, hCD4, hCD8, and hCD11c cells (Fig 3A and 3B). Given the stability of Rhodamine encapsulation in this type of nanoparticles [40], these results demonstrate that PLGA nanoparticles, delivered in thermosensitive gel, can reach the vaginal epithelium and the location of HIV target cells, and persist for 24h. To test the effectiveness of PLGA/RPV NPs in preventing HIV-1 vaginal transmission, humanized BLT mice were topically treated with PLGA/RPV NPs in thermosensitive gel (20μl of gel, 17.5μg of RPV per mouse; n = 12), thermosensitive gel containing PLGA NP without RPV, or vehicle (n = 8). Treated animals were challenged 1.5h (n = 4) or 24h (n = 8) after gel application with a high dose of HIV-1RHPA, a CCR5-tropic transmitted/founder virus (3.1×105 TCID) [16]. The presence of plasma viral RNA in peripheral blood was determined at intervals thereafter. No plasma viral RNA was found in the peripheral blood of any of the animals challenged with HIV-1 1.5h after the administration of PLGA/RPV NPs in thermosensitive gel (0/4) nor in 4/8 of the animals that received PLGA/RPV NPs in thermosensitive gel 24h prior to exposure to HIV-1 (p = 0.0084 and p = 0.0582 respectively). Seven to eight weeks post-exposure, cells isolated from multiple organs were analyzed for the presence of viral DNA. The lack of detectable cell-associated viral DNA in tissues confirmed the absence of HIV infection in these animals (Fig 3C–3E, S1 Table). The presence of plasma viral RNA and cell-associated viral DNA in peripheral blood and tissues of all the control mice confirmed efficient HIV transmission (Fig 3C and 3D). The protection of all BLT mice challenged with HIV-1 1.5h after treatment with PLGA/RPV NPs in thermosensitive gel, and the protection of 50% of the mice treated 24h prior to challenge with HIV-1, demonstrated the effectiveness of PLGA/RPV NPs in HIV prevention. To evaluate the efficacy of a coitus-independent systemic RPV LA formulation to prevent HIV-1 infection, we first determined the plasma RPV concentrations in mice for 28 days after intramuscular injections of either 15mg (50μl of 300mg/ml RPV LA nanosuspension) or 7.5mg (25μl of 300mg/ml RPV LA nanosuspension) of RPV LA (n = 4 per group). As shown in Fig 4A, high levels of RPV were detected in the plasma of all treated animals 24h post-injection. Plasma drug levels decreased rapidly over the following 4 days. After this point, the levels of RPV in plasma remained relatively constant for 10 days and then decreased gradually over the next two weeks. Throughout the 28 days of the study, RPV plasma levels exceeded the protein-adjusted IC90 of 12ng/ml [41]. These results demonstrated that a single intramuscular injection of RPV LA resulted in sustained levels of drug in plasma in mice and established its potential to serve as a prevention strategy for intermittent use. BLT mice (n = 12) were injected intramuscularly with 15mg of RPV LA (n = 6), vehicle or left untreated (n = 6). One week later, mice were challenged vaginally with a high dose (3.5×105 TCID) of HIV-1CH040, a transmitted/founder virus [17]. Viral RNA in plasma was evaluated over the following 8 weeks. Five out of six control mice were infected within 2 weeks after challenge, as evidenced by the presence of viral RNA in plasma. In sharp contrast, no viral RNA was detected in the plasma of the animals treated with RPV LA (Fig 4B, S2 Table). Analysis of tissue DNA of RPV LA-treated mice demonstrated the absence of viral DNA in all samples analyzed (S2 Table, Fig 4C). These results demonstrated that a single administration of RPV LA 7 days prior to challenge offered significant protection from vaginal HIV-1CH040 infection. It should be noted that longitudinal flow cytometry analysis of the mouse peripheral blood confirmed that the absence of viral RNA and DNA in the mice treated with RPV LA and exposed to HIV-1CHO40 was not due to a loss or to reduced levels of human CD45+ cells or human CD3+CD4+ cells throughout the course of the experiment. Only in the infected mice were we able to demonstrate a gradual decrease in the levels of human CD3+CD4+ cells (Fig 4D). The overall experimental approach to evaluate the ability of RPV LA to prevent vaginal HIV transmission after two high dose challenges is shown in Fig 5A. BLT mice received a single 15mg intramuscular injection of RPV LA (n = 10) or vehicle (n = 4). Mice were challenged one week later with a high dose of either HIV-1JR-CSF [an early passage CCR5-tropic primary isolate] (n = 3, TCID 3.5×105), HIV-1CH040 (n = 4, TCID 3.5×105) or HIV-1RHPA (n = 3, TCID 3.1×105) transmitted/founder CCR5-tropic viruses. Control mice were challenged with HIV-1CH040 or HIV-1RHPA (n = 2 each). Plasma viral RNA was monitored over time. Viral RNA was detected in 4/4 of the control mice. Two weeks post-challenge, no viral RNA was detected in any of the animals that received RPV LA (Fig 5b, S2 Table). Four weeks after RPV LA administration (3 weeks after the first challenge), mice were challenged vaginally with HIV-1THRO, a different CCR5-tropic transmitted/founder virus (TCID 3.5×105), in parallel with 6 additional control (no drug) BLT mice (also challenged with HIV-1THRO). Mice were monitored for the presence of plasma viral RNA for an additional 5 weeks. Seven of ten RPV LA-treated mice became infected within 4 weeks after the second HIV-1 challenge. In order to identify the virus that resulted in the infection of the RPV LA-treated mice, plasma viral RNA from each infected mouse was sequenced. Sequence analysis revealed that, despite the fact that no viral load was detected in plasma for over 2 weeks after the first challenge, 2 of these mice had actually acquired infection from the first challenge. A third mouse acquired infection after both challenges (dually infected mouse). Sequence analysis of the other 4 mice showed that they were only infected with the second challenge virus (Fig 5B, S1 Fig, S3 Table). No mutations associated with RPV resistance were found in the virus present in any of the RPV LA-treated and infected mice. Analysis of DNA from tissues of the three remaining uninfected mice treated with RPV LA revealed the absence of viral DNA in all tissues analyzed and confirmed their protection from HIV transmission after two high dose challenges (Fig 5C). In summary, during the dual challenge experiment, 7/10 RPV LA-treated animals were protected from the first challenge and 4/9 from the second challenge. These results demonstrate that RPV LA offered significant (>80%; p<0.0001) protection from a high dose of virus administered one week later, and partial (44%; p = 0.0038) protection from a second high dose HIV challenge 4 weeks after drug administration. There have been numerous attempts to prevent HIV infection with the topical application of microbicides [42]. Topical PrEP is based on the premise that blocking HIV at the site of entry offers the best opportunity to prevent HIV infection and avert systemic toxicity. Topical application of non-specific HIV inhibitors has failed to show protection against HIV infection in several large clinical trials [43]. In contrast, the first clinical trial, evaluating the potential of tenofovir in a vaginal gel formulation (CAPRISA 004), demonstrated significant protection [2]. Subsequent trials, using topically applied tenofovir, were not able to demonstrate protection, most likely due to a lack of product use [44, 45]. Here, we evaluated the efficacy of RPV formulated in PLGA nanoparticles suspended in a thermosensitive gel that remains liquid at room temperature but solidifies at body temperature. We evaluated the distribution of nanoparticles in the vagina of BLT mouse using PLGA nanoparticles with encapsulated rhodamine as a surrogate for rilpivirine. We chose rhodamine because it had been previously shown to be released slowly from the PLGA nanoparticles, with initial burst release seen in the first several hours followed by a more sustained, uniform release [40]. Our results showed that rhodamine-PLGA NPs formed a layer on the lumen of the vaginal epithelium. Some of the particles persisted at the vaginal epithelium 24h post-administration. However, PLGA NPs also infiltrated into the tissue, as fluorescence signal was also found in close proximity to HIV target cells. BLT mice, vaginally treated with PLGA/RPV NPs in the thermosensitive gel, were protected against a high dose challenge with a transmitted/founder virus administered 1.5h later. Protection, diminished (by 50%) when BLT mice were challenged 24h after gel administration. These results demonstrated that topical administration of this novel PLGA/RPV NP thermosensitive gel formulation efficiently prevented vaginal HIV transmission in this animal model. An alternative to topical PrEP is the systemic administration of antiretrovirals for the prevention of HIV acquisition. The results from multiple clinical trials (iPrEx, TDF2 and Partners PrEP study) demonstrate that systemic PrEP can prevent HIV acquisition via rectal and vaginal exposure [3–5]. The results from virtually all clinical trials of HIV prevention using antiretrovirals demonstrate that efficacy depends on product usage. Long-acting injectable formulations of antiretrovirals are one of several approaches that are being considered to enhance adherence to preventive measures. GSK744 (Cabotegravir) LA is an injectable nanosuspension formulation of an integrase inhibitor that has been shown to be safe and to sustain adequate levels of drug when administered intermittently. Recent results from investigations in non-human primates also demonstrate that GSK744 LA can efficiently prevent infection after repeated low dose viral challenges [13–15]. However, it should be noted that this formulation has demonstrated lower efficacy at preventing infection after a high dose challenge [14]. RPV LA is a long-acting injectable nanosuspension formulation of the NNRTI rilpivirine. Sustained levels of RPV in plasma after single intramuscular injection of RPV LA were reported in dogs, rats and mice, and in plasma, cervico-vaginal fluid and vaginal tissue of humans [10, 12]. However, an assessment of RPV LA in preventing HIV-1 vaginal transmission in a relevant animal model has not yet been reported. Here, we showed that a single administration of RPV LA in BLT mice conferred significant protection against a high dose vaginal challenge with HIV-1 one week after drug administration. Our results also demonstrated that a single administration of RPV LA offered protection, albeit reduced, from a second virus challenge 1 month after drug administration. Importantly, while all our control mice were infected within 2 weeks after challenge, breakthrough infections in treated animals were delayed and occurred 3 weeks after first challenge (4 weeks after drug administration) or 2–4 weeks after second challenge (6–8 weeks after drug administration). These results are consistent with a model in which the initial infection occurs at the site of exposure, but is contained by the presence of sustained levels of drug, preventing systemic replication. Once the levels of drug are unable to efficiently inhibit virus replication, viral spread can occur. Due to faster clearance in mice, a sustained level of RPV in plasma after single injection of RPV LA lasts significantly longer in humans than mice (3 months in humans vs. 3 weeks in mice, [10, 11]). Therefore, it is reasonable to anticipate a longer protective effect of RPV LA in humans. In summary, our results in humanized BLT mice highlight the potential of RPV as a candidate for HIV pre-exposure prophylaxis in both a topical coitus-dependent thermosensitive gel formulation as well as in a long acting injectable nanosuspension. Resomer 752 H (acid terminated poly-lactic-co-glycolic acid; Avg. Mol. Wt. 15000Da) was purchased from Sigma Chemicals (St. Louis, MO, USA). Rilpivirine (RPV) was purchased from Sequoia Research Ltd. (Pangbourne, UK). Potassium dihydrogen phosphate (HPLC grade), acetonitrile (HPLC grade), dimethyl sulfoxide (DMSO, AR Grade), ethyl acetate (AR grade), citric acid (AR grade) and trisodium citrate (AR grade) were purchased from Fisher Scientific Ltd (NJ, USA). Pluronic F127 and Pluronic F68 (BASF, NJ, USA) were received as gift samples. The ultra-pure water was obtained for all the experiments with the use of PURELAB Ultra system (Elga LLC, IL, USA). RPV-loaded PLGA nanoparticles were prepared using a previously described emulsion-solvent evaporation method [46]. Briefly, Resomer 752H (200mg) and Pluronic F127 (200mg) were dissolved in 3ml ethyl acetate by heating at 40°C in an incubating shaker bath. RPV (10mg) was dissolved in DMSO (50μl) by heating at 40°C in an incubating shaker bath and then transferred to the Resomer 752H solution to obtain a homogenous organic phase. The organic phase was added drop-wise to 10ml ultrapure water and homogenized using a probe sonicator (UP100H; Hielscher USA, Inc., NJ, USA). The resultant oil-in-water emulsion was stirred for 4h using a magnetic stirrer. Due to the photosensitive nature of RPV, the contents in the beaker were protected from light during the evaporation of ethyl acetate. The mean particle size, polydispersity index and zeta potential of the PLGA/RPV nanoparticles were measured in triplicate using dynamic light scattering at an angle of 90° at 25°C, by a “ZetaPlus” Zeta Potential Analyzer (Brookhaven Instruments Corp, NY, USA). For fabricating rhodamine-6G-labeled fluorescent PLGA nanoparticles, RPV was replaced with rhodamine-6G (1mg) and nanoparticles were formulated as described earlier. A reverse phase-HPLC method was developed and validated for determination of RPV from various matrices derived in the topical gel studies. The HPLC apparatus (Shimadzu Corporation, Columbia, MD) consisted of a pump (LC-20AB), system controller (CBM-20A), degasser unit (DGU-20A), refrigerated auto-sampler (SIL-20AC), a UV-Vis detector (SPD-20A) and a column heater (CTO-20A). Samples were run through a C18 pre-column and a Gemini C18 reverse-phase [150mm × 4.5mm (I.D.)] with 5μm particle size packing (Phenomenex, Torrance, CA). The mobile phase consisted of acetonitrile and 25mM KH2PO4 solution (50:50). For HPLC analysis, the flow rate of the mobile phase was at 0.6ml/min, column oven was set at 35°C, injection volume was 20μl and the analysis was carried out at 290nm. The retention time for the RPV was 12.9min. For standard curve, RPV stock solution (1mg/ml) was prepared in methanol. The stock solution was diluted with acetonitrile to obtain solutions of various concentrations. A standard curve was obtained by injecting 0.025–2μg/ml of RPV. The limit of detection for RPV was 8ng/ml. Intra-day and inter-day variability of the analytical method was <10%. A thermosensitive vaginal gel containing PLGA/RPV-NPs was prepared as described earlier [32]. Briefly, the pH of RPV NPs was adjusted to 4.5 with citric acid and sodium citrate. Glycerol (0.225g) was added to PLGA/RPV NPs (10ml) to adjust the osmolarity of nanoparticles. PLGA/RPV NPs were transferred to a screw-capped bottle and Pluronic F127 (2g) and Pluronic F68 (100mg) were added to PLGA/RPV-NPs with intermittent stirring. The screw-capped bottle containing PLGA/RPV NPs and Pluronics was stored overnight in the refrigerator to dissolve Pluronics. On the next day, the dispersions were gently stirred to obtain a homogenous translucent solution. The solution was observed for signs of nanoparticle aggregation and/or phase separation. Thermosensitivity of the gel was confirmed by incubating the gel in a 37°C water bath. The preparation of thermosensitive gel containing Rhodamine-6G-labeled fluorescent PLGA nanoparticles was carried out in a similar manner. Human cervical (HeLa) cells were purchased from the American Type Culture Collection (ATCC, Manassas, VA), TZM-bl cells were procured through the NIH AIDS Research and Reference Reagent Program. HeLa and TZM-bl cells were maintained in complete Dulbecco’s Modified Eagle’s Media ((DMEM, MediaTech Inc., Manassas, VA) supplemented with 10% fetal bovine serum (FBS, Hyclone Inc., Utah), 4mM L-glutamine, 100U/ml penicillin and 100μg/ml streptomycin (MP Biomedical Inc., Solon, OH) and maintained in a logarithmic growth phase. All cells were grown at 37°C and 5% CO2. TZM-bl HIV indicator cells were seeded in 24-well plates at a density of 2 x 105 per well. After 24h, the cells were treated with different concentrations of PLGA/RPV NPs and RPV solution (concentration range: 10μg/ml to 100pg/ml) for 24h. Cells were washed and cultured for 24h in fresh complete DMEM. Cells were infected with HIV-1 NL4-3 virus (25μl) for 4h and incubated for an additional 48h. One-Glo reagent (Promega, Madison, WI), supplemented with Triton X-100 (final concentration 0.01%) was added to inactivate virus and to allow for the measurement of luciferase activity. Results were normalized to the luciferase activity of cells infected with virus incubated with plain RPMI medium. RPV LA was prepared as previously described [11]. Briefly, a sterile isotonic nanosuspension, consisting of rilpivirine particles, was prepared by wet nanomilling of the rilpivirine base, surfactant, and buffer to ensure neutral pH under aseptic conditions. The median particle size was 200nm. Poloxamer 338 (Pluronics F108), a hydrophilic, nonionic surfactant, was used to enhance solubility and stabilize the colloidal suspension against aggregation. The final drug concentration was 300mg/ml. Plasma was isolated from 0.05–0.1ml peripheral blood samples on EDTA collected from mouse retro-orbital venous sinus and stored at -80°C until analysis. Plasma samples were analyzed individually for unchanged rilpivirine by liquid chromatography-tandem mass spectrometry (LC/MS-MS) as described previously [11]. BLT mice were generated as described previously [19, 23, 26, 27, 47, 48]. Briefly, a 1–2mm piece of human fetal liver tissue was sandwiched between two pieces of autologous fetal thymus tissue (Advanced Bioscience Resources, Alameda, CA) under the kidney capsule of sublethally irradiated (0.250Sv) 6–8 wk old NOD.Cg- Prkdcscid Il2rgtm1Wjl/SzJ mice (NSG; The Jackson Laboratory, Bar Harbor, ME). Following implantation, mice were transplanted intravenously with hematopoietic CD34+ stem cells isolated from autologous human fetal liver tissue. Human immune cell reconstitution was monitored by flow cytometrical analysis of the peripheral blood every 2 weeks, as previously described [23, 26, 27, 48]. At the end of experiments, mice were euthanized by exposure to avertin followed by euthanasia. Mice were maintained at the Division of Laboratory Animal Medicine, University of North Carolina at Chapel Hill (UNC-CH). All animal experiments were carried out in accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the Institutional Animal Care and Use Committee guidelines of the University of North Carolina (protocol number:12–170). Female humanized-BLT mice were anesthetized with Nembutal. Rhodamine-labeled NPs or control nanoparticles (without rhodamine) in thermosensitive gel (20μl) was instilled into the mouse vagina. 1.5h or 24h later, mice were sacrificed and the FRT harvested, fixed in 4% paraformaldehyde solution (SafeFix, Fisher Science) and embedded in OCT compound (Sakura). Mouse vaginal frozen sections (5μm) were stained with monoclonal antibody for hCD45 (Dako, mouse IgG1), hCD3 (Thermo Scientific, rabbit IgG), hCD4 (GenWay, Rabbit IgG), hCD8 (Dako, Mouse IgG1), hCD11c (Leica, mouse IgG2a), mouse IgG1 (Dako), mouse IgG2a (Dako) or rabbit Ig (Dako) after blocking with Background Sniper (Biocare Medical). The sections were then stained with either DyLight 488-conjugated donkey anti-mouse IgG or DyLight 488-conjugated donkey anti-rabbit IgG (Jackson Immunoresearch). All sections were finally counterstained with DAPI (Sigma) and analyzed by confocal microscopy (TCS SP2, Leica). For topical administration of PLGA/RPV NP, BLT mice were administered intravaginally with 17.5μg RPV in the form of PLGA/RPV NPs in 20μl thermosensitive gel containing PLGA/RPV NPs. 1.5h or 24h later, the animals were anesthetized with Nembutal and challenged with 4.5×105 TCID HIV transmission/founder virus HIVRHPA. Control BLT mice (n = 4) received vehicle or thermosensitive gel with blank nanoparticles, and were challenged with the same transmission/founder virus. For systemic administration of RPV LA, female BLT mice received single injection of 15 mg of nanosuspension intramuscularly. One week later, mice were anesthetized with Nembutal and intravaginally challenged with transmission/founder viruses (HIVCHO40 3.0×105 TCID or HIVRHPA 4.5×105 TCID) or HIVJR-CSF (7.0×105 TCID). 3 weeks later, uninfected mice were challenged vaginally with transmission/founder HIVTHRO (4.0×105 TCID). Viral stocks were generated by transfecting proviral DNA into 293T cells using Lipofectamine 2000 (Invitrogen) and tissue culture infectious units (TCID) were determined using TZM-bl cells, essentially as we have previously reported [49, 50]. HIV-1 JR-CSF, CHO40, THRO and RHPA were obtained from Dr. Irving Chen and John Kappes via the AIDS Research and Reagent Repository Program. Infection of BLT mice with HIV-1 was monitored in peripheral blood by determining levels of viral RNA in plasma by one-step real-time reverse transcriptase PCR assay, using the following primers: CATGTTTTCAGCATTATCAGAAGGA, TGCTTGATGTCCCCCCACT, and the MGB-probe carboxyfluorescein (FAM)-CCACCCCACAAGATTTAAACACCATGCTAA-Q (nonfluorescent quencher) (Applied Biosystems) (sensitivity of 400 HIV RNA copies/ml). The percentage of human CD4+ T cells in peripheral blood of BLT mice before challenge (0–2 weeks prior to exposure) and after challenge was determined by flow cytometry with respective antibodies: hCD45-APC, hCD3-FITC, hCD4-PE and hCD8-PerCP (eBioscience). Flow cytometry data were collected using a BD FACSCanto cytometer and analyzed using BD FACSDiva software. The presence of viral DNA in tissues and peripheral blood collected from BLT mice was determined by real-time PCR analysis of DNA extracted from 5×104–4×106 cells from harvested tissue (spleen, lymph nodes, bone marrow, liver, lung, female reproductive tract) or from 15–50μl peripheral blood cells, as previously described [23, 26, 27, 51]; (assay sensitivity of 10 DNA copies per sample). Viruses replicating in infected animals were identified by sequence analysis. Viral RNA was isolated from plasma using QIAamp viral RNA columns (Qiagen) according to the manufacturer’s protocol, and cDNA was generated using Superscript III Reverse Transcriptase (Invitrogen) with the primer GTGGGTACACAGGCATGTGTGG. cDNA was amplified by nested PCR using the Expand High Fidelity PCR System (Roche). PCR primers were designed to anneal in regions with the fewest possible primer mismatches to HIVJR-CSF, HIVCH040, HIVRHPA and HIVTHRO sequences. Primer sequences were as follows: outer forward primer, TGCATATTGTGAGTCTGTTACTATGTTTACT; reverse prime CAGGAGCAGATGATACAG; inner forward primer, GTAGGACCTACACCTGTCAAC; reverse primer CCTGCAAAGCTAGGTGAATTGC. Amplified viral DNA was sequenced and compared to sequences of transmitted/founder viruses. Drug concentrations in plasma over time were compared using Tukey’s multiple comparison test. Statistical differences between treated and control animals in the efficacy of tested nanoformulations in protection protecting from vaginal HIV-1 transmission were determined by log-rank/ Mantel–Cox test. All statistical analyses were performed using GraphPad Prism software (version 6).

10.1371/journal.ppat.1002361

Signal Transduction through CsrRS Confers an Invasive Phenotype in Group A Streptococcus

The CsrRS (or CovRS) two component system controls expression of up to 15% of the genome of group A Streptococcus (GAS). While some studies have suggested that the sensor histidine kinase CsrS responds to membrane perturbations as a result of various environmental stresses, other data have implicated the human antimicrobial peptide LL-37 and extracellular Mg2+ as specific signals. We now report that Mg2+ and LL-37 have opposite effects on expression of multiple genes that are activated or repressed by the transcriptional regulator CsrR. Using a GAS isolate representative of the recently emerged and widely disseminated M1T1 clone implicated in severe invasive disease, we found marked up-regulation by CsrRS of multiple virulence factors including pyrogenic exotoxin A, DNase Sda1, streptolysin O, and the hyaluronic acid capsular polysaccharide, among others. Topology and surface protein labeling studies indicated that CsrS is associated with the bacterial cell membrane and has a surface-exposed extracellular domain accessible to environmental ligands. Replacement of a cluster of three acidic amino acids with uncharged residues in the extracellular domain of CsrS abrogated LL-37 signaling and conferred a hyporesponsive phenotype consistent with tonic activation of CsrS autokinase activity, an effect that could be overridden by mutation of the CsrS active site histidine. Both loss- and gain-of-function mutations of a conserved site in the receiver domain of CsrR established an essential role for lysine 102 in CsrS-to-CsrR signal transduction. These results provide strong evidence that Mg2+ and LL-37 are specific signals that function by altering CsrS autokinase activity and downstream phosphotransfer to CsrR to modulate its activity as a transcriptional regulator. The representation of multiple antiphagocytic and cytotoxic factors in the CsrRS regulon together with results of in vitro phagocytic killing assays support the hypothesis that CsrRS mediates conversion of GAS from a colonizing to an invasive phenotype in response to signaling by host LL-37.

Group A Streptococcus (S. pyogenes or GAS) is exclusively a human pathogen that can inhabit the human throat as a harmless commensal, cause localized, self-limited infection in the form of pharyngitis or strep throat, or invade local tissues or the bloodstream to produce life-threatening disease states such as necrotizing fasciitis or streptococcal toxic shock. We present evidence that the GAS CsrRS (or CovRS) two component system governs the transition from a colonizing to an invasive phenotype by transducing a specific signal from the antimicrobial peptide LL-37 that is secreted as part of the human innate immune response to GAS infection. We show that LL-37 signaling requires specific domains of both the CsrS sensor kinase and the CsrR response regulator, and that signaling results in a coordinated and marked increase in expression of multiple bacterial factors that confer resistance to phagocytic killing, a hallmark of GAS virulence.

Human beings are thought to be the principal if not exclusive host for group A Streptococcus (S. pyogenes, GAS). The organism's primary environmental niche is the human pharynx where GAS can colonize the epithelium without evoking any clinical symptoms, or it can produce local inflammation and symptomatic streptococcal pharyngitis [1], [2]. GAS also causes impetigo, a superficial skin infection, and, less commonly, severe invasive infections such as necrotizing fasciitis, bacteremia, and streptococcal toxic shock [3], [4]. The regulated expression of a variety of gene products enhances GAS survival in the human host through a dynamic process of adaptation to stresses that may change depending on the precise anatomic location of the bacteria in the body, environmental factors, and engagement of host defense mechanisms [5], [6]. Two component regulatory systems (TCS) play an important role in such dynamic adaptation of many bacteria to changing environmental conditions [7], [8]. CsrRS (also called CovRS) is the most extensively characterized TCS in GAS. First identified as a regulator of the has operon that encodes the enzymes required for synthesis of the hyaluronic acid capsular polysaccharide, CsrRS has since been shown to affect expression of as much as 15% of the GAS genome including genes encoding many virulence factors [9]–[12]. Genetic evidence and similarity to TCS in other species have suggested that CsrS is a sensor histidine kinase whose phosphorylation state is influenced by environmental signals, while CsrR is a transcriptional regulator whose activity at target promoters is controlled by phosphorylation. It is presumed, but not proven, that phosphorylation of CsrR results from phosphotransfer from CsrS. It has also been proposed that CsrS has a phosphatase activity and can dephosphorylate CsrR [13]. Transcriptional profiling of CsrR- or CsrRS-mutants has indicated that CsrR acts primarily, although not exclusively, as a repressor of gene expression, as mutants exhibit increased expression of most CsrRS-regulated genes, and phosphorylation of CsrR in vitro enhances its binding to regulated promoters [9], [10], [14], [15]. While it is clear that CsrRS influences expression of many important GAS products, a unifying explanation of the adaptive role of the CsrRS system is still unproven. One proposal is that CsrRS represents a system to detect and respond to a variety of environmental stresses, such as elevated temperature, acidic pH, and high osmolarity, all of which might result in alterations in physical properties of the bacterial cell membrane and consequent signaling through CsrS [13]. An alternative model is that CsrS recognizes specific ligands, and that interaction of these ligands with its extracellular domain (ECD) results in changes in CsrS autokinase activity and/or phosphatase activity for CsrR. The latter model is based on the findings that increased concentrations of extracellular Mg2+ result in widespread down-regulation of CsrR-repressed genes, an effect dependent on a functional CsrS and not reproduced by other cations [11], [16]. Thus, Mg2+ may serve as a specific stimulus for activation of CsrS kinase activity with downstream phosphorylation of CsrR. The human antimicrobial peptide LL-37 has been shown to have effects on CsrRS signaling opposite to those of elevated Mg2+. Concentrations of LL-37 far below those that inhibit GAS growth were shown to stimulate increased expression of the has operon and three other CsrR-repressed genes in a CsrS-dependent fashion [17]. While these two models are not necessarily mutually exclusive, it is difficult to reconcile LL-37 signaling with a model of non-specific membrane perturbation since the effects of LL-37 on gene expression were not reproduced by a broad range of doses of other antimicrobial peptides, including other cathelicidins, of similar or greater antibacterial potency [17]. The highly specific effect of LL-37 to stimulate up-regulation of CsrR-repressed genes suggests that CsrRS functions to detect and counteract host immune effectors that mediate bacterial clearance from the infected host. Circumstantial evidence for such a role comes from isolation of spontaneous CsrRS mutants in the setting of invasive GAS infection, both in patients with severe invasive GAS infection and in experimental animals [18]–[21]. Exposure of wild type GAS to LL-37 or inactivation of CsrRS by mutation results in increased expression of factors that dramatically enhance GAS resistance to opsonophagocytic killing [12], [17]. These observations suggest that a physiologic role of CsrRS is to detect relatively low concentrations of LL-37 as a signal of mobilization of host defenses including the recruitment of phagocytic leukocytes and to trigger a global transcriptional response that enhances GAS resistance to phagocytosis. We now report the results of further investigation that provides strong support for this hypothesis. LL-37 not only activates expression of the four previously identified loci, but also stimulates either activation or repression of multiple CsrRS-regulated genes. Signaling by LL-37 is dependent on CsrS, which is shown to have a surface-exposed domain on the bacterial cell. Transduction of the LL-37 signal requires specific domains of both CsrS and its cognate regulator CsrR to induce changes in gene expression. A critical consequence of LL-37-mediated CsrRS-signaling is enhanced resistance to phagocytic killing by human blood leukocytes, a bacterial phenotype that is central to both persistence of GAS in the human host and pathogenesis of invasive infection. Earlier work by Gryllos et al. found that exposure of GAS to subinhibitory concentrations of LL-37 up-regulated expression of hasB, spyCEP/scpC/prtS, mac/IdeS, and SPy0170, genes that were shown previously to be down-regulated by extracellular Mg2+ in a CsrRS-dependent manner [11], [17]. Furthermore, the stimulatory effect of LL-37 on CsrRS-regulated gene expression could be blocked by high concentrations of Mg2+. These findings suggested the hypothesis that Mg2+ and LL-37 act as opposing extracellular signals for the CsrS sensor histidine kinase. To test if other CsrRS-regulated genes also respond to both stimuli, we investigated ten additional genes for their responsiveness to LL-37 and Mg2+ in GAS strain 854. This strain was chosen for further analyses because initial experimentation showed marked up-regulation of the four previously characterized CsrRS target genes by LL-37, signaling that was completely abrogated in an isogenic csrS deficient mutant [17]. Furthermore, strain 854 is representative of the widely disseminated M1T1 clone associated with invasive GAS infections over the past three decades [22]–[25]. In the present study, we found that exposure of strain 854 to 100 nM LL-37 resulted in up-regulation of speA (pyrogenic exotoxin A), sda1 (DNase), ska (streptokinase), slo (streptolysin O), nga (NAD-glycohydrolase), and SPy0136 (hypothetical protein; N.B., throughout this paper, unnamed open reading frames are designated by SPy numbers according to the SF370 or MGAS315 genome sequences [26], [27]), as assessed by quantitative RT-PCR (qRT-PCR) analysis of RNA samples from LL-37-treated and untreated bacteria (Figure 1). Culture of strain 854 in 15 mM Mg2+ had the opposite effect from that evoked by LL-37. That is, Mg2+ exposure resulted in down-regulation of these genes relative to their expression at baseline in unsupplemented medium (Figure 1). Conversely, expression of several genes in strain 854 was repressed by LL-37 and up-regulated by Mg2+. Genes in the latter category included metB (putative cystathionine beta-lyase), SPy1414 (putative cation (potassium) transport protein), grab (protein G-related α2-macroglobulin-binding protein), and speB (cysteine protease) (Figure 1). To verify that the changes in gene expression observed in response to LL-37 resulted in corresponding changes in production of the encoded proteins, we assayed four representative virulence determinants from this group of CsrRS-regulated genes. Growth of strain 854 in the presence of LL-37 resulted in marked increases in SLO and NADase and repression of SpeB, as assessed by western blot, and increased DNase activity (Figure S1). DNase activity associated with invasive M1T1 isolates such as strain 854 has been shown to be due predominantly to the enzyme encoded by the prophage-associated sda1 gene (also called sdaD2), a member of the CsrRS regulon [28]. These results corroborate the qRT-PCR data and, together, they extend earlier findings that LL-37 can up-regulate gene expression to include several additional CsrRS-controlled genes. Moreover, they show that expression of certain CsrRS-regulated genes is repressed, rather than stimulated, by LL-37. For both categories of genes, the effect of Mg2+ is opposite to that of LL-37, an observation that supports the hypothesis that the two molecules act as functionally antagonistic stimuli for signaling through CsrRS. The predicted histidine kinase CsrS is thought to represent a cell-surface sensor component of the CsrRS TCS that detects and responds to environmental signals. According to secondary structure and membrane protein model predictions, CsrS contains two membrane-spanning domains near the N-terminus that flank a predicted ECD of 151 amino acids [16]. To test these model predictions, membrane and cytoplasmic fractions of wild type GAS 854 and control csrS-deficient strain 854csrSΩ were isolated from whole cell lysates, fractionated by SDS-PAGE, and analyzed by western blot with anti-CsrS serum. CsrS was found exclusively in membranes of wild type bacteria and, as expected, was absent from csrS mutant preparations (Figure 2A). Like CsrS, the unrelated membrane protein OpuABC [11] was also in wild type 854 membranes, but not in the cytoplasmic fractions (Figure 2A). Consistent with its predicted cytosolic localization, the CsrR protein was mainly detected in the cytoplasmic fraction. These results localized CsrS to the GAS cell membrane. In order to test whether CsrS is accessible to signaling molecules in the extracellular environment, we labeled proteins exposed on the bacterial surface with biotin via a disulfide linker. Biotinylated proteins were isolated from bacterial cell lysates using NeutrAvidin resin affinity chromatography. Resin-bound proteins were released by reduction of the disulfide bond linking biotin to the GAS surface proteins, fractionated by SDS-PAGE, and analyzed by western blot with CsrS antiserum. CsrS was detected predominantly in this eluted fraction (Figure 2B), a result that indicates CsrS was accessible to biotinylation, i.e., that a portion of the protein is exposed to the extracellular environment. CsrR, used here as a control cytosolic protein, did not react with biotin, and was detected only in the unbound flow-through fraction (Figure 2B). These data demonstrate that CsrS is a membrane-associated protein and includes a surface-exposed domain, conclusions consistent with our hypothesis that the ECD of CsrS functions as the sensor domain for environmental signals. We noted previously that the predicted ECD of CsrS includes a cluster of negatively charged amino acids that corresponds to a similar cluster in PhoQ of S. typhimurium and E. coli implicated in binding of cationic ligands [17], [29], [30]. Preliminary experiments using wild type or mutant forms of csrS to complement in trans a csrS mutant of M-type 3 GAS strain DLS003 suggested that three charged residues in the ECD were required for LL-37 signaling. However, these experiments were not definitive as the level of CsrS protein expressed from the mutant construct was higher than that observed in the wild type strain [17]. To examine more thoroughly the role of the predicted CsrS ECD in LL-37 sensing by CsrS, we introduced point mutations into the chromosomal csrS locus of GAS strain 854 by allelic replacement. Four independent mutant strains were constructed in which one or all three negatively charged amino acids localized in a small cluster of acidic residues (148DHIED152, Figure 3A) were substituted with similar uncharged residues (D148N, E151Q or D152N). Mutation of these three amino acids did not affect expression levels or surface localization of mutant CsrS, as similar quantities of CsrS were detected in western blots of membrane fractions obtained from the csrS triple point mutant strain 854csrSTM and wild type 854 (Figure 2A), and similar amounts of CsrS were labeled by biotinylation on the mutant strain cell surface (Figure 2B). The four resulting isogenic csrS mutants 854csrSD148N, 854csrSE151Q, 854csrSD152N, and 854csrSTM were tested for LL-37-mediated up-regulation of hasB, spyCEP, mac, and SPy0170 expression. Wild type strain 854 and each of the mutant strains were grown to early exponential phase in the presence or absence of 100 nM LL-37, and gene expression was assessed by qRT-PCR. In contrast to wild type, the isogenic csrS triple mutant showed little or no change in gene expression in response to LL-37 (Figure 3B). The csrS mutants with single amino acid substitutions (D148N, E151Q or D152N) all showed moderate LL-37-mediated up-regulation of the four target genes, but less than that observed in wild type (Figure 3B). Mutation of this region of the CsrS ECD also abrogated or severely blunted the effect of Mg2+ to repress, or in the case of grab, to activate, CsrRS-regulated gene expression (Figure S2). The results above provide evidence that the mutated cluster of acidic residues in the predicted ECD is critical for LL-37 and Mg2+ signaling through CsrS in strain 854. To confirm these findings and to test their generality for other GAS strains, we constructed an analogous csrS triple point mutant of M-type 49 strain NZ131 and examined its response to LL-37 by qRT-PCR. Similar to the results in the 854 background, we observed almost complete loss of LL-37-stimulated up-regulation of hasB, mac, and SPy0170 in NZ131csrSTM, and a marked reduction in spyCEP up-regulation (Figure 3C, left panel). In wild type NZ131, expression of these four genes was repressed during growth in 15 mM Mg2+, but no such repression was observed for hasB, spyCEP, or Spy0170 in the NZ131csrSTM (Figure 3C, right panel). Thus, similar findings in two independent strain backgrounds highlight the importance of a small cluster of negatively charged amino acids in the predicted CsrS ECD in LL-37 and Mg2+ signaling through CsrRS. During characterization of the csrS triple mutant, we noted that mutant bacteria formed compact, glossy colonies similar to wild type 854, but distinctly different from the mucoid colony appearance of 854csrSΩ lacking CsrS. To verify that the distinctive colony morphology reflected a difference in capsule gene expression, we compared relative expression of hasB (from the hyaluronic acid capsule biosynthetic operon) in the three strains. As expected, in the absence of supplemental Mg2+ or LL-37, expression of hasB was increased more than 50-fold in strain 854csrSΩ relative to that in wild type 854, whereas hasB expression in the csrS triple point mutant was actually reduced by 40% compared to wild type (Table 1). This finding of reduced capsule gene expression suggested that the ECD mutations in the triple mutant resulted not only in refractoriness to regulation by LL-37, but also in increased activity of the CsrR response regulator, presumably by enhancing its phosphorylation in the absence of signaling from an external ligand. Such an effect could result from increased autokinase activity of CsrS or reduced phosphatase activity of CsrS for phospho-CsrR. To test this hypothesis, we compared expression of additional CsrRS-regulated genes in the csrS triple mutant relative to wild type 854. As observed for hasB, in the absence of supplemental Mg2+ or LL-37, expression of spyCEP, mac, and SPy0170 was down-regulated in the triple mutant compared to wild type, whereas expression of each of these genes was significantly up-regulated in 854csrSΩ relative to wild type expression levels (Table 1). Furthermore, expression of grab was increased in the triple mutant relative to wild type levels (data not shown). As grab is activated by supplemental Mg2+ and repressed by LL-37 (Figure 1), this result is also consistent with the proposed model of increased CsrR activity in the triple mutant. To test directly whether the altered ECD of the triple mutant changed gene regulation by affecting autokinase activity of CsrS, we inactivated the kinase by replacing the active site histidine residue with alanine (H280A). As expected, when introduced in strain 854, this mutation resulted in a mucoid colony morphology, and the mutant strain 854csrSH280A displayed marked up-regulation of CsrRS-repressed genes in a pattern very similar to that observed in 854csrSΩ (Table 1). Similarly, introduction of the H280A mutation into the CsrS triple mutant resulted in mucoid colonies and a comparable derepression of CsrRS-repressed genes as in 854csrSΩ and in 854csrSH280A (Table 1). Since mutation of the active site histidine of CsrS abrogated the suppressive effect of the ECD triple mutant, the most parsimonious model is that these mutations in the ECD affect gene expression by altering the autokinase activity of CsrS. While an effect on phosphatase activity is not excluded by these experiments, the results suggest strongly that the ECD triple point mutant expresses a constitutively active CsrS histidine kinase that is relatively refractory to signaling induced by external stimuli. CsrRS regulates the expression of several genes that encode products implicated in GAS resistance to opsonophagocytic killing and cytotoxicity: hasABC, slo, nga, spyCEP, sda1, mac, and speB. Upregulation of antiphagocytic factors by host LL-37 is expected to enhance virulence in vivo; however, testing this hypothesis directly in an animal model is not possible, currently, since cathelicidins of other mammalian species do not share the CsrRS-signaling activity of human LL-37 [17]. Because in vitro resistance to phagocytic killing by human blood leukocytes correlates with GAS virulence in vivo [31], we used an in vitro assay to assess the effect of LL-37 on phagocytic resistance as a proxy for effects on in vivo virulence. As would be predicted by the effects of LL-37 on regulation of antiphagocytic factors, exposure of four unrelated wild type GAS strains to LL-37 increased resistance of all four strains to phagocytic killing in vitro [17]. Inactivation of CsrR in the M-type 3 strain DLS003 also resulted in increased resistance to phagocytic killing by human peripheral blood leukocytes, consistent with the marked up-regulation of CsrRS-regulated antiphagocytic factors in the mutant strain [12]. Because deletion of CsrS results in a similar, although less marked, up-regulation of CsrRS-controlled genes, we expected that deletion of CsrS or inactivation of its histidine kinase activity would also lead to increased resistance to phagocytic killing. In vitro opsonophagocytic assays of 854csrSΩ and 854csrSH280A confirmed these predictions: both mutant strains were highly resistant to phagocytic killing by human blood leukocytes in vitro similar to a ΔcsrR mutant (Figure 4). In marked contrast, the csrS triple mutant was as susceptible to killing as wild type 854 in the absence of LL-37, but did not show any increase in phagocytic resistance in response to LL-37 unlike wild type 854 (Figure 4). These observations further support the proposed model that the csrS triple mutant exhibits constitutive activation of CsrS autokinase activity and tonic phosphorylation of CsrR. An important consequence is down-regulation of CsrRS-controlled antiphagocytic factors and hyporesponsiveness to the stimulatory effect of LL-37. The current working model for the CsrRS TCS is that of a classical sensor histidine kinase linked by phosphotransfer to a response regulator whose activity is controlled by its phosphorylation state. The experiments described above provide new evidence to support the surface location and stimulus-regulated histidine kinase activity of CsrS. We investigated also the role of CsrR in this model by further characterizing strain 950771, a GAS M-type 3 strain that exhibits a high level of capsular polysaccharide production that does not increase upon exposure to LL-37. Sequencing the csrRS locus in 950771 revealed a point mutation in csrR (K102R) in the highly conserved CsrR receiver domain [17]. The lysine residue that is mutated in 950771 is conserved not only in the csrR locus of all sequenced GAS strains, but also in response regulators of many TCS in a wide variety of bacterial species where it occupies a location near the conserved aspartic acid residue that is the site of phosphorylation [32]. In E. coli CheY, substitution of arginine for lysine at the corresponding site (K109R) did not prevent phosphorylation, but abrogated induction of tumbling motility that normally results from CheY phosphorylation, a finding interpreted to mean that the conserved lysine is required for phosphorylation to produce the active conformation of the response regulator [33]. To test whether K102 is required for GAS CsrR to transmit a signal from CsrS, we replaced R102 in strain 950771 with the consensus lysine residue (R102K). Whereas isolate 950771 (R102) showed no response to LL-37 or to Mg2+, strain 950771csrRR102K restored both up-regulation of hasB, spyCEP, mac, and SPy0170 in response to LL-37 and down-regulation in response to 15 mM Mg2+ (Figure 5A). In addition, correction of CsrR to the consensus K102 sequence markedly reduced expression of all four genes during growth in unsupplemented medium (Figure 5B), a result that implies that K102 is necessary for transduction of the signal mediated by tonic phosphorylation of CsrR by CsrS under standard laboratory growth conditions. To further test whether the K102R CsrR mutation is sufficient to prevent CsrRS-mediated modulation of target gene expression, we also introduced the K102R mutation in wild type strain NZ131. In contrast to wild type NZ131 that exhibited a 2- to 25-fold increase in expression of hasB, spyCEP, mac, and SPy0170 in response to 100 nM LL-37, mutant strain NZ131csrRK102R showed no response (Figure 5C, left panel). Moreover, the mutant failed to repress expression of these genes in response to 15 mM Mg2+ (Figure 5C, right panel). Together, these results indicate that the conserved lysine residue at position 102 in CsrR is required for signal transduction from CsrS to modulate target gene expression in response to extracellular LL-37 or Mg2+. Previous studies have demonstrated that CsrRS regulates expression of more than 100 GAS genes including those encoding many important virulence determinants [9]-[11]. In addition, evidence has been presented that elevated levels of extracellular Mg2+ result in down-regulated expression of CsrR-repressed genes, whereas exposure of GAS to subinhibitory concentrations of LL-37 has the opposite effect [11], [16], [17]. Results of the present study demonstrate the same reciprocal pattern of regulation by LL-37 and Mg2+ for an expanded repertoire of GAS genes. While the predominant pattern of regulation is one of up-regulation of gene expression by LL-37 and down-regulation by Mg2+, we report several instances of the opposite pattern, that is, repression of gene expression by LL-37 and activation by Mg2+. The current investigation provides new experimental evidence that supports a model of CsrRS as a classical TCS that responds to these environmental signals through modulation of CsrS autokinase activity, with downstream signaling that depends on phosphotransfer from CsrS to the CsrR transcriptional regulator. A critical consequence of CsrRS signaling by LL-37 is the coordinated modulation of expression of multiple genes in a fashion that dramatically increases GAS resistance to killing by phagocytes, a bacterial phenotype that enhances virulence and promotes invasive infection in vivo. In addition to the previously demonstrated up-regulation by LL-37 of the hyaluronic acid capsule synthesis operon, mac/IdeS (Mac/IgG protease), and spyCEP (IL-8 protease), we found that LL-37 activated and Mg2+ repressed expression of genes encoding several important virulence factors including ska (streptokinase), slo (streptolysin O), and nga (NAD-glycohydrolase), as well as speA (pyrogenic exotoxin A) and sda1 (DNase), two virulence determinants encoded by prophages associated with the invasive M1T1 GAS clonal group [34]. The opposite pattern of regulation was observed for speB (cysteine protease) and grab (protein G-related α2-macroglobulin-binding protein). Repressed expression of speB in response to LL-37 may also contribute to an invasive phenotype, as the speB-encoded cysteine protease has been proposed to degrade the anti-phagocytic M1 protein and to inactivate the sda1 gene product, a DNase that itself enhances GAS virulence by degrading neutrophil extracellular traps (NETs) [21], [25], [35], [36]. Protein modeling of CsrS indicates the presence of two membrane-spanning regions that flank a domain predicted to form an extracellular loop that represents a potential site for interaction with environmental stimuli [12], [16]. In cell-fractionation experiments, we found that CsrS is physically associated with the bacterial cell membrane, as predicted by this model. Furthermore, CsrS on intact bacterial cells was accessible to biotin-labeling, a result that implies that a domain of the protein lies in the extracellular space. We also investigated the role in CsrS-mediated signal transduction of a small cluster of negatively charged amino acids in the CsrS ECD. Because a similar cluster of acidic residues has been implicated in binding of cationic ligands to E. coli and S. typhimurium PhoQ, we tested in GAS the effect of substituting uncharged amino acids for these three residues. While our intent had been to disrupt binding of Mg2+ and/or LL-37 to the CsrS ECD, we discovered that this relatively small alteration in the ECD not only abrogated ligand signaling, but also resulted in a global effect on the CsrRS regulon consistent with tonic activation of CsrS autokinase activity. Support for this hypothesis also came from the observation that the effects of the csrS ECD mutations were overridden by mutating the active site histidine of the CsrS kinase domain, a result that implies that the effects of the former mutations on target gene regulation are mediated through CsrS kinase activity. Thus, in the absence of increased extracellular Mg2+ or exposure to LL-37, expression of CsrR-repressed genes was reduced in the csrS triple mutant. The finding that neutralizing the charge of three amino acids in the CsrS ECD leads to an apparent activation of CsrS kinase activity and hyporesponsiveness to ligand signaling suggests that the mutations result in a conformational change in the cytoplasmic domain of CsrS that mimics that induced by binding of Mg2+ to the ECD. It is tempting to speculate that binding of Mg2+ and/or LL-37 to the same region of the ECD also modulates kinase activity by this mechanism, although attempts to demonstrate specific binding of either ligand to the ECD have, so far, been unsuccessful. The data summarized above suggest strongly that LL-37 signaling depends on direct interaction of the peptide and/or Mg2+ with the extracellular domain of CsrS. However, we cannot exclude an alternative signaling mechanism such as membrane disruption by LL-37 that secondarily results in altered CsrS autokinase activity. We found that deletion of CsrS or inactivation of its kinase activity produced a similar pattern of altered gene expression as deletion of CsrR, although the magnitude of change in gene expression was somewhat smaller for some genes. These observations imply that, under laboratory growth conditions, CsrS activates CsrR, presumably by phosphorylation, increasing its activity as a transcriptional regulator. Activation of CsrR by CsrS can be increased by exposure to elevated extracellular Mg2+ or reduced by exposure to LL-37. The results discussed above support a model in which expression of the CsrRS regulon depends on the equilibrium between the phosphorylated and unphosphorylated states of CsrR. Increased extracellular Mg2+ or mutation of critical residues in the CsrS ECD increases CsrS phosphorylation and enhances phosphotransfer to CsrR, shifting the equilibrium toward phospho-CsrR with consequent repression of CsrR-repressed genes and activation of CsrR-activated genes. Conversely, exposure to LL-37 or deletion of CsrS shifts the equilibrium toward unphosphorylated CsrR, which is less active in regulating target promoters. Transduction of these modulating signals to altered transcriptional regulation depends also on the presence of a conserved lysine residue at position 102 in the receiver domain of CsrR. A natural mutant with a conservative arginine substitution at this position was refractory to signaling by extracellular Mg2+ or exposure to LL-37 and exhibited a pattern of gene expression similar to that of a CsrR deletion mutant. These phenotypes were confirmed by repairing the natural mutant to the consensus K102 and by introducing the K102R mutation into an unrelated wild type strain. On the basis of these findings and work by others on the role of the corresponding lysine residue in bacterial TCS, we conclude that CsrR K102 is critical to transducing the signal of CsrR phosphorylation and to modulation of CsrR-mediated transcriptional regulation at target promoter sequences. Several studies have documented the emergence of GAS strains with spontaneous inactivating mutations in CsrS or CsrR in the setting of invasive infection [18]–[20]. Because such mutants have a gene expression profile that results in a multifactorial enhancement of resistance to clearance by host phagocytes, these mutant variants have a strong selective advantage for survival in microenvironments such as the bloodstream or deep tissue sites where they are exposed to attack by host phagocytes. However, analysis of a collection of GAS pharyngeal isolates indicated that CsrRS mutants are distinctly rare in this setting, in marked contrast to isolates from patients with severe systemic infection [19]. The predominance of strains with a functional CsrRS system in the pharynx implies that CsrRS-mediated dynamic regulation of gene expression in response to environmental cues contributes to adaptation of GAS to its preferred environmental niche. During initial colonization, the low concentration of LL-37 on the resting pharyngeal epithelium is predicted to result in an intermediate level of CsrRS activation and a corresponding moderate expression of CsrRS-regulated virulence factors. This “colonizing” phenotype, however, can change quickly in response to increased local concentrations of LL-37. The striking up-regulation of an antiphagocytic phenotype upon exposure to LL-37 enables the organism to maintain the capacity to arm itself against host effectors and thus resist clearance. The coordinated program of altered gene expression induced by LL-37 signaling can tip the balance of pathogen-host interaction from one of asymptomatic colonization to uncontrolled invasive infection. Paradoxically, secretion of LL-37 from injured epithelial cells or from degranulation of recruited neutrophils as part of the host innate immune response may trigger local or systemic invasion by GAS as a result of CsrRS-mediated virulence factor expression. The human subjects aspects of this study were approved by the institutional review board of Children's Hospital Boston. Written informed consent was provided by study participants. Wild type GAS strains used in this study and isogenic mutants derived from them are described in Table 2. GAS M-type 1 strain 854 is a clinical isolate from a patient with a retroperitoneal abscess [17]. GAS M-type 49 strain NZ131 is a skin isolate from a patient with glomerulonephritis [37]. GAS strain 950771 is an M-type 3 clinical isolate from a child with necrotizing fasciitis and sepsis [38]. GAS strains were grown at 37°C in Todd-Hewitt broth (Difco) supplemented with 0.5% yeast extract (THY) or on THY agar or trypticase-soy agar (BD Bioscience) supplemented with 5% defibrinated sheep blood. Escherichia coli (E. coli) strains DH5α (New England Biolabs) and StrataClone (Stratagene) were used for cloning. Recombinant protein overexpression for antisera production was carried out using E. coli strain BL21(DE3) (Novagen). Antibiotics were used when necessary at the following concentrations: for GAS, erythromycin 1 µg/ml; for E. coli, erythromycin 200 µg/ml, kanamycin 50 µg/ml, carbenicillin 100 µg/ml, and ampicillin 100 µg/ml. GAS cultures were grown in THY broth supplemented with or without 100 nM LL-37 or 15 mM MgCl2, and cells were harvested either at early exponential (A600 nm 0.25), mid-exponential (A600 nm 0.5), late exponential (A600 nm 0.8) or early stationary (A600 nm ∼1) growth phase. Total RNA extraction from bacterial cells was performed as described [11] during the growth phase at which target gene expression was maximal. RNA concentration and purity were determined using a NanoDrop spectrophotometer ND-1000 (Thermo Fisher Scientific). Quantitative RT-PCR was performed on an ABI PRISM 7300 Real-Time PCR system (Applied Biosystems) using the QuantiTect SYBR Green RT-PCR kit (Qiagen). Primers used are listed in Table S1. Expression level of each target gene was normalized to recA (spyM3_1800/SPy2116) and analyzed using the ΔΔCt method as described [11]. Replicate experiments were performed from at least three independent RNA preparations in triplicate. Statistical analysis was performed using the paired Student's t-test for expression level comparison under different growth conditions in a single strain and the unpaired t-test for testing differences between strains. The human cathelicidin LL-37 (a gift of Robert I. Lehrer, UCLA, CA, USA) was synthesized as described previously and its purity was confirmed by high-performance liquid chromatography and mass spectrometry [39]. To introduce single point and triple point mutations in the CsrS ECD region, vector pORIcsrS containing the wild type csrS sequence [11] served as template to amplify the entire plasmid by PCR with primer pair HTW 13/14 for csrS(D148N) substitution, HTW 15/16 for csrS(E151Q) substitution, HTW 17/18 for csrS(D152N) substitution, or csrS418-F(muNHIQN)/csrS480-R(muNHIQN) for csrS triple point substitution (D148N,E151Q,D152N) as described in the Quikchange site-directed mutagenesis protocol (Stratagene). From the resulting plasmids, csrS fragments used for allelic replacement were PCR-amplified with Phusion high-fidelity DNA polymerase (Finnzymes, New England Biolabs) by using primers csrS-F(PshAI) and rt0245-R and cloned into vector pSC-B (StrataClone blunt PCR cloning kit, Stratagene). To introduce a csrS H280A mutation into GAS strain 854 and into the isogenic triple mutant strain 854csrSTM, a csrS fragment was amplified by PCR from wild type 854 chromosomal DNA with primer pair 5005_149F/5005_1204R and cloned into pGEM-T (Promega). The resulting plasmid was used for Quikchange site-directed mutagenesis to csrS H280A with primer pair H280A-F/H280A-R. For generating a csrR deletion mutant in strain 854 (854ΔcsrR), an overlap PCR using Phusion DNA polymerase was performed of the region upstream of csrR with primer pair Reg1P-F/HTW52 and the region downstream of csrR including about 500 bp of csrS with primer pair HTW 53/54. The hybridized strands of the two resulting PCR products were used as a template for the second PCR amplifying a 1 kb fragment encompassing a CsrR deletion of amino acids 3–223. The final product was ligated into vector pSC-B. Subsequently, the csrS or ΔcsrR fragments described above were released from pSC-B or pGEM-T by SalI/BamHI digestion and were subcloned into the temperature-sensitive shuttle vector pJRS233 [40]. To introduce the consensus CsrR (csrR(K102)) into GAS strain 950771 and non-consensus CsrR (csrR(R102)) into GAS strain NZ131, csrR R102 was amplified from 950771 chromosomal DNA was amplified by using Platinum Taq high fidelity DNA polymerase (Invitrogen) and primers CsrP-F and csrS176-R. The PCR product was cloned into pGEM-T and then subcloned into pJRS233 using PstI and XbaI restriction sites to generate pJRS-csrR(R102). Resulting plasmid pJRS-csrR(R102) was then used for Quikchange site-directed mutagenesis to convert csrR(R102) to csrR(K102) by using primer pair HTW 71/72, creating plasmid pJRS-csrR(K102). All primers are described in Table S1. Recombinant pJRS233 shuttle plasmids were electroporated into GAS strains 854, 854csrSTM, NZ131, or 950771 and then subjected to allelic gene replacement as described [38]. To confirm the genotype of mutant strains, csrR and csrS loci were PCR-amplified with Easy-A high fidelity DNA polymerase (Stratagene) from chromosomal DNA and the sequences confirmed by DNA sequencing (DNA Sequencing Core, Brigham and Women's Hospital, Boston, MA, USA). Cultures of 854, 854csrSΩ, and 854csrSTM were grown in THY at 37°C to an A600 nm of ∼0.4, cells were collected (1250 × g, 8 min) and washed once with 10 mM Tris-HCl, pH 8.0, and resuspended in 360 µl hypotonic TEG buffer (10 mM Tris-HCl, 1 mM EDTA, 20% glucose, pH 8.0) supplemented with protease inhibitor cocktail III (Calbiochem). For peptidoglycan degradation, mutanolysin (∼500 units, Sigma) and lysozyme (∼17,700 units, Sigma) were added and samples were shaken at 1000 rpm at 37°C for 1 h in an Eppendorf thermomixer. Cells were washed once in 500 µl TEG buffer and resuspended in 500 µl TE buffer (10 mM Tris-HCl, 5 mM EDTA, pH 8.0) supplemented with protease inhibitor cocktail (Roche). Cells were lysed by ultrasonication (5×3 sec bursts on level 5, Sonic Dismembrator model 60, Fisher Scientific) on ice followed by centrifugation (Eppendorf 5417C 10,000 × g) for 20 min at 4°C to remove cell debris. Membranes were separated from the cytoplasmic fraction by ultracentrifugation of supernatants (Beckman Coulter Ultima, TLA-100.3 rotor) for 1 h at 90,000 × g at 4°C. Membranes and cytoplasmic fractions were resuspended in SDS-PAGE sample buffer and heated to boiling. GAS strain 854 was grown in liquid culture to A600nm 0.7 (late exponential phase) or A600nm 1.2 (stationary phase) in the absence or presence of 100 nM LL-37. Bacteria were removed by centrifugation (21,000 × g, 5 min). Cell-free supernatants were used for assays of DNase activity or mixed with sample buffer and heated to boiling before SDS-PAGE and western blot analysis. Samples were fractionated on 10% (membrane and cytoplasmic fractions) or 4–12% gradient (supernatant proteins) NuPAGE Novex Bis-Tris gels and then transferred to nitrocellulose membranes for western blotting as previously described [11]. Blots were incubated with specific rabbit antiserum against GAS CsrS ECD [11], CsrR, SLO [41], NADase [41], or SpeB (Toxin Technology, Sarasota, FL) at a 1∶1000 dilution, or with mouse antiserum against GAS membrane protein OpuABC (courtesy of Giuliano Bensi, Novartis Vaccines) at a 1∶3000 dilution, each followed by horseradish-peroxidase-linked secondary antibody [11]. Signal development was carried out using the SuperSignal West Pico chemiluminescence substrate (Thermo Scientific Pierce). For surface biotinylation the Cell Surface Protein Isolation Kit (Pierce) was used according to the manufacturer's protocol with the following modifications. GAS were grown in 30 ml THY at 37°C to A600 nm ∼ 0.3, cells were harvested (Centra CL3, Thermo IEC, 1250 × g, 8 min), washed twice in 1.5 ml PBS (Eppendorf 5417C, 9800 × g, 1 min), and resuspended in 1.5 ml Sulfo-NHS-SS-Biotin labeling solution. The biotinylation is reversible by cleavage of the disulfide bond in Sulfo-NHS-SS-Biotin. As a negative control, wild type 854 cells were incubated with PBS instead of biotin labeling solution. After 30 min agitation at 4°C, 100 µl quenching solution was added to treated cells and the cells were centrifuged at 6800 × g for 4 min. Cells were washed twice in 1.5 ml TBS (25 mM Tris-HCl, 0.15 M NaCl, pH 7.2) and frozen at −20°C. Frozen cells were resuspended in 250 µl lysis buffer supplemented with 2.5 µl protease inhibitor cocktail III (Calbiochem) and lysed by two rounds of ultrasonication (5×1 s bursts, level 1) with an incubation on ice in between. Lysates were centrifuged at 20,800 × g for 4 min at 4°C. Supernatants were incubated with 250 µl immobilized NeutrAvidin resin for 60 min in spin-columns with end-over-end rotation. Flow-through samples were retained and mixed with SDS-PAGE sample buffer. Resin was washed four times with 500 µl wash buffer supplemented with protease inhibitor and incubated with 200 µl SDS-PAGE sample buffer with 50 mM DTT for 60 min with end-over-end rotation at RT. Eluates were collected by brief centrifugation of uncapped spin-columns. Samples were heated at 95°C for 5 min and stored at −20°C until needed for western blot analysis. A 1∶3 mixture of antiserum to CsrS ECD and antiserum to N-terminal truncated CsrS was used to detect CsrS protein on the blots. Full-length CsrR and N-terminal truncated CsrS (CsrSΔ1-231) and were fused separately to a N-terminal His6 tag by cloning into overexpression vector pET-28a (Novagen) PCR-amplified DNA fragments obtained with primer pairs JL-48/JL-49 and HTW 37/46, respectively. Following overexpression by IPTG induction, recombinant proteins were affinity purified using Ni2+-NTA resin (Qiagen) under native conditions (His6-CsrR) or under denaturing conditions (His6-CsrSΔ1-231) according to the manufacturer's protocol. Purified proteins were used to immunize rabbits (LAMPIRE Biological Laboratories, Inc., Pipersville, Pennsylvania, USA). Reactivity of immune sera against CsrS or CsrR was evaluated by western blotting of GAS lysates. DNase activity in GAS culture supernatants was assayed as described by Aziz et al. with modifications [42]. Supernatants were diluted 1∶125 in sterile deionized water, and 10 µL samples were mixed with 1 µg plasmid DNA in 100 mM Tris, pH 7.5, supplemented with 1 mM CaCl2 and 1 mM MgCl2 in a 15 µL total reaction volume. Samples were incubated at 37°C for 20 min and then were stopped by the addition of 20 mM EDTA. Samples were analyzed on 1% agarose gels and DNA was visualized with SYBR Safe DNA stain (Invitrogen). GAS resistance to phagocytic killing was evaluated by an in vitro assay as described [43]. In brief, GAS strains grown to early exponential phase with or without 100 nM LL-37 were mixed with freshly isolated human peripheral blood leukocytes at a multiplicity of infection of 3 – 4 in the presence of 10% human serum as complement source. Aliquots were withdrawn for quantitative culture immediately after mixing and after 1 h end-over-end rotation at 37°C. Results were reported on a log scale as the fold-change in cfu defined as the total cfu after incubation divided by the total starting cfu. Statistical significance of differences in the capacity of GAS strains to resist opsonophagocytic killing were evaluated by one-way ANOVA with Bonferroni's post-test analysis.

10.1371/journal.pgen.1003296

Ancient DNA Reveals Prehistoric Gene-Flow from Siberia in the Complex Human Population History of North East Europe

North East Europe harbors a high diversity of cultures and languages, suggesting a complex genetic history. Archaeological, anthropological, and genetic research has revealed a series of influences from Western and Eastern Eurasia in the past. While genetic data from modern-day populations is commonly used to make inferences about their origins and past migrations, ancient DNA provides a powerful test of such hypotheses by giving a snapshot of the past genetic diversity. In order to better understand the dynamics that have shaped the gene pool of North East Europeans, we generated and analyzed 34 mitochondrial genotypes from the skeletal remains of three archaeological sites in northwest Russia. These sites were dated to the Mesolithic and the Early Metal Age (7,500 and 3,500 uncalibrated years Before Present). We applied a suite of population genetic analyses (principal component analysis, genetic distance mapping, haplotype sharing analyses) and compared past demographic models through coalescent simulations using Bayesian Serial SimCoal and Approximate Bayesian Computation. Comparisons of genetic data from ancient and modern-day populations revealed significant changes in the mitochondrial makeup of North East Europeans through time. Mesolithic foragers showed high frequencies and diversity of haplogroups U (U2e, U4, U5a), a pattern observed previously in European hunter-gatherers from Iberia to Scandinavia. In contrast, the presence of mitochondrial DNA haplogroups C, D, and Z in Early Metal Age individuals suggested discontinuity with Mesolithic hunter-gatherers and genetic influx from central/eastern Siberia. We identified remarkable genetic dissimilarities between prehistoric and modern-day North East Europeans/Saami, which suggests an important role of post-Mesolithic migrations from Western Europe and subsequent population replacement/extinctions. This work demonstrates how ancient DNA can improve our understanding of human population movements across Eurasia. It contributes to the description of the spatio-temporal distribution of mitochondrial diversity and will be of significance for future reconstructions of the history of Europeans.

The history of human populations can be retraced by studying the archaeological and anthropological record, but also by examining the current distribution of genetic markers, such as the maternally inherited mitochondrial DNA. Ancient DNA research allows the retrieval of DNA from ancient skeletal remains and contributes to the reconstruction of the human population history through the comparison of ancient and present-day genetic data. Here, we analysed the mitochondrial DNA of prehistoric remains from archaeological sites dated to 7,500 and 3,500 years Before Present. These sites are located in North East Europe, a region that displays a significant cultural and linguistic diversity today but for which no ancient human DNA was available before. We show that prehistoric hunter-gatherers of North East Europe were genetically similar to other European foragers. We also detected a prehistoric genetic input from Siberia, followed by migrations from Western Europe into North East Europe. Our research contributes to the understanding of the origins and past dynamics of human population in Europe.

Our current knowledge of the origins of human populations and their migratory history relies on archaeological, anthropological, linguistic and genetic research. The study of genetic markers, especially the maternally inherited mitochondrial DNA (mtDNA), has allowed important events in the genetic history of humans to be reconstructed [1]–[11]. However, reconstructions based solely on present-day genetic diversity can be biased by a variety of evolutionary mechanisms, such as genetic drift and/or past population events. The ability to accurately reconstruct recent human evolutionary events can be significantly improved through the direct analysis of ancient human remains from representative time periods. The mtDNA diversity of prehistoric populations has been previously described for Palaeolithic/Mesolithic hunter-gatherers from Central, Eastern and Scandinavian Europe [12]–[14], and for Neolithic farmers from Southern and Central Europe (CE) [15]–[20]. These studies have uncovered an unexpected and substantial heterogeneity in the geographical, temporal and cultural distribution of the mtDNA diversity. However, little is known about past mtDNA diversity in North East Europe (NEE), including the Baltic region, the Volga-Ural Basin (VUB), and sub-Arctic Europe. It is likely that different demographic events have been involved in shaping the gene pools of the populations of Western/Central Europe and NEE, due to the geographical position and distinct climatic conditions of the latter. During the Upper Palaeolithic (∼30,000–40,000 years before present, yBP), the northernmost latitudes of Europe were covered by an ice sheet that prevented settlement by anatomically modern humans. With the glacial retreat at the end of the Ice Age (∼11,500 yBP) [21], small foraging groups progressed into NEE from southern periglacial refuges [22]–[23]. As climatic conditions improved in the early Holocene (8,000–10,000 yBP), the first human settlements appeared in the Kola Peninsula [24], and foraging activities intensified in the steppe-forest zone of Northern Europe leading to the widespread establishment of complex Mesolithic societies of fishermen and hunter-gatherers [23], [25]–[26]. At the same time, Western Europe and CE were undergoing the Neolithic transition, during which an agricultural lifestyle spread rapidly, largely due to favorable climatic and ecological conditions. The Neolithic transition is thought to have been slower and more gradual in NEE than in Western/Central Europe and to have involved little migration of early farmers from CE [27]. From the Neolithic onwards, contacts between populations of NEE and groups living in the South are evident in archaeological and historical records [24]. Around the Baltic, historical records describe numerous population movements that originated in Scandinavia (e.g., Viking incursions ∼800 Anno Domini, AD [28]), Western/Central Europe (e.g., the Slavic migrations ∼700–1,000 AD [29]) or Central/East Siberia (e.g., the Mongol invasions ∼500–700 AD [30]). The geographical position of NEE makes it subject to influences from both Western and Eastern Eurasia, which could explain the linguistic and cultural diversity, observed in the area today. Two different linguistic families are spoken: Indo-European languages (Slavic, Baltic and Germanic) and Finno-Ugric languages (e.g., Estonian, Finnish, Mari, Saami [31]). Saami people of Fennoscandia (northern Norway, Sweden, Finland and Russia) are considered unique among Europeans in terms of their nomadic lifestyle and their livelihood, which is mainly based on fishing and reindeer herding. The ethnogenesis of the Saami remains unclear and two origins in Western and Eastern Europe were proposed [24], [32]–[33]. The Saami differ from the rest of the European populations in their reduced genetic diversity [1], [34]–[35], and mtDNA lineages that are otherwise very rare in European populations (haplogroups or hgs, U5b1b1a, V, Z1 and D5). In particular, the Saami-specific U5b1b1a clade is defined by the so-called hypervariable region I (HVR-I) ‘Saami motif’ 16144C-16189C-16270T (numbering according [36]) [37]. These lineages are also detected at low frequencies in adjacent NEE populations [32], [38]–[40], which on the other hand fall within the European mtDNA diversity and appear rather homogeneous irrespective of their languages [3]–[5], [38], [40]. Subtle mtDNA differences are however observed among them due to variable influences from genetically differentiated neighboring populations: central Europeans in the West, Saami in the North, and people from the VUB in the East. The absence of strong structure in the present-day mtDNA gene pool of NEE stands in contrast to the variety of languages and cultures, and to the complex history of how and when these were formed. Modern mtDNA data does not resolve the origins of the Saami either. Our aim was to provide answers to these questions and reconstruct events in the genetic history of NEE by generating and analyzing ancient DNA (aDNA) data from prehistoric human remains collected in northwest Russia (Figure 1). In particular, our objective was to characterize the genetic relationships between hunter-gatherer populations in NEE and Central/Northern Europe and to estimate the genetic legacy of ancient populations to present-day NEE and Saami. The oldest samples were collected in the Mesolithic graveyards of Yuzhnyy Oleni Ostrov (aUz; ‘Southern Reindeer Island’ in Russian) and Popovo (aPo), both dated around 7,000–7,500 uncalibrated. yBP, uncal. yBP. The sites of aUz and aPo are located along one of the proposed eastern routes for the introduction of Saami-specific mtDNA lineages [32]. Results from odontometric analyses suggested a direct genetic continuity between the Mesolithic population of Yuzhnyy Oleni Ostrov and present-day Saami [41]. We also analyzed human remains from 3,500 uncal. yBP site Bol'shoy Oleni Ostrov (aBOO; ‘Great Reindeer island’ in Russian) in the Kola Peninsula. This site is located within the area currently inhabited by the Saami. We compared the ancient mtDNA data from NEE with a large dataset of ancient and modern-day Eurasian populations to search for evidence of past demographic events and temporal patterns of genetic continuity and discontinuity in Europe. The skeletal remains from aUz, aPo, and aBOO were genetically analysed by i) direct sequencing of the mtDNA hyper-variable region I (HVR-I, nucleotide positions, np 16056–16409) and ii) typing of 22 haplogroup-diagnostic single nucleotide polymorphisms (SNPs) in the coding-region using the GenoCore22 reaction [16]. Strict criteria were followed to authenticate aDNA data and detect contamination by exogenous DNA or artefactual mutations caused by post-mortem DNA damage (see Materials and Methods). In total, 34 ancient genotypes were obtained that were considered unambiguous on the basis of these authenticity criteria (Table 1). Sequences have been deposited in Genbank (http://www.ncbi.nlm.nih.gov/genbank/; accession numbers KC414891-KC414924). The success of DNA amplification reactions varied among archaeological sites as follows: 9/42 individuals (21.5%) for aUz, 2/3 (66.7%) for aPo, and 23/23 (100.0%) for aBOO. The higher success rates (100%) observed for samples from aBOO were consistent with their younger age and excellent macroscopic preservation, probably due to the cold climatic conditions of the Kola Peninsula (Figure S1). The presence of naturally crushed marine shells in the burial grounds of aBOO has also been proposed to explain the exceptional preservation of the remains [42]. In contrast, and in accordance with their poorer macroscopic preservation, aDNA from the samples of aUz and aPo was more difficult to amplify, with a lower amplification success and some contaminated results that had to be excluded. In order to identify the genetic affinities of the two ancient populations with other ancient and present-day Eurasian populations, mtDNA hg distributions were compared by Principal Component Analysis (PCA). The PCA plot of the first two components (41.5% of the total variance, Figure 2) showed that present-day populations largely segregate into three main clusters: Europeans (in yellow), Middle Easterners (in grey) and Central/East Siberians (in blue). The spread of extant populations of Europe and Central/East Siberia along the first component axis (28.5% of the variance) appeared to reflect their longitudinal position, whereas Europeans and Middle Easterners were separated along the second component axis (13.0% of the variance). As shown previously, populations of the ‘Central/East Siberian’ cluster were predominantly composed of hgs A, B, C, D, F, G, Y, and Z, while in contrast populations of the ‘European’ cluster were characterized by higher frequencies of hgs H, HV, V, U, K, J, T, W, X, and I (e.g., [43]–[47]). The two ancient groups - aUzPo and aBOO - from two individual time periods appeared remarkably distinct on the basis of the PCA, suggesting a major genetic discontinuity in space and time. The hg distribution in the Mesolithic aUzPo population: U4 (37%), C (27%), U2e (18%), U5a (9%), and H (9%), indicated an ‘admixed’ composition of ‘European’ (U4, U2e, U5a and H, 73%) and ‘Central/East Siberian’ (C, 27%) hgs, based on the PCA plot (Figure 2). Interestingly, the population of aUzPo did not group with modern NEE populations, including Saami, but fell instead between the present-day ‘European’ and ‘Central/East Siberian’ clusters on the PCA graph, and more precisely between populations of the VUB (in light green) and West Siberia (in dark green). The high frequency of hg U4 is a feature shared between Mesolithic aUzPo, present-day VUB (Komi, Chuvashes, Mari), and West Siberian populations (Kets, Selkups, Mansi, Khants, Nenets), with the latter group also being characterized, like aUzPo, by the presence of hg C. The genetic affinity between Mesolithic aUzPo and present-day West Siberian populations could be visualized on the genetic distance map of North Eurasia (Figure 3A), on which locally lighter colorings indicated low values of genetic distances, and therefore an affinity between aUzPo and extant West Siberians. In order to test the potential population affinities formulated on the basis of the hg-frequency PCA and the distance map, we examined the present-day geographical distribution of the haplotypes found in aUzPo via haplotype sharing analyses (Figure 4). These analyses are less impacted by biases due to small population sizes or unidentified maternal relationships in ancient populations, and thus are less prone to artefacts. Although the highest percentages of shared haplotypes for aUzPo were observed in pools of West Siberian Khants/Mansi/Nenets/Selkups (2.8%), South Siberian Altaians/Khakhassians/Shors/Tofalars (2.2%) and Urals populations (Chuvash/Bashkirs, 2.0%), matches were widely distributed across Eurasia. This was consistent with the observation that most haplotypes sequenced in aUzPo were basal and hence, not informative in terms of geographical population affinity. Haplogroup-based analyses suggested that the genetic affinity between aUzPo and present-day West Siberians was partly due to the presence of hg C, implying that the non-basal haplotype C1 found in aUzPo (16189C-16223T-16298C-16325C-16327T, detected in three individuals) could be a clear genetic link with extant Siberian populations. However, the C1 haplotype found in aUz did not belong to hg C1a, the only C1 clade restricted to Asia (characterized by a transition at np 16356 [48]). Indeed, no exact match was found for the C1 haplotype in the comparative database of Eurasian populations (comprising 168,000 haplotypes), although 47 derivatives (showing one to three np differences) were found in extant populations broadly distributed throughout Eurasia (Table S1). Therefore, the C1 haplotype sequenced in aUzPo is currently uninformative about population affinity. In addition, all three aUzPo individuals showed identical C1 haplotypes, which meant that a close maternal kinship between these individuals could not be rejected. Biases due to the overestimation of the hg C1 frequency and small sample size of aUzPo may have led to an overestimation of the genetic affinity with modern-day West Siberians in the hg-based analyses. To account for this, we assumed a scenario of extreme maternal kinship, in which identical haplotypes found in several individuals at the same site (redundant haplotypes) were only counted once (Figure S2A). Under this scenario, the genetic affinity between aUzPo and present-day Western Siberians was less distinctly pronounced (Figure S2B). To further evaluate the apparent significant genetic discontinuity between aUzPo and modern extant populations of NEE and Saami, we analyzed Bayesian Serial SimCoal (BayeSSC) coalescent simulations [49] using Approximate Bayesian Computation (ABC, [50]) and tested whether discontinuity could be better explained by genetic drift or by migration. Models of genetic continuity between aUzPo and the present-day population of NEE or Saami (H0a) were compared to models in which genetic discontinuity between aUzPo and the extant population of NEE was introduced by migration (H1a, Figure 5). Ancestors of individuals from CE were selected as a source population for the migration on the basis of the PCA plot (Figure 2) showing that present-day populations of NEE shared the most genetic similarities with those of CE. The model of genetic discontinuity between aUzPo and present-day Saami was not tested since no source population for a potential migration could be identified from the PCA plot. The model of genetic continuity between aUzPo and present-day Saami was found to fit the observed data better than the model of genetic continuity between aUzPo and present-day NEE. This can be attributed to the low haplotypic diversities (0.74 and 0.81, respectively, in contrast to 0.98 for NEE; Table 2) of both aUzPo and Saami populations. The migration model provided a better fit for the genetic data than the model of genetic continuity (H0a), as indicated by a low Akaike Information Criterion (AIC, [51]) and a high Akaike weight ω [52]–[53]. The lowest AIC (Figure 5) and highest Akaike's ω (Table 3) were obtained for migration models, the best fit being obtained for the model involving 10% of migrants over the last 7,500 years (H1b; ω = 1.00E+0 as opposed to ω = 2.57E-7 for the continuity model H0a). Our analyses of coalescent simulations therefore supported a genetic discontinuity between aUzPo and the present-day population of NEE, which was better explained by a migration from CE than by genetic drift. At the 3,500 uncal. yBP site of aBOO, we observed 39% ‘European’ hgs: U5a (26%), U4 (9%), T (4%), and 61% ‘Central/East Siberian’ hgs: C (35%), Z (13%), D (13%). Concordant with this admixed hg make-up, PCA indicated a position close to present-day Siberians (Figure 2). This position did not change when potential maternal relationships among individuals were accounted for by excluding redundant haplotypes (Figure S2B). The genetic relationship between aBOO and Siberians was also evident on the genetic distance map, where the area representing the lowest genetic distance covered a broader area of Siberia than for aUzPo (Figure 3B). The extant populations that showed most genetic similarity to aBOO were found in Central and East Siberia. In contrast, the area of maximum similarity for aUzPo lay in West Siberia (Figure 3A); this observation however could be influenced by low sample size in aUzPo. Haplotype sharing analyses for aBOO confirmed the genetic affinity with modern-day West and Central/East Siberians inferred from the PCA (Figure 4), but also identified a close relationship with the VUB population pool. The distribution of haplotype matches observed in pools of the VUB, West Siberia and Central/East Siberia was partly due to the presence of basal C* (16223T-16298C-16327T) and D* (16223T-16362C) haplotypes in these pools, whereas these types were absent in Middle Eastern and European pools. Central Siberian Tuvinians displayed the highest percentage of shared haplotypes with aBOO (12.2%) although all shared haplotypes belong to hgs C* and D*. A more explicit genetic link between aBOO and extant East Siberians was seen in the presence of the derived C5 haplotype (16148T-16223T-16288C-16298C-16311C-16327T) in aBOO and in one single Buryat individual of Central Siberia [54]. The Z1a haplotype (16129A-16185T-16223T-16224C-16260T-16298C) detected in aBOO had a broad but interesting distribution in Eurasia. It was found in all Central/East Siberian pools except in Tuvinians, but also in the Bashkirs of the Urals, in the VUB pool, as well as in Scandinavian and Baltic populations (Norwegians, Swedes, Finns, Ingrians, Karelians, and the Saami). Although haplotype sharing analyses revealed genetic links between aBOO and extant populations of NEE, a strong genetic differentiation was obvious between aBOO, modern-day NEE and Saami. This genetic discontinuity was further supported by BayeSSC analyses (Figure 5; Table 3). Similarly to aUzPo, a better fit was obtained for the model involving a 10% migration from CE over the last 3,500 years (H1b; ω = 1.00E+0) than for the model of genetic continuity between aBOO and NEE (H0b; ω = 3.86E-10). Previously described populations of hunter-gatherers of Central/East Europe (aHG [12], [14]) and Scandinavia (aPWC, [13]) were characterized by high frequencies and diversity of hg U4, U5a and U5b, which caused the two ancient datasets to group outside the cluster of extant European populations on the PCA plot (Figure 2). This matches previous studies that have shown that genetic continuity between hunter-gatherers and present-day Europeans can be rejected [12]–[13]. Like other European hunter-gatherers, aUzPo is characterized by high frequencies and diversity of hgs U4 and U5, but was genetically differentiated from aHG and aPWC due to the occurrence of hg C. Despite the fact that high frequencies of hgs U5b and V cluster the aHG and aPWC hunter-gatherer groups on the PCA plot (Figure 2), and that these hgs are also common in modern-day Saami, the ‘Saami motif’ is absent from aPWC and genetic continuity between aPWC and modern-day Saami was rejected [13]. Although the aBOO individuals were also characterized by high frequencies of hg U, the group appeared less close to the Palaeolithic/Mesolithic hunter-gatherers aHG and aPWC on the PCA plot than aUzPo. Haplotype sharing analyses (Figure 6) also showed that aBOO shared less haplotypes with aHG and aPWC than aUzPo (4.76% and 0.00%, respectively, versus 9.52% and 36.84%). This observation was confirmed by the analyses of our coalescent simulations, in which a model of genetic continuity between aHG, aPWC and aUzPo (ω = 9.91 E-1; H0d) was better supported than a model of genetic continuity between aHG, aPWC and aBOO (ω = 1.10 E-4; H0e). As demonstrated above, aBOO exhibited greater genetic affinities with extant populations of Siberia than aUzPo. Accordingly, aBOO shared more haplotypes with ancient samples from Siberia aEG (10.87% [55]) and aKUR (7.69% [56]) than aUzPo (0.00% and 7.69%, respectively; Figure 6). To date, all studies on ancient Mesolithic/Palaeolithic hunter-gatherers from Europe have reported large proportions of hg U: 64% in aUzPo, 73% in aHG, 74% in aPWC; and hg U was also found in three out of five Mesolithic individuals of Spain [20], [57]. On the basis of the distribution of hg U5b, it was proposed that the Mesolithic population has remained genetically homogeneous over a wide geographical area and for a long period of time [57]. The new data from aUzPo suggests that hg U5a may be a representative of Central and North East Europe's Mesolithic mtDNA diversity, whereas elevated frequencies of hg U4 appear more characteristic of populations of the peri-Baltic area (aUzPo and aPWC). Haplogroup U also represents a significant genetic component of aBOO (35%), as well as Bronze Age Central Asians (14% in aKAZ; 2,700–3,400 yBP), and pre-Iron Age Siberians (54% in aKUR; Andronovo and Karasuk cultures; 2,800–3,800 yBP). Today, hg U is found in 7% of Europeans and displays a wide distribution in Europe, West Siberia, south west Asia, the Near East and North Africa [5]. Both the widespread distribution and high variability of hg U in extant and prehistoric populations are consistent with the description of hg U as one of the oldest hgs in Europe. On the basis of modern genetic data, hg U was proposed to have originated in the Near East and spread throughout Eurasia during the initial peopling by anatomically modern humans in the early Upper Palaeolithic (around 45,000 yBP, [5]). It is then plausible that hg U constituted the major part of the Palaeolithic/Mesolithic mtDNA substratum from Southern, Central and North East Europe to Central Siberia. It can also be suggested that the Palaeolithic/Mesolithic mtDNA substratum has been preserved longer in NEE than in Central and southern parts of Europe, where new lineages arrived with incoming farmers during the Neolithisation from the Near East [16]. This is supported by ancient genomic data obtained from hunter-gatherers of Scandinavia [58] and Spain [57], that shows a genetic affinity between Mesolithic individuals and present-day northern Europeans and supports genetic discontinuity between Mesolithic and Neolithic populations of Europe. The detection of haplogroup H in the Mesolithic site of aUz (one haplotype) is noteworthy. To date, haplogroup H has either been rare or absent in groups of hunter-gatherers previously described. It has not been found in hunter-gatherer mtDNA datasets of eastern Europe [12] and Scandinavia [13], but has been found in two hunter-gatherers of the Upper Palaeolithic sites of La Pasiega and La Chora in northern Spain [20]. The closest match to the ancient H haplotype in aUzPo belongs to sub-haplogroup H2a2 [59], which is more common in eastern Europe [60] with highest frequencies in the Caucasus. Current ancient data is too scarce to investigate the past phylogeography of haplogroup H in full detail. However, together with U4, U5 haplotypes this H haplotype suggests continuity of some maternal lineages in (North) East Europe since the Mesolithic. While the Mesolithic aUzPo site showed genetic affinities with extant populations of West Siberia in hg-based analyses, the precise genetic origins of aUzPo individuals was more difficult to identify from haplotypic data due to the high number of basal haplotypes. At the archaeological level also, the Siberian connection with aUzPo is less clear. The material culture present in the burials of aUz links these populations with the neighboring regions in the West but also in the East and South-East [26], [61]. As for Siberia, it has undergone a complicated early and mid-Holocene migration history due to repeated environmental changes [62]. With the data at hand, it is therefore difficult to make any definite statement about sixth millennium connections between Karelia and Siberia. Interestingly, samples from aBOO, which are 4,000 years younger and located further North-West than aUzPo, were characterized by a large proportion and elevated diversity of mtDNA lineages showing a clear ‘Central/East Siberian’ origin (hgs C, D, and Z). Haplogroups C and D are the most common hgs in northern, central and eastern Asia. They are thought to have originated in eastern Asia and expanded through multiple migrations after the Late Glacial Maximum (∼20,000 yBP [63]). Notably, haplotypic matches were observed between aBOO and modern-day central Siberian Buryats of the peri-Baikal region, which was proposed to be the origin of ancient migrations that disseminated hgs C and D [63]. Today, the sharp western boundary for the distribution of hgs C, D and Z lies in the VUB, where they display intermediate frequencies: C (0.3–11.8%), Z (0.2–0.9%), and D (0.6–12%) [64]. Sub-hgs Z1 and D5 are also present in modern-day Saami, with highest cumulated frequencies (15.9%) in the Saami of Finland, the easternmost part of the Saami geographical distribution [32]. A precise date for the arrival of these ‘Central/East Siberian’ lineages in NEE is difficult to estimate, although the presence of ‘Central/East Siberian’ lineages in the 3,500 year-old aBOO site indicates that an eastern genetic influence pre-dates historical westward expansions from Central/East Siberia of, e.g., the Huns and the Mongols (∼400–1,500 AD). We present here direct genetic evidence for a prehistoric gene-flow from Siberia. On the basis of modern-day genetic data, hg Z1 was proposed to have been introduced into populations of the VUB and Saami by migrations from Siberia via the southern Urals to the Pechora and Vychegda basins (northwest Urals), associated with the appearance of the Kama culture ∼8,000 yBP [22], [32]. The presence of hg Z1 in aBOO establishes a direct genetic link between aBOO and modern-day populations of the VUB and Saami, and possibly indicates the trajectory of the migration that brought ‘Central/East Siberian’ lineages into NEE. The fact that aBOO did not contain any other Saami-specific haplotypes, suggests an independent origin and contribution of Z1 to the Saami gene pool. The genetic links between the sample populations of aUzPo/aBOO and the extant populations of Siberia follow a general pattern discussed for the early and mid-Holocene (6,000–10,000 yBP). Facilitated by the East-West extension of vegetation zones between the Russian Far East and Eastern Europe [65], long-distance contacts and connections across Eurasia have been proposed for a number of cases. For example, the North East and East European hunter-gatherer pottery is thought to have originated in the early ceramic traditions of the Russian Far East and Siberia [66]–[68]. An eastern Asian origin followed by a westward expansion was also discussed for domesticated broomcorn millet (Panicum miliaceum L.) [69]. While the exact scenario behind these two examples of long-distance connections is unclear, migrations are a common interpretative model for evidence from later periods [70]. In any case, long-distance connections across Eurasia are not unusual. A later migration from the East was associated with the spread of the Imiyakhtakhskaya culture from Yakutia (East Siberia) through northwestern Siberia to the Kola Peninsula during the Early Metal Age (3,000–4,000 yBP, [24]). Interestingly, one individual of the aBOO site (grave 10, not sampled for aDNA here) was archaeologically associated with this culture, but its cultural relationships to other individuals of the same site remain unclear. The apparent genetic discontinuity between aUzPo and aBOO is consistent with craniometrical analysis that have proposed a genetic discontinuity between the two groups despite the finding of ‘caucasoid’ and unusual ‘mongoloid’ cranial features at both sites [28]. Samples of aBOO were also shown to display craniometrical affinities with ancient populations of West Siberia and the Altai, in line with the ancient genetic data presented here [42]. The ‘admixed’ nature of the aUzPo and aBOO populations is supported by the apparent random distribution of mtDNA lineages within the corresponding graveyards, i.e., there is no structure in the sites reflecting the ‘Western’ or ‘Eastern’ origins of the buried individuals [26]. The present-day Saami populations display clear haplotypic differences from all the ancient populations sampled for DNA so far (prehistoric hunter-gatherer populations of North/South/Central/East Europe, aUzPo and aBOO) where none of the hg V and U5b1b1a lineages distinctive of the Saami could be detected. We show here that the mitochondrial ancestors of the Saami could not be identified in the ancient NEE populations of aUzPo or aBOO, despite the latter site being within the area occupied by Saami today. The widespread modern-day distribution of U5b1 and V lineages makes it difficult to identify the origins of the Saami [32]. Sub-haplogroup U5b1b1 to which the ‘Saami motif’ belongs was proposed to have originated and spread from southern/central Europe after the Late Glacial Maximum [32]–[33]. Despite its clear association with Saami ancestry, the ‘Saami motif’ also occurs at low frequency (below 1%) in a wide range of non-Saami populations in Europe, and haplotypes closely related to the ‘Saami motif’ have even been found in modern Berbers of North Africa [33]. Two origins have been proposed on the basis of archaeological and genetic evidence [24], [32]. First, ancestors of the Saami were suggested to have reached Fennoscandia from Western Europe along the Atlantic cast of Norway as part of the expansion of Mesolithic post-Ahrensburgian cultures (Fosna-Hensbacka and Komsa) in the early Holocene (∼10,000–11,000 yBP). Alternatively, the Saami were proposed to find their origins in Mesolithic post-Swiderian cultures (Kunda, Veretye, Suomusjärvi), which had moved from Poland into NEE also in the early Holocene [24]. The data from aUzPo, in which neither U5b1 or V could be detected, does not support the latter hypothesis. If migrations brought U5b1 and V to Fennoscandia from the East, they must have occurred after 7,500 yBP or have had a weak genetic impact on surrounding populations of NEE. Saami mtDNA diversity has been influenced by a combination of founder event(s), (multiple) bottlenecks, and reproductive isolation, which are likely due to the challenging conditions of life in the subarctic taiga/tundra [32]. The complex demographic history of Saami renders their population history difficult to reconstruct on the basis of modern genetic data alone. Further temporal population samples will be required, especially along the proposed alternative western migration route into sub-arctic Europe. Individuals from 7,500 year-old aUzPo and 3,500 year-old aBOO show remarkable genetic dissimilarities with present-day North East Europeans: high frequencies of hg U, the presence of mtDNA lineages of ‘Central/East Siberian’ origin, and near absence (one out of 34 samples) of hg H which comprises up to ∼50% in extant European populations [5]. The results of our coalescent simulation analyses show that the models that take account of genetic input(s) from CE are better supported and could explain the genetic discontinuity observed between either aUzPo or aBOO and the modern population of NEE (Figure 5). The mtDNA lineages with a clear Central/Western European signature and currently prevalent in NEE might have reached the western Baltic and southern Scandinavia during the continuing influx of farming populations from Central or lastly southeastern Europe [13], [58], as from 6,000 yBP onwards [71]–[74]. However, intruding Neolithic farmers never reached Karelia and Fennoscandia [75], so the change in population would have to be a post-Neolithic process or to be due to migrations from other sources. The major prehistoric migration in this area was associated with the spread of early pottery from the East into the Baltic, Karelia and Fennoscandia starting around 7,000 yBP. This migration might have contributed to an early population change in Karelia and Fennoscandia as well, but the mtDNA characteristics of the populations involved is presently unknown [76]–[78]. As for Siberia, a general push-back of populations by an expansion of populations from the South-West is discussed [62]. Thus, the present-day distribution of populations similar to aUzPo and aBOO might just be a remnant of a once much larger extension across western and Central northern Eurasia, which is consistent with frequencies of hgs U4 and U5, i.e. the Palaeolithic/Mesolithic genetic substratum, have remained higher in extant populations of NEE, the VUB and Western Siberia than in central Europeans, where these were largely replaced at the onset of the Neolithic [16], [79]. Genetic discontinuity between aUzPo, aBOO and present-day populations of NEE was also observed at the haplotype level, as seen by the lack of matches between lineages from ancient individuals and from present-day NEE (e.g., ‘Central/East Siberian’ lineages in aBOO), or by their total absence in all Eurasian populations of the comparative dataset. A good example is the haplotype C1 found in aUzPo, which is absent in modern-day Eurasians and in all other foraging populations of Europe. This indicates that hg C1 was rare and probably preserved in aUzPo by a relative reproductive isolation, previously proposed for Mesolithic hunter-gatherers of NEE on the basis of odontometric [41] and craniometric [80] analyses. These results do not exclude a common origin for European foragers but highlight differentiating consequences of post-glacial founder effects followed by reproductive isolation among Palaeolithic/Mesolithic groups. Genetic discontinuity between prehistoric populations of Europe may have been caused by the random loss of genetic diversity through drift, which is likely to have been accelerated in small and isolated groups, such as aUzPo and aBOO. In the Kola Peninsula, the scarcity in the archaeological records observed in the Kola Peninsula for the Early Metal Age was interpreted as an indication of drastic size reductions of human groups, as a response to deteriorating climatic conditions ∼2,500 yBP [24]. This could have lead to the local extinction of mtDNA lineages of Siberian origin detected in aBOO in the Kola Peninsula. Overall, our results illustrate the power of aDNA to reconstruct the complex genetic history of NEE, which is made of past migrations from both Siberia and Europe. Ancient DNA also reveals the plasticity of demographic events in human populations at both the scale of NEE and Eurasia. Future accumulation of genetic data from ancient populations will make it possible to establish more genetic relationships between past human populations in space and time. A total of 146 human teeth—representing 74 individuals—were collected from three archaeological sites in northwestern Russia: Yuzhnyy Oleni Ostrov, Popovo, and Bolshoy Oleni Ostrov (under custody of the Kunstkamera Museum, St Petersburg, Russia; Figure S1, Table S2). The oldest samples were collected in the Mesolithic graveyards of Yuzhnyy Oleni Ostrov (aUz; ‘Southern Reindeer Island’ in Russian) and Popovo (aPo). Ninety-six teeth representing 48 individuals were obtained from the Yuzhnyy Oleni Ostrov archaeological site, which is located on Yuzhnyy Oleni Island, Onega Lake, Karelia (61°30′N 35°45′E). The site was first discovered in the 1920s during quarrying excavations, which led to the subsequent destruction of most parts of the graveyard. Scientific excavation of the site by Soviet archaeologists in the 1930s and the 1950s eventually unearthed a total of 177 individuals in 141 different mortuary features [81]. The population size of the burial ground before its partial destruction was estimated at around 500 individuals [82]. The Yuzhnyy Oleni Ostrov graveyard stands out from other Mesolithic sites in Europe by its abundance and diversity of mortuary features. First identified as a Neolithic graveyard, a later reanalysis and radiocarbon dating revealed an age of around 7,000–7,500 uncal. yBP [83]. For Popovo, 6 teeth belonging to 3 individuals were obtained from the archaeological site located on the bank of the Kinema River, in the Archangelsk region (64°32′N 40°32′E). The wide range of dates obtained for this site (9,000–9,500 uncal. yBP and 7,500–8,000 uncal. yBP [84]). We expect that the radiocarbon dates for both the sites of Popovo and Yuzhnyy Oleni Ostrov will be revised, as potential freshwater-derived reservoir effects impacting the dates are currently investigated (T. Higham, personal communication). The sites of aUz and aPo are located along one of the proposed eastern routes for the introduction of Saami-specific mtDNA lineages [32]. Results from odontometric analyses suggested a direct genetic continuity between the Mesolithic population of Yuzhnyy Oleni Ostrov and present-day Saami [41]. Due to the small sample size, and the temporal and geographic proximity of aPo and aUz, the specimens from these sites were pooled for statistical analyses (aUzPo). We also analyzed human remains from the Early Metal Age archaeological site of Bol'shoy Oleni Ostrov (aBOO; ‘Great Reindeer island’ in Russian) in the Kola Peninsula. This site is located within the area currently inhabited by Saami individuals. Fourty-five teeth representing 23 individuals were obtained from this archaeological site, located in the Murmansk region, Kola Peninsula (68°58′N 33°05′E). Several excavation campaigns have been undertaken between 1927 and 2006. Radiocarbon dates for two graves were obtained from the Oxford Radiocarbon Accelerator Unit, United Kingdom, and revealed an age of around 3237±32 yBP (calibrated dates in years before 1950, 3525–3440 BC (68.2%) and 3610–3420 BC (95.4%)) and 3195±39 yBP; calibrated dates, 3500–3430 BC (68.2%) and 3530–3390 BC (95.4%)) for grave 12 and grave 13, respectively, corresponding to the Early Metal Age. The organic preservation of artifacts made of bone, antlers and wood in this site is exceptional for this time period and geographical location [42]. DNA isolation, amplification and quantitation were performed at the aDNA laboratory of the Australian Centre for Ancient DNA (ACAD), University of Adelaide. Whenever possible, two distinct teeth were analyzed for each ancient individual. The outer surface of each tooth was decontaminated, first, through exposure to ultra-violet (UV) light for 20 min on each side, then, through gentle wiping using a paper towel soaked in sodium hypochlorite (bleach). The protocol described in [85] was followed to isolate DNA from powdered teeth. Given the archaeological and anthropological value of the samples from aUz, aPo and aBOO, their morphological integrity had to be maintained: tooth powder was collected by cutting off the crown of each tooth and drilling inside the root using a dental drill at low speed. Collecting material from only the dental pulp and dentin may prevent the risk of contamination by exogenous DNA, as the inside of the teeth may be protected from the environment by the enamel. The mtDNA HVR-I was amplified and sequenced between np 16056 and 16410 as described in [85]. The GenoCore22 reaction described in [85] was used to type 22 haplogroup-diagnostic SNPs in the mtDNA coding-region (Table S3). Twenty-two fragments of mtDNA were amplified simultaneously in a multiplex reaction and SNPs were detected using Single-Base Extension (SNaPshot kit, Applied Biosystems). The copy-number of two HVR-I fragments - L16209/H16303 (133 bp) and L16209/H16348 (179 bp) - was estimated in selected aDNA extracts (individuals UZOO-43, UZOO-79, BOO72-1, and BOO72-9) by quantitative real-time PCR following the protocol detailed in [16] (Table S4). Six individuals were randomly selected (UZOO-77, BOO57-1, BOO72-1, BOO72-4, BOO72-7, and BOO72-15), for which the second sample was sent to G.B. at the Johannes Gutenberg University of Mainz for independent replication of DNA extraction, HVR-I amplification and direct sequencing. PCR products were cloned and sequenced. Ancient DNA work at the Johannes Gutenberg University was carried out according to protocols described in [85]. Strict precautions were taken in order to minimize the risk of contamination by modern DNA and detect artefactual mutations arising from contamination and aDNA degradation. Seven criteria support the authenticity of the mtDNA data presented here. Mitochondrial DNA data from aUzPo and aBOO were compared to data obtained from other ancient and present-day populations. Data for extant populations were compiled in the MURKA mtDNA database and integrated software, which currently contains 168,000 HVR-I records from published studies and is curated by co-authors V. Z., O.B. and E.B. of the Russian Academy of Medical Sciences, Moscow. A sub-sample of 91 ancient and modern Eurasian populations (∼28,652 individuals) was used for comparative analysis. Names of modern-day populations were abbreviated using ISO codes in capital letters, and in small letters when ISO codes were not available. Unless specified otherwise, the same population labels were used for all the maps and analyses in this study, i.e., PCA, haplotype sharing and analysis of coalescent simulations (Table S5). PCA was performed using the hg frequency database for ancient and modern-day populations described in Table S5. We used a total of 19 variables to perform the PCA. Seventeen of these variables were frequencies of hgs C, D, H, HV, I, J, K, N1, T, U2, U4, U5a, U5b, V, W, X, and Z. In addition, the frequencies of six ‘east Eurasian’ hgs were pooled into one ‘EAS’ group including hgs A, B, E, F, G, and Y. Finally, frequencies of eight hgs found at lower frequencies in Eurasia were pooled into the ‘misc’ group including hgs L, M*, N*, U1, U6, U7, U8. By pooling and removing rare hgs (with frequencies below 1%) we could reduce statistical noise. In order to assess the impact of potential maternal kinship within the sites of aUzPo and aBOO, we performed an additional PCA, in which redundant haplotypes, i.e. haplotypes found in more than one individual at a given site, were counted only once (Figure S2A and Figure S2B). PCA was carried out using a customized script based on the function prcomp in R version 2.12 (R Development Core Team, http://www.R-project.org). The genetic distances between 144 pools of extant Eurasian populations and each of the aUzPo and aBOO populations were calculated using the software DJ (written by Yuri Seryogin, and freely available at http://genofond.ru, also see [29]). The software GeneGeo written by S.K. was used to plot genetic distances onto geographic maps (as described in [16]). In haplotype sharing analyses, we calculated the percentages of shared haplotypes between 29 extant populations and the ancient populations of aUzPo and aBOO. A database of mtDNA haplotypes was collated for modern-day populations, each containing 500 individuals. We pooled populations of less than 500 individuals on the basis of their geographical and/or linguistic similarities. For extant populations of more than 500 individuals, we randomly sub-sampled 500 individuals from the populations. For each haplotype of aUzPo and aBOO, we counted the number of haplotype matches found in each of the extant populations of the comparative database. This number was divided by the sample size in order to obtain the percentage of shared haplotypes. The same procedure was applied to calculate the percentage of shared haplotypes between the ancient populations studied here (aUzPo and aBOO) and previously described ancient populations. Percentages of shared haplotypes between ancient and present-day populations were represented in a bar plot. Percentages of shared haplotypes among ancient populations were represented in a table, the cells of which were colored according to a gradient reflecting the haplotypic similarities between the populations compared. In coalescent simulation analyses we considered the ancient populations of aUzPo, aBOO, Central/East/Scandinavian European hunter-gatherers (aHG [12], [14], aPWC [13]), and the modern populations of NEE, CE, and Saami (saa). Population statistics (haplotype diversity and fixation indexes, FST) for the ancient and extant populations were calculated in Arlequin version 3.11 (Table 2, [91]). In BayeSSC [49], genealogies were simulated under the following model of sequence evolution: a generation time of 25 years, a mutation rate of 7.5.10−6 substitutions/per site/per generation [92], a transition/transversion ratio of 0.9841, and parameters for the gamma distribution of rates along the sequence of 0.205 (theta) and 10 (kappa) [16]. Under the models of genetic continuity H0, the effective population size (Ne) of a single population was allowed to grow exponentially. The values of the growth rate were drawn from a uniform prior distribution, such that the population has evolved from a Palaeolithic population of Ne 5,000 that lived 1,500 generations ago. The values for the modern-day (NEE or saa) Ne were drawn from a uniform distribution: we explored present-day Ne between 100,000 and 30,000,000 for NEE and 1,000 to 500,000 for saa. Population statistics were estimated at various points in time, corresponding to the age of the ancient populations considered in models H0a to H0e (aUzPo, aBOO, aHG and aPWC). Under the models of migration H1, we assumed a single NEE population undergoing an exponential growth in Ne and being the recipient (sink population) of a migration from CE (source population). Population sizes of each of the present-day sink population (NEE) and source population (CE) were drawn from a uniform distribution of Ne varying from 100,000 to 15,000,000 individuals. Migration and divergence times were estimated from uniform distributions (from 2 to 139 generations for migration and from 620 to 2,600 generations for divergence). Three different percentages of the source population size were tested for the value of percentages of migrants: 10%, 50% and 75%. Population statistics were calculated for 100,000 genealogies simulated using BayeSSC (available at http://www.stanford.edu/group/hadlylab/ssc/index.html). The distribution of six selected population statistics (haplotype diversity and fixation indices FST) were drawn from the simulations and compared to the corresponding observed population statistics in an ABC framework [50], [93]. The 1% of the simulations for which simulated population statistics exhibited the smallest Euclidian distance with observed population statistics was retained to construct posterior distributions of population parameters. From these distributions, values of population parameters that optimized the likelihood of a given model were estimated and used in replacement of priors in demographic models. We finally generated 10,000 genealogies in BayeSSC for these models. BayeSSC outputs were analyzed in R version 2.12 using scripts available on request at http://www.stanford.edu/group/hadlylab/ssc/index.html. Goodness of fit of the different models tested was compared using AICs [51] and Akaike's weights ω [52]–[53] (Table 3). The Genbank accession numbers for the 34 mtDNA sequences reported in this paper are KC414891–KC414924.

10.1371/journal.pcbi.1005962

A comprehensive and quantitative comparison of text-mining in 15 million full-text articles versus their corresponding abstracts

Across academia and industry, text mining has become a popular strategy for keeping up with the rapid growth of the scientific literature. Text mining of the scientific literature has mostly been carried out on collections of abstracts, due to their availability. Here we present an analysis of 15 million English scientific full-text articles published during the period 1823–2016. We describe the development in article length and publication sub-topics during these nearly 250 years. We showcase the potential of text mining by extracting published protein–protein, disease–gene, and protein subcellular associations using a named entity recognition system, and quantitatively report on their accuracy using gold standard benchmark data sets. We subsequently compare the findings to corresponding results obtained on 16.5 million abstracts included in MEDLINE and show that text mining of full-text articles consistently outperforms using abstracts only.

Text mining has become an integral part of all fields in science. Owing to the large number of articles published every day, it is necessary to employ automated systems to assist in curation, knowledge management and discovery. To date, most systems make use of information collected from abstracts only. Moreover, studies on smaller collections of abstracts and full-text articles have demonstrated some information is available only in the full-text body. Nonetheless, to date there has been no large-scale comprehensive comparison of abstracts and full-text articles. In this work, we analyze a hitherto unprecedented collection of 15 million full-text articles. Through quantitative benchmarks we assess the difference between full-text articles and abstracts. Our findings confirm what has long been discussed, namely that access to the full-text body improved text mining greatly.

Text mining has become a widespread approach to identify and extract information from unstructured text. Text mining is used to extract facts and relationships in a structured form that can be used to annotate specialized databases, to transfer knowledge between domains and more generally within business intelligence to support operational and strategic decision-making [1–3]. Biomedical text mining is concerned with the extraction of information regarding biological entities, such as genes and proteins, phenotypes, or even more broadly biological pathways (reviewed extensively in [3–9]) from sources like scientific literature, electronic patient records, and most recently patents [10–13]. Furthermore, the extracted information has been used as annotation of specialized databases and tools (reviewed in [3,14]). In addition, text mining is routinely used to support manual curation of biological databases [15,16]. Thus, text mining has become an integral part of many resources serving a wide audience of scientists. The main text source for scientific literature has been the MEDLINE corpus of abstracts, essentially due to the restricted availability of full-text articles. However, full-text articles are becoming more accessible and there is a growing interest in text mining of complete articles. Nevertheless, to date no studies have presented a systematic comparison of the performance comparing a very large number of abstracts and full-texts in corpora that are similar in size to MEDLINE. Full-text articles and abstracts are structurally different [17]. Abstracts are comprised of shorter sentences and very succinct text presenting only the most important findings. By comparison, full-text articles contain complex tables, display items and references. Moreover, they present existing and generally accepted knowledge in the introduction (often presented in the context of summaries of the findings), and move on to reporting more in-depth results, while discussion sections put the results in perspective and mention limitations and concerns. The latter is often considered to be more speculative compared to the abstract [3]. While text-mining results from accessible full-text articles have already become an integral part of some databases (reviewed recently for protein-protein interactions [18]), very few studies to date have compared text mining of abstracts and full-text articles. Using a corpus consisting of ~20,000 articles from the PubMed Central (PMC) open-access subset and Directory of Open Access Journals (DOAJ), it was found that many explicit protein–protein interactions only are mentioned in the full text [19]. Additionally, in a corpus of 1,025 full-text articles it was noticed that some pharmacogenomics associations are only found in the full text [20]. One study using a corpus of 3,800 articles with focus on Caenorhabditis elegans noted an increase in recall from 45% to 95% when including the full text [21]. Other studies have worked with even smaller corpora [17,22,23]. One study have even noted that the majority of claims within an article is not reported in the abstract [24]. Whilst these studies have been of significant interest, the number of full-text articles and abstracts used for comparison are nowhere near the magnitude of the actual number of scientific articles published to date, and it is thus unclear if the results can be generalized to the scientific literature as a whole. The earlier studies have mostly used articles retrieved from PMC in a structured XML file. However, full-text articles received or downloaded directly from the publishers often come in the PDF format, which must be converted to a raw unformatted text file. This presents a challenge, as the quality of the text mining will depend on the proper extraction and filtering of the unformatted text. A previous study dealt with this by writing custom software taking into account the structure and font of each journal at that time [21]. More recent studies typically provide algorithms that automatically determines the layout of the articles [25–27]. In this work, we describe a corpus of 15 million full-text scientific articles from Elsevier, Springer, and the open-access subset of PMC. The articles were published during the period 1823–2016. We highlight the possibilities by extracting protein–protein associations, disease–gene associations, and protein subcellular localization from the large collection of full-text articles using a Named Entity Recognition (NER) system combined with a scoring of co-mentions. We quantitatively report the accuracy and performance using gold standard benchmark data sets. Lastly, we compare the findings to corresponding results obtained on the matching set of abstracts included in MEDLINE as well as the full set of 16.5 million MEDLINE abstracts. The MEDLINE corpus consists of 26,385,631 citations. We removed empty citations, corrections and duplicate PubMed IDs. For duplicate PubMed IDs we kept only the newest entry. This led to a total of 16,544,511 abstracts for text mining. The PubMed Central corpus comprises 1,488,927 freely available scientific articles (downloaded 27th January 2017). Each article was retrieved in XML format. The XML file contains the article divided into paragraphs, article category and meta-information such as journal, year published, etc. Articles that had a category matching Addendum, Corrigendum, Erratum or Retraction were discarded. A total of 5,807 documents were discarded due to this, yielding a total of 1,483,120 articles for text mining. The article paragraphs were extracted for text mining. No further pre-processing of the text was done. The journals were categorized according to categories (described in the following section) by matching the ISSN number. The number of pages for each article was also extracted from the XML, if possible. The Technical Information Center of Denmark (DTU Library) TDM corpus is a collection of full-text articles from the publishers Springer and Elsevier. The corpus covers the period from 1823 to 2016. The corpus comprises 3,335,400 and 11,697,096 full-text articles in PDF format, respectively. An XML file containing meta-data such as publication date, journal, etc. accompanies each full-text article. PDF to TXT conversion was done using pdftotext v0.47.0, part of the Poppler suite (poppler.freedesktop.org). 192 articles could not be converted to text due to errors in the PDF file. The article length, counted as the number of pages, was extracted from the XML file. If not recorded in the XML file we counted the number of pages in the PDF file using the Unix tool pdfinfo v0.26.5. Articles were grouped into four bins, determined from the 25%, 50%, and 75% quantiles, respectively. These were found to be 1–4 pages (0–25%), 5–7 pages (25–50%), 8–10 pages (50–75%) and 11+ pages (75%-100%). Each article was, based on the journal where it was published, assigned to one or more of the following seventeen categories: Health Sciences, Chemistry, Life Sciences, Engineering, Physics, Agriculture Sciences, Material Science and Metallurgy, Earth Sciences, Mathematical Sciences, Environmental Sciences, Information Technology, Social Sciences, Business and Economy and Management, Arts and Humanities, Law, Telecommunications Technology, Library and Information Sciences. Due to the large number of categories, we condensed anything not in the top-6 into the category “Other”. The top-six categories health science, chemistry, life sciences, engineering, physics and agricultural sciences make up 74.8% of the data (S1 Fig). The assignment of categories used in this study was taken from the existing index for the journal made by the librarians at the DTU Library. For the temporal statistics, the years 1823–1900 were condensed into one. Following the PDF-to-text conversion of the Springer and Elsevier articles we ran a language detection algorithm implemented in the python package langdetect v1.0.7 (https://pypi.python.org/pypi/langdetect). We discarded 902,415 articles that were not identified as English. We pre-processed the remaining raw text from the articles as follows: We assumed that acknowledgments and reference lists are always at the end of the article. Upon encountering either of the terms: “acknowledgment”, “bibliography”, “literature cited”, “literature”, “references”, and the following misspellings thereof: “refirences”, “literatur”, “références”, “referesces”. In some cases the articles had no heading indicating the start of a bibliography. We tried to take these cases into account by constructing a RegEx that matches the typical way of listing references (e.g. [1] Westergaard, …). Such a pattern can be matched by the RegEx “^\[\d+\]\s[A-Za-z]”. The other commonly used pattern, “1. Westergaard, …”, was avoided since it may also indicate a new heading. Keywords were identified based on several rounds of manual inspection. In each round, 100 articles in which the reference list had not been found was randomly selected and inspected. We were unable to find references in 286,287 and 2,896,144 Springer and Elsevier articles, respectively. Manual inspection of 100 randomly selected articles revealed that these articles indeed did not have a reference list or that the pattern was not easily describable with simple metrics, such as keywords and RegEx. Articles without references were not discarded. The PDF to text conversion often breaks up paragraphs and sentences, due to new page, new column, etc. Paragraph and sentence splitting was performed using a ruled-based system. If the previous line of text does not end with a “.!?”, and the current line does not start with a lower-case letter, it is assumed that the line is part of the previous sentence. Otherwise, the line of text is assumed to be a new paragraph. A number of Springer and Elsevier documents were removed due to technical issues post pre-processing. An article was removed if: Some PDF files without texts are scans of the original article (point 1). We did not attempt to make an optical character recognition conversion (OCR) as the old typesetting fonts often are less compatible with present day OCR programs, and this can lead to text recognition errors [28,29]. For any discarded document, we still used the meta-data to calculate summary statistics. In some cases the PDF to text conversion failed, and produced non-sense data with a white space between the characters of a majority of the words (point 2). To empirically determine a cutoff we gradually increased the cutoff and repeatedly inspected 100 randomly selected articles. At the 2% quantile we saw no evidence of broken text. Articles with the following keywords in the article were discarded: Author Index, Key Word Index, Erratum, Editorial Board, Corrigendum, Announcement, Books received, Product news, and Business news (point 3). These keywords were found as part of the process of identifying acknowledgments and reference lists. Further, any article that was available through PubMed Central was preferentially selected by matching doi identifiers. This left a total of 14,549,483 full-text articles for further analysis. Some articles were not separable, or were subsets of others. For instance, conference proceedings may contain many individual articles in the same PDF. We found 1,911,365 articles in which this was the case. In these cases we removed the duplicates, or the shorter texts, but kept one copy for text mining. In total, we removed 898,048 duplicate text files. The majority of articles had a separate abstract. We matched articles from PubMed Central to their respective MEDLINE abstract using the PMCID to PubMed ID conversion file available from PMC. Articles from Springer and Elsevier typically had a separate abstract in the meta-data. Any abstract from an article that was part of the 1,911,365 articles that could not be separated was removed. This led to a total of 10,376,626 abstracts for which the corresponding full text was also included downstream, facilitating a comparative analysis. References for the full text articles analyzed can be found at 10.6084/m9.figshare.5419300. An article is preferentially referenced by its Digital Object Identifier (DOI) (98.8%). However, if that was not available, we used the PubMed Central ID for PMC articles (0.005%), or the list of authors, article title, journal name, and year. (0.006%) We performed text mining of the articles using a Named Entity Recognition (NER) system, described earlier [30–33]. The software is open source and can be downloaded from https://bitbucket.org/larsjuhljensen/tagger. The NER approach is dictionary based, and thus depends on well-constructed dictionaries and stop word lists. We used the gene names from the STRING dictionary v10.0 [30], disease names from the Disease Ontology (DO) [34] and compartment names from the Gene Ontology branch cellular component [35]. Stop word lists were all created and maintained in-house. Pure NER based approaches often struggles with ambiguity of words. Therefore, we included additional dictionaries that we do not report the results from. If any identified term was found in multiple dictionaries, it was discarded due to ambiguity. The additional dictionaries include small molecule names from STITCH [36], tissue names from the Brenda Tissue Ontology [37], Gene Ontology biological process and molecular function [35], and the mammalian phenotype ontology [38]. The latter is a modified version made to avoid clashes with the disease ontology. The dictionaries can be downloaded from https://doi.org/10.6084/m9.figshare.5827494. In the cases where the dictionary was constructed from an ontology co-occurrences were backtracked through all parents. E.g. the term type 1 diabetes mellitus from the Disease Ontology is backtracked to its parent, diabetes mellitus, then to glucose metabolism disease, etc. Co-occurrences were scored using the scoring system described in [39]. In short, a weighted count for each pair of entities (e.g. disease-gene) was calculated using the formula, C(i,j)=∑k=1nwdδdk(i,j)+wpδpk(i,j)+wsδsk(i,j) (1) where δ is an indicator function taking into account whether the terms i,j co-occur within the same document (d), paragraph (p), or sentence (s). w is the co-occurrence weight here set to 1.0, 2.0, and 0.2, respectively. Based on the weighted count, the score S(i,j) was calculated as, S(i,j)=Cijα(CijC..Ci.C.j)1−α (2) where α is set to 0.6. All weights were optimized using the KEGG pathway maps as benchmark (described further below). The S scores were converted to Z scores, as described earlier [40]. PPIs were benchmarked using pathway maps from the KEGG database [41–43]. Any two proteins in the same pathway were set to be a positive example, and any two proteins present in at least one pathway, but not the same, were set as a negative example. This approach assumes that the pathways are near complete and includes all relevant proteins. The same approach has been used for the STRING database [39]. The disease–gene benchmarking set was created by setting the disease-gene associations from UniProt [44] and Genetics Home Reference (https://ghr.nlm.nih.gov/, accessed 23th March 2017) as positive examples. The positive examples were then shuffled, and the shuffled examples were set as negative examples. Shuffled examples that ended up overlapping with the positive examples were discarded. This approach has previously been described [31]. The protein–compartment benchmark set was created by extracting the compartment information for each protein from UniProt and counting these as positive examples. For every protein found in at least one compartment, all compartments where it was not found were set as negative examples. The same approach has been used previously [33]. Receiver Operating Characteristic (ROC) curves were created by gradually increasing the Z-score and calculating the True Positive Rate (TPR) and False Positive Rate (FPR), as described in eqs (3) and (4). We compare the ROC curves by the Area Under the Curve (AUC), a metric ranging from 0 to 1. ROC-AUC curves provide a quantitative way of comparing benchmarks of classifiers, and is often used in machine learning and text mining. A perfect classifier will have an AUC = 1, and a classifier that performs equal to or worse than random will have an AUC ≤ 0.5. Individual mentions of entities used for the benchmark in each article can be downloaded from 10.6084/m9.figshare.5620417. We analyzed and compared four different corpora comprising all full-text articles (14,549,483 articles, All Full-texts), full-text articles that had a separate abstract (10,376,626 articles, Core Full-texts), the abstract from the full-text articles (10,376,626 abstracts, Core Abstracts), and the MEDLINE corpus (16,544,511 abstracts, MEDLINE) (see Methods for a detailed description of the pre-processing). The growth of the data set over time is of general interest in itself, however, it is also important to secure that the concepts used in the benchmarks are likely to be present in a large part of the corpus. We found that the number of full-text articles has grown exponentially over a long period (Fig 1A, a log-transformed version is provided in S2 Fig). We also observed that the growth represents a mixture of two components: one from 1823–1944, and another from 1945–2016. Through linear regression of the log2-transformed counts for the period 1945–2016 we found that the growth rate is 0.103 (p < 2 * 10−16, R2 = 0.95). Thus, the doubling time for the full-text corpus is 9.7 years. In comparison, MEDLINE had a growth rate of 0.195 (p < 2 * 10−16, R2 = 0.91) and a doubling time of 5.1 years. We noticed that there was a drop in the number of full-text publications around the years 1914–1918 and 1940–1945. Likewise, we see a decrease in the number of publications indexed by MEDLINE in the entire period 1930–1948. In the full-text corpora we found a total of 12,781 unique journal titles. The most prevalent journals are tied to health or life sciences, such as The Lancet, Tetrahedron Letters, and Biochemical and Biophysical Research Communications, or the more broad journals such as PLoS ONE (see S1 Table for the top-15 journals). The Lancet publishes only very few articles per issue, it was established in 1823 and has been active in publishing since then, thus explaining why it so far has nearly published 400,000 articles. In contrast, PLoS ONE was launched in 2006, and has published more than 172,000 articles. Of the 12,781 journal titles, 6,900 had one or more category labels assigned by librarians at the Technical University of Denmark. The vast majority of the full-texts, 13,343,040, were published in journals with one or more category labels. The frequency of each category within the corpus can be seen in S1 Fig. We observed that before the 1950’s health science dominated and made up almost 75% of all publications (Fig 1B). At the start of the 1950’s the fraction started to decrease, and to date health science makes up approximately 25% of all publications in the full-text corpus. Inspecting the remaining eleven categories in a separate plot we found that there was no single category that was responsible for the growth (S3 Fig). We binned the full-text articles into four categories based on the number of pages (see Methods). The average length of articles has increased considerably during the almost 250 years studied (Fig 1C). Whereas 75% of the articles were 1–3 pages long at the end of the 20th century, less than 25% of the articles published after year 2000 are that short. Conversely, articles with ten or more pages only made up between 0.7%-7% in the 19th century, a level that had grown to 20% by the start of the 21st century. We also observed that the average number of mentioned entities changed over time (S4 Fig). Mentions of genes and compartments were nearly non-existing prior to 1950, and has been increasing at an exponential rate since year 2000. Moreover, disease mentions dropped around year 1950, which correlates well with the decreasing proportion of published articles from health science journals in our corpus (Fig 1C). We ran the textmining pipeline on the two full-text and two abstract corpora. In all cases we found that the AUC-value was far greater than 0.5, from which we conclude that the results were substantially better than random (Fig 2) (see Methods for a definition of the AUC). The biggest gain in performance when using full-text was seen in finding associations between diseases and genes (AUC increase from 0.85 to 0.91) (Table 1). Compared to MEDLINE, the traditional corpus used for biomedical text mining, there was an increase in the AUC from 0.85 to 0.91. The smallest gain was associations between proteins, which increased from 0.70 to 0.73. Likewise, the Core Full-texts always performed better than Core Abstracts, signifying that some associations are only reported in the main body of the text. Consequently, traditional text mining of abstracts will never be able to find this information. All Z-scores used for benchmarking can be downloaded from https://doi.org/10.6084/m9.figshare.5340514. It has previously been speculated if text mining of full-text articles may be more difficult and lead to an increased rate of false positives [3]. To investigate this we altered the weights of the scoring system (See Methods, Eqs 1 and 2). The scoring scheme used here has weights for within sentence, within paragraph and within document co-occurrences (see Methods). When setting the document weight to zero versus using the previously calibrated value found in an earlier study we found that having a non-zero small value does indeed improve extraction of known facts in all cases (S5 Fig) [33]. We observed that the increase in AUC is less than when using a lower document weight (S2 Table). In one case, protein–protein associations, the MEDLINE abstract corpus outperforms the full-text articles. Abstracts are generally unaffected by the document weight, mainly because abstracts are almost always one paragraph. Overall, the difference in performance gain is largest for full-texts and lowest for abstracts and MEDLINE. Hence, all the full-text information is indeed valuable and necessary. For practical applications, it is often necessary to have a low False Positive Rate (FPR). Accordingly, we evaluated the True Positive Rate (TPR) of the different corpora at the 10% FPR (TPR@10%FPR) (Fig 3). We found that full-texts have the highest TPR@10%FPR for disease-gene associations (S2 Table). When considering protein–protein associations and protein-compartment associations, full-texts perform equivalently to Core Abstracts and Core Full-texts. The result was similar to when we evaluated the AUC across the full range, removing the document weight has the biggest impact on the full-texts (S5 Fig and S6 Fig), while abstracts remain unaffected. We have investigated a unique corpus consisting of 15 million full-text articles and compared the results to the most commonly used corpus for biomedical text mining, MEDLINE. We found that the full-text corpus outperforms the MEDLINE abstracts in all benchmarked cases, with the exception of TPR@10%FDR for protein–compartment associations. To our knowledge, this is the largest comparative study to date of abstracts and full-text articles. We envision that the results presented here can be used in future applications for discovering novel associations from mining of full-text articles, and as a motivation to always include full-text articles when available and to improve the techniques used for this purpose. The corpus consisted of 15,032,496 full-text documents, mainly in PDF format. 1,504,674 documents had to be discarded for technical reasons, primarily because they were not in English. Further, a large number of documents were also found to be duplicates or subsets of each other. On manual inspection we found that these were often conference proceedings, collections of articles etc., which were not easily separable without manual curation. We also managed to identify the list of references in the majority of the articles thereby reducing some repetition of knowledge that could otherwise lead to an increase in the false positive rate. We have encountered and described a number of problems when working with full-text articles converted from PDF to TXT from a large corpus. However, the majority of the problems did not stem from the PDF to TXT conversion, which could potentially be solved using a layout aware conversion tool. Examples include PDFX [25], SectLabel [26] and LA-PDFText [27] of which the first is not practical for very large corpora as it only exists as an online tool. Nonetheless, to make use of the large volume of existing articles it is necessary to solve these problems. Having all the articles in a structured XML format, such as the one provided by PubMed Central, would with no doubt produce a higher quality corpus. This may in turn further increase the benchmark results for full-text articles. Nevertheless, the reality is that many articles are not served that way. Consequently, the performance gain we report here should be viewed as a lower limit as we have sacrificed quality in favor of a larger volume of articles. The solutions we have outlined here will serve as a guideline and baseline for future studies. The increasing article length may have different underlying causes, but one of the main contributors is most likely increased funding to science worldwide [45,46]. Experiments and protocols are consequently getting increasingly complex and interdisciplinary–aspects that also contribute to driving meaningful publication lengths upward. The increased complexity has also been found to affect the language of the articles, as it is becoming more specialized[47]. Moreover, we observed a steep increase in the average number of mentions of genes and compartments. This finding can most likely be attributed to recent developments in molecular biology, such as the sequencing of the human genome, Genome Wide Association Studies (GWAS), and other high-throughput technologies in ‘omics [48,49]. It was outside the scope of this paper to go further into socio-economic impact. We have limited this to presenting the trends from what could be computed from the meta-data. Previous papers are–in terms of benchmarking–only making qualitative statements about the value of full-text articles as compared to text in abstracts. In one paper a single statement is made on the potential for extracting information, but no quantitative evidence is presented [50]. In a paper targeting pharmacogenomics it is similarly stated that that there are associations that only are found in the full-text, but no quantitative estimates are presented [20]. In a paper analyzing around 20,000 full-text papers a search for physical protein interactions was made, concluding that these contain considerable higher levels of interaction [19]. Again, no quantitative benchmarks were made comparing different sources. In this paper, we have made a detailed comparison of four different corpora that provides a strong basis for estimating the added value of using full-text articles in text mining workflows. We have used quite difficult, but still well established benchmarks, to illustrate the differences in performance when comparing text mining of abstracts to full-text articles. Within biology, and specifically in the area of systems biology, macromolecular interactions and the relationships between genes, tissues and diseases are key data that drive modeling and the analysis of causal biochemical mechanisms. Knowledge of interactions between proteins is extremely useful when revealing the components, which contribute to mechanisms in both health and disease. As many biological species from evolution share protein orthologs, their mutual interactions can often be transferred, for example from an experiment in another organism to the corresponding pair of human proteins where the experiment has not yet been performed. Such correspondences can typically be revealed by text mining as researchers in one area often will not follow the literature in the other and vice versa. The results presented here are purely associational. Through rigorous benchmarking and comparison of a variety of biologically relevant associations, we have demonstrated that a substantial amount of relevant information is only found in the full body of text. Additionally, by modifying the document weight we found that it was important to take into account the whole document and not just individual paragraphs. The improvement in AUC that we present here were not overwhelming. One reason could be that associations have a higher probability of being curated if they are mentioned in the abstract. Moreover, most tools are geared towards abstracts. Thus, what we present is a lower limit on the performance gain. Consequently, as more full-text articles become available and text-mining methods improve, the quantitative benchmarks will improve. However, because the literature is highly redundant diminishing returns in terms of performance gain should be expected when adding evermore text. Event-based text mining will be the next step for a deeper interpretation and extending the applicability of the results [5]. With more development it may also be possible to extract quantitative values, as has been demonstrated for pharmacokinetics [51]. Other work is also going into describing the similarity between terms, and how full-text articles can augment this [52]. However, this is beyond the scope of this article. The Named Entity Recognition (NER) system used depends heavily on the dictionaries and stop word lists. A NER system is also very sensitive to ambiguous words. To combat this we have used dictionaries from well-known and peer-reviewed databases, and we have included other dictionaries to avoid ambiguous terms. Other approaches to text mining have previously been extensively reviewed [10,14,51]. The full-text corpus presented here consists of articles from Springer, Elsevier and PubMed. However, we still believe that the results presented here are valid and can be generalized across publishers, to even bigger corpora. Preprocessing of corpora is an ongoing research project, and it can be difficult to weed out the rubbish when dealing with millions of documents. We have tried to use a process where we evaluate the quality of a subset of randomly selected articles repeatedly and manually, until it no longer improves.

10.1371/journal.ppat.1001106

The HA and NS Genes of Human H5N1 Influenza A Virus Contribute to High Virulence in Ferrets

Highly pathogenic H5N1 influenza A viruses have spread across Asia, Europe, and Africa. More than 500 cases of H5N1 virus infection in humans, with a high lethality rate, have been reported. To understand the molecular basis for the high virulence of H5N1 viruses in mammals, we tested the virulence in ferrets of several H5N1 viruses isolated from humans and found A/Vietnam/UT3062/04 (UT3062) to be the most virulent and A/Vietnam/UT3028/03 (UT3028) to be avirulent in this animal model. We then generated a series of reassortant viruses between the two viruses and assessed their virulence in ferrets. All of the viruses that possessed both the UT3062 hemagglutinin (HA) and nonstructural protein (NS) genes were highly virulent. By contrast, all those possessing the UT3028 HA or NS genes were attenuated in ferrets. These results demonstrate that the HA and NS genes are responsible for the difference in virulence in ferrets between the two viruses. Amino acid differences were identified at position 134 of HA, at positions 200 and 205 of NS1, and at positions 47 and 51 of NS2. We found that the residue at position 134 of HA alters the receptor-binding property of the virus, as measured by viral elution from erythrocytes. Further, both of the residues at positions 200 and 205 of NS1 contributed to enhanced type I interferon (IFN) antagonistic activity. These findings further our understanding of the determinants of pathogenicity of H5N1 viruses in mammals.

Highly pathogenic H5N1 influenza A viruses have caused more than 500 human infections with approximately 60% lethality in 15 countries and continue to pose a pandemic threat. The recent worldwide spread of pandemic H1N1 influenza A viruses raises the concern of reassortment between the H5N1 viruses and other influenza viruses. However, the molecular determinants for high virulence of the H5N1 viruses in mammals are not fully understood. We, therefore, investigated their virulence in a ferret model, which is a widely accepted animal model for assessing human influenza virus replication. We identified an amino acid in hemagglutinin and four amino acids in nonstructural proteins that are associated with high virulence of a human H5N1 virus, A/Vietnam/UT3062/04. We also found that the amino acid in hemagglutinin changes its receptor-binding property and the amino acids in nonstructural protein 1 affect its interferon antagonistic ability. These findings provide insight into the pathogenesis of H5N1 viruses in mammals.

In 1997, the first human case of influenza caused by an H5N1 virus occurred in Hong Kong [1], [2]. In 2003, a new outbreak of H5N1 virus was identified in Vietnam. Since then, H5N1 viruses have spread across Asia, Europe and Africa. As of July 22, 2010, 501 cases of H5N1 virus infections in humans have been reported by the World Health Organization (WHO; http://www.who.int/en/), 297 of which were fatal. The mortality is, therefore, approximately 60%. H5N1 viruses have been characterized by using a variety of mammalian models [3]. In mice, enhanced HA cleavability, as well as lysine at position 627 of the polymerase subunit PB2, plays an important role in the virulence of H5N1 viruses [4]. Viruses possessing these properties replicate systemically and cause death in mice. Ferrets are considered suitable for evaluating infection of human influenza viruses because these viruses replicate in the upper respiratory tract without adaptation in ferrets, and some strains cause severe pneumonia in these animals. Some of the H5N1 viruses isolated from humans can kill ferrets, whereas H5N1 viruses isolated from birds tend to cause mild disease in this animal model [5], [6]. Systemic infection, high replication efficiencies, and neurovirulence are associated with the high lethality of human H5N1 viruses in ferrets. Salomon et al. [7] reported that the genes encoding the nonstructural proteins (NS) and polymerase complex are important for the lethality of the human H5N1 virus A/Vietnam/1203/04 in ferrets, compared with the avian H5N1 virus A/quail/Vietnam/36/04. However, the molecular bases for the high virulence of H5N1 viruses in ferrets are not fully understood. To advance our understanding of the pathogenicity of H5N1 viruses, we compared the virulence of H5N1 influenza viruses isolated from humans in a ferret model. By generating reassortant viruses between the most virulent A/Vietnam/UT3062/04 (UT3062) virus and the avirulent A/Vietnam/UT3028/03 (UT3028) virus, we identified the genes responsible for high virulence in ferrets. We also performed in vitro studies to determine the molecular mechanisms by which H5N1 viruses exhibit high virulence in mammals. To compare the virulence of H5N1 influenza viruses isolated from humans in ferrets, we intranasally inoculated 5- to 7-month-old male animals (n = 3) with 107 plaque-forming units (PFU) of virus and observed the lethality, changes in body weight and body temperature, clinical signs, and virus shedding in the upper respiratory tract of the virus-infected animals (Table 1). The UT3062, A/Vietnam/UT3040/04, A/Vietnam/UT3028II/03, A/Vietnam/UT30850/05, A/Vietnam/UT3030/03, A/Vietnam/UT3040II/04, and A/Vietnam/UT3047III/04 were virulent in ferrets, causing the deaths of the virus-infected animals. These virulent viruses, with the exception of A/Vietnam/UT3028II/03, caused mean maximum weight loss of 9.1%–18.4% and anorexia, consistent with previous studies [5], [6]. Systemic viral infection was observed in most of the fatally infected animals (Table S1). Notably, inoculation of animals with UT3062 resulted in 100% lethality with 15.4±2.7% mean maximum weight loss (Table 1). These results demonstrate that UT3062 is the most virulent in ferrets of the viruses we tested. By contrast, six other human H5N1 viruses, A/Vietnam/UT30259/04, A/Vietnam/UT3035/03, A/Indonesia/UT3006/05, UT3028, A/Vietnam/UT30408III/05, and A/Vietnam/UT30262III/04 did not kill any ferrets and all, except A/Vietnam/UT30259/04, caused limited body weight loss (1.6%–6.6% mean maximum weight loss, median 4.2%) (Table 1), indicating that H5N1 viruses isolated from humans differ in their virulence in ferrets. Among the H5N1 viruses listed in Table 1, two viruses, A/Vietnam/UT3035/03 and A/Vietnam/UT30408III/05, which did not kill any ferrets, were isolated from patients who recovered from their H5N1 virus infections. The rest of the viruses used were isolated from patients who ultimately died. Of note, the virulence of test viruses in ferrets generally correlated with that in mice [8]. On days 3 and 6 post-infection (p.i.), we collected nasal washes from the virus-infected animals and titrated them in Madin-Darby canine kidney (MDCK) cells. On day 6 p.i., the virus titers in the nasal washes of animals infected with the virulent viruses (except for A/Vietnam/UT3028II/03) were generally higher than those of animals infected with viruses that were not lethal (Table 1). Sequence comparisons of the viruses used in this study revealed that the most virulent virus UT3062 was most closely related to an avirulent virus UT3028, with 18 amino acid differences in their 9 proteins (Table S2); there were no amino acid differences in the matrix (M) proteins of the two viruses. Therefore, we generated UT3062 and UT3028 by reverse genetics and confirmed their virulence by intranasally inoculating 5- to 6-month-old male ferrets with 107 PFU of the viruses. Animals infected with the UT3062 virus died on days 6–8, resulting in 67% lethality and showed −13.4±3.1% mean maximum weight loss (Figures 1 and 2). Conversely, all animals infected with the UT3028 virus survived and showed appreciably less mean maximum weight loss (−1.1±0.3%) (Figures 1 and 2). These results were similar to those obtained with the respective original viruses (Figure 2 and Table 1). To determine the molecular basis for the high virulence of UT3062, we generated reassortant viruses between the UT3062 and UT3028 viruses using reverse genetics and tested their virulence by intranasally inoculating 5- to 6-month-old male ferrets with 107 PFU of the viruses and observing them for 10 days for clinical manifestations. Since no amino acid differences were identified in M proteins of the two viruses, we used the UT3028 M gene for generating all reassortant viruses. The reassortants were named according to the origin of their UT3062 or UT3028 genes. For example, 3062(Pol+NP+HA+NA) indicates a virus possessing the polymerase complex (PB1, PB2, and PA), nucleoprotein (NP), HA, and neuraminidase (NA) genes from UT3062 and the rest of its genes from UT3028 (Figure 2). Reassortant viruses possessing both the UT3062 HA and NS genes, 3062(NP+HA+NA+NS), 3062(Pol+HA+NA+NS), 3062(HA+NA+NS), 3062(HA+NS), and 3062(NP+HA+NS), were lethal to animals (33%–100% lethality). By contrast, those possessing either the HA or NS genes of UT3028 were not lethal to any animals. Mean maximum body weight loss in the animals infected with the former viruses (8.5%–14.4%, median 11.4%) tended to be greater than that in the animals infected with the latter viruses (0.7%–8.6%, median 4.1%). Animals infected with 3062(HA+NS) resulted in 83% lethality and 17.8%±2.7% mean maximum weight loss. These results show that the difference in virulence between UT3062 and UT3028 is mainly attributable to both of the HA and NS genes. Sequence comparison between UT3062 and UT3028 revealed only one amino acid difference in their HA proteins and four amino acid differences in their NS proteins, two in NS1 and two in NS2 (Table 2). Since NS1 and NS2 mRNAs are produced from the same gene segment, with the NS1 mRNA being unspliced and the NS2 mRNA being spliced, the nucleotide alterations can affect both proteins; i.e., the amino acid differences at positions 200 and 205 of NS1 were coupled to those at positions 47 and 51 of NS2, respectively. Therefore, it was impossible to substitute the amino acids only in the NS1 or the NS2 protein. Here, we generated two reassortant viruses with a mutation in the NS segment by reverse genetics. 3062(HA)+NS1N200S possesses the UT3062 HA gene, a mutant NS segment encoding the UT3028 NS1 protein, which has an asparagine-to-serine substitution at position 200 (and encodes NS2 with a threonine-to-alanine substitution at position 47) and the rest of its genes from UT3028. 3062(HA)+NS1G205R possesses the UT3062 HA gene, a mutant NS segment encoding the UT3028 NS1 protein, which has glycine-to-arginine substitution at position 205 (and encodes NS2 with a methionine-to-isoleucine substitution at position 51) and the rest of its genes from UT3028. We then tested their virulence in ferrets as described above. As shown in Figure 2, both of the reassortant viruses, 3062(HA)+NS1N200S and 3062(HA)+NS1G205R, were not lethal to any animals, with 4.8%±1.4% and 2.3%±1.4 mean maximum weight loss, respectively. These results suggest that all of the amino acids in HA and NS proteins contribute to the virulence in ferrets, although it is unclear whether changes in NS1, NS2, or both affect virulence. In addition, 3062(HA+NA+NS) (33% lethality and 8.5±3.4% mean maximum body weight loss) was attenuated compared to 3062(NP+HA+NA+NS) (67% lethality and 14.4±4.9% mean maximum body weight loss). Further, 3062(NP+HA+NS) and 3062 (HA+NS) killed 100% and 83% of animals, respectively. These results suggest that the UT3062 NP gene may enhance virus virulence in ferrets. These results are consistent with previous findings that virulence of influenza virus is multigenic [7], [9], [10]. To understand the basis for the difference in virulence in ferrets among the viruses, we examined the in vitro and in vivo replication of the parental UT3062, UT3028, and the reassortant viruses. For in vitro testing, we compared their growth kinetics in mink lung epithelial (Mv1Lu) cells by infecting these cells with viruses at a multiplicity of infection (MOI) of 0.001 and monitoring the growth kinetics for 48 h. All of the viruses replicated to more than 108 PFU/ml at 36 or 48 h p.i. and the differences in their viral titers were less than one log PFU/ml at each time point (Figure S1), indicating that there were no substantial differences in their replicative ability in these cells. To examine viral replication in ferrets, we infected animals with 107 PFU of the parental UT3062, UT3028, and selected reassortant viruses (3062(HA+NS), 3062(NP+HA+NS), and 3062(HA+NA+NS)). Virus titers in nasal and tracheal swabs, and organs were examined. On days 3 and 7 p.i., three animals from each infected group were sacrificed for virus titration. As shown in Table 3, UT3062, 3062(HA+NS), 3062(NP+HA+NS), and 3062(HA+NA+NS) were detected systemically on days 3 and 7 p.i., whereas UT3028 was detected mainly in the upper respiratory tracts of ferrets on day 3, but not 7, p.i. The differences in replicative ability of these viruses in ferrets thus correlate with lethality in this animal model. When we compared the histopathology between ferrets infected with UT3062 and those infected with UT3028, we found three major differences (Figures 3 and 4): (1) host reaction to viral exposure in the lungs on day 1 p.i., (2) viral infection in the tracheobronchial lymph node, and (3) distribution of viral antigens and the inflammatory reaction in the lungs on day 3 and beyond p.i. Firstly, cells infiltrating the lung lesions differed between animals infected with UT3062 and those infected with UT3028. Although substantial numbers of viral antigen-positive cells were detected in the lungs of ferrets infected UT3062 or UT3028, the lungs of ferrets infected with UT3062 had marked infiltration of eosinophils around/in the bronchi (Figure 3A). By contrast, the lung lesions of ferrets infected with UT3028 contained many neutrophils (Figure 3B). Secondly, viral infection in the tracheobronchial (pulmonary regional) lymph node at 1 day p.i. differed between the two viruses. Although we did not detect viral antigen in the tracheobronchial lymph node of ferrets infected with UT3028, we did find viral antigen at this site in all three ferrets infected with UT3062 (Figure 3D and E). Thirdly, in animals infected with UT3028, we did not detect viral antigen beyond 3 days p.i., with the exception of one ferret, which was euthanized at 5 days p.i. (Figure 4A). The numerous neutrophils observed on 1 day p.i. were replaced by lymphocytes, macrophages and regenerative epithelial cells during the course of infection (data not shown). On the other hand, in animals infected with UT3062, a substantial number of viral antigen-positive cells were detected in the lungs even 3 days p.i. and the areas in the lungs where the viral antigen-positive cells were detected expanded widely by 5 and 7 days p.i. in some ferrets (Figure 4B). Moreover, when compared to ferret lung lesions with less viral antigen-positive cells, the lesions of ferrets with extensive viral antigen-positive cells had fewer lymphocytes and substantial pulmonary edema, hemorrhaging and fibrinous exudates (data not shown). These findings indicate that there was a tendency for delay in viral clearance in UT3062-infected ferrets and consequently some animals progressed to death. The virus was, however, completely eliminated in some animals, presumably because of individual animal variability. Next, to evaluate the effect of the UT3062 HA and NS genes in vivo, we examined the pathogenicity of 3062(HA+NS) virus, which possesses UT3062 HA and NS genes and its remaining genes from UT3028. When we examined ferrets infected with 3062(HA+NS) on days 3 and 7 p.i., we found that they had pathological lesions that more closely resembled those of ferrets infected with UT3062 than those of ferrets infected with UT3028. Namely, the ferrets had viral infection in the tracheobronchial lymph node and widely distributed viral pneumonia by 3 days p.i. (Figure 3G to I). Pulmonary edema, hemorrhages and fibrinous exudates were obvious in the lung lesions rather than recruitment of lymphocytes and regenerative changes, which were characteristic of ferrets infected with UT3028. Therefore, the UT3062 HA and NS gene products play a critical role in viral pathogenicity in this ferret model. Viruses first replicated in the lungs (at the primary site of viral exposure), and infection then expanded into the tracheobronchial lymph node. Viral infection in the regional lymph node may negatively affect viral exclusion from the host, leading to continued viral replication. UT3062, like almost all other human H5N1 viruses, has alanine at position 134 of HA (H3 numbering). UT3028, however, has threonine at this position (Table 2). These findings suggest that a single substitution at position 134 (A134T) of HA affects virulence in ferrets. Previously, Auewarakul et al. [11] showed that substitutions at positions 129 and 134 (L129V/A134V) allow virus recognition of both sialic acid liked to galactose by α2,3 linkage (SAα2,3Gal) and SAα2,6Gal, unlike the parent virus, which recognizes only SAα2,3Gal. Yamada et al. [12], however, found that a single substitution at position 134 (A134T) did not change receptor-binding preference with the same sialylglycopolymers used by Auewarakul et al. [11]. We, therefore, performed virus elution assays using chicken and horse erythrocytes. From chicken erythrocytes, which express both SAα2,3Gal and SAα2,6Gal [13], UT3062 and a reassortant possessing the UT3062 HA were not eluted even after 20h of incubation at 37°C. By contrast, UT3028 and a reassortant possessing the UT3028 HA were gradually released from chicken erythrocytes (Figure 5). Since the viruses possessing the UT3028 HA were eluted regardless of the origin of the NAs (either UT3028 or UT3062), this difference in elution from erythrocytes is due to the difference in the amino acid residue at position 134 of HA. When we used horse erythrocytes, all of the viruses were more rapidly eluted from these erythrocytes than from chicken erythrocytes; however, UT3028 and a reassortant possessing the UT3028 HA were eluted more efficiently from horse erythrocytes than were UT3062 and reassortants possessing the UT3062 HA (data not shown). These results suggest that UT3062 HA differs from UT3028 HA in its receptor-binding property. NS1 mediates type I IFN antagonism and affects viral growth in cells. We, therefore, assessed the IFN antagonistic activity of these NS1s by using an IFN bioassay [14], [15], [16], [17], [18], [19], [20]. Briefly, Mv1Lu cells were infected with each virus at an MOI of 1.25 and the supernatants were collected at 12–24 h p.i. H5N1 viruses in the supernatants were inactivated with UV and neutralizing antibody (A1A1, [21]) treatment. The supernatants were added to fresh Mv1Lu cells and cultured for 22 h, followed by infection with vesicular stomatitis virus (VSV) to determine VSV infectivity of the above-described supernatant-pretreated Mv1Lu cells. As a control, we used a recombinant influenza virus expressing an RNA-binding- and IFN-antagonism-defective NS1 protein within which two basic amino acids were substituted to alanines (R38A/K41A) on the UT3062 backbone [17], [22]. We also generated reassortant viruses possessing the mutant NS1 of UT3028 that has either the N200S or the G205R mutation, on the UT3028 backbone and designated them 3028NS1-N200S and 3028NS1-G205R, respectively. As described above, these viruses also possessed amino acid substitution in NS2, T47A, or M51I, respectively. At 18 h and 24 h p.i., the supernatant from Mv1Lu cells infected with UT3028 or a virus possessing the UT3028 NS gene and the remaining genes from UT3062 (i.e., 3028(NS)) inhibited VSV plaque formation more efficiently than did the supernatants from cells infected with UT3062 (statistically significant difference at P<0.05, Tukey Honestly Significant Difference [HSD] test) (Figure 6). Furthermore, the supernatant of cells infected with either 3028NS1-N200S or 3028NS1-G205R inhibited VSV plaque formation more efficiently than did that from viruses possessing the UT3062 NS gene (i.e., UT3062 and 3062(NS)) (statistically significant difference, P<0.05, Tukey HSD test), but less efficiently than that from viruses possessing the UT3028 NS gene (i.e., UT3028 and 3028(NS)) (Figure 6). These results indicate that both serine at position 200 and arginine at position 205 of NS1 contribute to the enhanced type I IFN antagonistic property of UT3062 NS1, which, in turn, leads to high virulence in ferrets. To further assess the IFN antagonistic property of NS1, we investigated the effects of the amino acids in NS1 on the expression of the firefly luciferase reporter gene under the control of an interferon-stimulated response element (ISRE) in 293 cells treated with IFNβ. Briefly, 293 cells were transfected with pISRE-Luc, pRL-TK, and pCAGGS NS1 or pCAGGS GFP (negative control). At 24 h post-transfection, the cells were treated with recombinant human IFNβ. At 30 h post-transfection, the cells were lysed and luciferase activities were measured by using the Dual-luciferase Reporter assay system. There were, however, no significant differences in expression from the ISRE between UT3062 NS1 and UT3028 NS1 (data not shown). We then investigated the effects of the amino acids in NS1 on the expression of the firefly luciferase reporter gene under the control of the IFNβ promoter in 293 cells treated with Sendai virus (SeV) as described previously [23]. Briefly, 293 cells were transfected with p125-Luc, pRL-TK, and pCAGGS NS1 or pCAGGS GFP (negative control). At 36 h post-transfection, the cells were treated with SeV (Cantell strain). At 48 h post-transfection, the cells were lysed and luciferase activities were measured by using the Dual-luciferase Reporter assay system. The results of this experiment also showed that there were no significant differences in expression from the IFNβ promoter between UT3062 NS1 and UT3028 NS1 (data not shown), indicating that other mechanisms affect the IFN antagonistic property. Here, using H5N1 viruses isolated from humans, we found that receptor-binding property and NS1 IFN antagonism play important roles in the high virulence of these viruses in ferrets. HA is a receptor-binding and fusion protein and, therefore, is required for virus entry. It is known to play a critical role in virulence [4], [24], [25], [26], [27]. In this study, we found that viruses possessing threonine at position 134 of HA were appreciably attenuated in ferrets compared to those possessing alanine. Although Yamada et al. [12] did not find differences in the receptor-binding preference between HAs with a single substitution at position 134 (134A or 134T) in a direct binding assay to sialylglycopolymers, we found that this substitution affected receptor-binding property as detected by a virus elution assay (Figure 5). Since the amino acid at position 134 is located near the receptor-binding pocket but does not directly interact with sialyloligosaccharides [11], the substitution at this residue may influence the receptor-binding property indirectly. Alanine at position 134 of HA is highly conserved in avian H5N1 viruses—only one virus is known to harbor serine at that position (the Influenza Sequence Database (ISD; https://flu.lanl.gov/, registration system [28])). Similarly, most human H5N1 viruses also have alanine at this position; however, UT3028 and two other H5N1 viruses that we isolated from humans have threonine at this position (Y. Sakai-Tagawa and Y. Kawaoka, unpublished). Further, eleven H5N1 viruses isolated from humans have valine and one has serine at this position (ISD; https://flu.lanl.gov/[28] and Y. Sakai-Tagawa and Y. Kawaoka, unpublished). These data indicate that an amino acid substitution at position 134 of HA is more frequently observed in human H5N1 viruses than in avian viruses, suggesting that viruses possessing a substitution at position 134 of HA may be selected during replication in humans. Although NS1 is a multifunctional protein, one of its main functions is to suppress type I IFN production [29]. Recent studies revealed that NS1 plays an important role(s) in antiviral responses via dsRNA-dependent protein kinase R (PKR) and 2′5′-oligoadenylate synthetase/RNase L [30], [31]. Here, using an IFN bioassay, we showed that both serine at position 200 and arginine at position 205 of NS1 contribute to the enhanced type I IFN antagonistic property of UT3062 that leads to high virulence in ferrets. However, we did not observe significant differences in IFNβ-stimulated expression from the ISRE or in SeV-stimulated expression from the IFNβ promoter in 293 cells between the UT3062 and UT3028 NS1 proteins. It may be that NS1 exhibits type I IFN antagonism by a mechanism other than tested in this study. Alternatively, the difference observed in the highly sensitive IFN bioassay using VSV is not detectable in other IFN assays. The amino acid residues at positions 200 and 205 of NS1 are not well conserved, although the residues at these positions in UT3062 have been observed in other human and avian H5N1 viruses (Table S3). These findings support the hypothesis that the amino acid residues determined to be important in this study are affected by the genetic background of the test viruses. Nonetheless, the HA amino acid at position 134 and the NS1 amino acids at positions 200 and 205 may now be included as virulence markers for H5N1 viruses. Our research protocol for the use of ferrets followed the University of Tokyo's Regulations for Animal Care and Use, which was approved by the Animal Experiment Committee of the Institute of Medical Science, the University of Tokyo (approval number: 19–29). The committee acknowledged and accepted both the legal and ethical responsibility for the animals, as specified in the Fundamental Guidelines for Proper Conduct of Animal Experiment and Related Activities in Academic Research Institutions under the jurisdiction of the Ministry of Education, Culture, Sports, Science and Technology, 2006. Madin-Darby canine kidney (MDCK) cells were maintained in minimal essential medium (MEM) with 5% newborn calf serum. Human embryonic kidney 293 and 293T cells were maintained in Dulbecco's modified Eagle's MEM (DMEM) with 10% fetal calf serum. Mink lung epithelial (Mv1Lu) cells were maintained in MEM with 10% fetal calf serum and 1% non-essential amino acids. All cells were grown at 37°C in 5% CO2. H5N1 viruses isolated from humans in Vietnam and Indonesia were used in this study (Table 1). Virus stocks were propagated through two passages in MDCK cells for 24–48 h at 37°C. The cell supernatants were harvested, clarified by centrifugation, aliquoted, and stored at −80°C. The frozen virus stocks were thawed and titrated for virus infectivity in MDCK cells by plaque assays. Virus titers were calculated as PFU/ml. All experiments were performed under biosafety level 3+ conditions. Viral RNA was extracted directly from the supernatants of H5N1 virus-infected MDCK cell cultures with a QIAamp Viral RNA Mini Kit (Qiagen, http://www1.qiagen.com/). Complementary DNA was generated by SuperscriptIII (Invitrogen, http://www.invitrogen.com/) with the universal primers for influenza A virus genes. The resulting products were PCR-amplified using PfuUltra High-Fidelity DNA polymerase (STRATAGENE, http://www.stratagene.com/) with specific primers for each virus gene and cloned into a plasmid under the control of the human RNA polymerase I (PolI) promoter and the mouse RNA PolI terminator (PolI plasmids). We altered the NS1 gene sequence that encodes the RNA-binding site of UT3062 to create the RNA-binding defective sequence (R38A/K41A) as previously described [17]. All reassortant viruses and the parental UT3062 and UT3028 viruses were generated by plasmid-based reverse genetics, as described by Neumann et al. [32]. Briefly, PolI plasmids and protein expression plasmids were mixed with a transfection reagent, TransIT 293T (Mirus Bio, http://www.mirusbio.com/); incubated at room temperature for 15 min; and then added to 293T cells. Transfected cells were incubated in OPTI-MEM I (Invitrogen) for 48 h. Supernatants containing infectious viruses were harvested and propagated in MDCK cells at 37°C for 48 h. The supernatants were harvested, aliquoted, and stored at −80°C. We used male ferrets, 5–7 months old (MarshallBioResources, http://www.marshallbioresources.com/) in this study. All ferrets were inoculated intranasally with 107 PFU of infectious virus in 500 µl of phosphate-buffered saline (PBS) under anesthesia with ketamine (25 mg/kg) and xylazine (2 mg/kg). Clinical signs, body weights, and body temperatures were recorded daily for 10 days post-infections (p.i.). The percent changes in body weights were calculated by comparing the weights of each ferret at each time point to its initial weight on day 0. Body temperatures were measured using a rectal thermometer. Changes in body temperature were calculated by comparing the body temperatures of each ferret at each time point to its initial body temperature on day 0. All animals exhibiting more than 20% weight loss, hemorrhage from any body orifice, or inability to remain upright were euthanized. Surviving ferrets were euthanized under deep anesthesia at 3 weeks p.i. On days 3 and 6 p.i., nasal washes were collected from anesthetized ferrets and titrated for virus infectivity in MDCK cells by plaque assays. Ferrets infected with the parental UT3062 and UT3028 and selected reassortant viruses (Table 3) were euthanized with deep anesthesia and necropsied on days 3 and 7 p.i. Tissue samples of the brain, olfactory bulb, lungs, hilar lymph node, liver, kidney, spleen, duodenum, and descending colon were collected. A portion of each was stored at −80°C for virus titration and the rest were preserved in 10% neutral buffered-formalin for pathological examination. To prepare 10% tissue emulsions, frozen tissue samples were thawed, weighed, and homogenized in 10 volumes (w/v) of sterile PBS using a multi-beads shocker (Yasui Kikai, http://www.yasuikikai.co.jp/). After centrifugation of the samples at 800× g for 5 min at 4°C, the supernatants were collected. In addition, swabs from nose and trachea were collected and suspended in 1 ml of sterile PBS containing 0.3% BSA and penicillin (200 U/ml). After centrifugation at 800× g for 5 min at 4°C, the supernatants were collected and stored at −80°C. Virus in nasal washes, swabs and tissue samples was titrated for virus infectivity in MDCK cells by plaque assays. Virus titer was expressed as PFU/g for tissue samples and PFU/ml for nasal washes and swabs. The limitations of virus detection were 102.0 PFU/g for tissue samples and 101.0 PFU/ml for nasal washes and swabs. The formalin-fixed tissues were processed for routine paraffin embedding. The paraffin-embedded tissues were cut into 5 µm thick slices and stained using hematoxylin-and-eosin (H&E). Additional sections were cut for immunohistological staining with rabbit polyclonal antibodies against an H5N1 virus (A/Vietnam/1203/04). Specific antigen-antibody reactions were visualized by means of 3,3′ diaminobenzidine tetrahydrochloride and the Dako EnVision system (Dako, http://www.dako.jp/). All animal experiments were approved by the Animal Research Committee of The University of Tokyo. The reassortant viruses and the parental UT3062 and UT3028 viruses were inoculated into Mv1Lu cell monolayers at an MOI of 0.001 PFU with MEM containing 0.3% bovine serum albumin, and incubated at 37°C. Cell supernatants were harvested at a given number of hours p.i. After centrifugation at 1,000× g for 5 min, samples were titrated for virus infectivity in Mv1Lu cells by plaque assays. The ability of viruses to be eluted from erythrocytes was assessed as previously described [33], [34], with some modifications. Briefly, virus stocks were diluted serially in calcium saline (0.9 mM CaCl2-154 mM NaCl in 20 mM borate buffer, pH 7.2) and 50 µl aliquots were incubated with 50 µl of 0.55% chicken erythrocytes at 4°C for 1 h in microtiter plates. The plates were then transferred to 37°C and monitored periodically for 20 h. Levels of IFN secreted by virus-infected Mv1Lu cells were assessed as previously described [14], [15], [16], [17], [18], [19], [20], with some modifications. Briefly, Mv1Lu cells were infected with each virus at an MOI of 1.25. Supernatants from infected cells were harvested 12–24 h p.i. To inactivate viral infectivity, the supernatants were treated with UV light for 20 min and then mixed with neutralizing α-VN1203HA monoclonal antibodies (A1A1, [21]). The supernatants were added to fresh Mv1Lu cells and incubated for 22 h. These pretreated Mv1Lu cells were then infected with VSV, and the VSV infectivity titers were determined. As a control, we used a recombinant influenza virus expressing an RNA-binding- and IFN antagonism-defective NS1 protein within which two basic amino acids were substituted to alanines (R38A/K41A). The experiments were carried out in triplicate and were independently repeated twice. 8×104 of 293 cells were transfected with 50 ng of pISRE-Luc (Clontech, http://www.clontech.com/) and 50 ng of pRL-TK (Promega, http://www.promega.com/) by using TransIT293 (Mirus, http://www.mirusbio.com/). 100 ng of pCAGGS NS1 or pCAGGS GFP were co-transfected. Cells were incubated for 24 h and were treated with 100 units of recombinant human IFNβ, 1a (PBL interferonSource, http://www.interferonsource.com/). At 30 h post-transfection, cells were lysed and luciferase activities were measured by using the Dual-luciferase Reporter assay system (Promega, http://www.promega.com/) according to the protocol provided by the manufacturer. Firefly luciferase values were divided by Renilla luciferase values to normalize for transfection efficiency. 2×105 of 293 cells were transfected with 250 ng of p125-Luc (the reporter plasmid carrying the firefly luciferase gene under the control of the IFNβ promoter was kindly provided by Takashi Fujita) and 250 ng of pRL-TK (Promega, http://www.promega.com/) by using TransIT293 (Mirus, http://www.mirusbio.com/). 5–500 ng of pCAGGS NS1 or pCAGGS GFP were co-transfected. After incubation for 24 h, cells were stripped with trypsin and divided into two wells of a 12-well plate. At 36 h post-transfection, cells were treated with approximately 5×1010 focus-forming unit of SeV (Cantell strain). At 48 h post-transfection, cells were lysed and luciferase activities were measured by using the Dual-luciferase Reporter assay system (Promega, http://www.promega.com/) according to the protocol provided by the manufacturer. Firefly luciferase values were divided by Renilla luciferase values to normalize for transfection efficiency.

10.1371/journal.ppat.1000312

Molecular Evolution of a Viral Non-Coding Sequence under the Selective Pressure of amiRNA-Mediated Silencing

Plant microRNAs (miRNA) guide cleavage of target mRNAs by DICER-like proteins, thereby reducing mRNA abundance. Native precursor miRNAs can be redesigned to target RNAs of interest, and one application of such artificial microRNA (amiRNA) technology is to generate plants resistant to pathogenic viruses. Transgenic Arabidopsis plants expressing amiRNAs designed to target the genome of two unrelated viruses were resistant, in a highly specific manner, to the appropriate virus. Here, we pursued two different goals. First, we confirmed that the 21-nt target site of viral RNAs is both necessary and sufficient for resistance. Second, we studied the evolutionary stability of amiRNA-mediated resistance against a genetically plastic RNA virus, TuMV. To dissociate selective pressures acting upon protein function from those acting at the RNA level, we constructed a chimeric TuMV harboring a 21-nt, amiRNA target site in a non-essential region. In the first set of experiments designed to assess the likelihood of resistance breakdown, we explored the effect of single nucleotide mutation within the target 21-nt on the ability of mutant viruses to successfully infect amiRNA-expressing plants. We found non-equivalency of the target nucleotides, which can be divided into three categories depending on their impact in virus pathogenicity. In the second set of experiments, we investigated the evolution of the virus mutants in amiRNA-expressing plants. The most common outcome was the deletion of the target. However, when the 21-nt target was retained, viruses accumulated additional substitutions on it, further reducing the binding/cleavage ability of the amiRNA. The pattern of substitutions within the viral target was largely dominated by G to A and C to U transitions.

RNA viruses are well-known for their tremendous capacity to evolve, a characteristic that threatens the development of effective antiviral strategies. A new antiviral strategy was recently proposed to control plant RNA viruses that relied on the expression in plants of artificial microRNAs (amiRNAs) targeting short sequences of 21-nt in the viral genome. Here, we have evaluated the likelihood that changes in the 21-nt target sequence would result in resistance breakdown. We found that changes at different sites in the target had different consequences on the ability of the virus to evade amiRNA surveillance. Then, we evolved viruses with a single substitution within the target under the selective pressure imposed by amiRNAs but without any selective pressure at the protein level. We found extra mutations accumulated in the target that further reduced base pairing with the amiRNA. These results showed that when allowed to replicate, RNA viruses would readily generate genetic variability that would facilitate evasion from the engineered miRNA-mediated virus resistance.

Plant miRNAs regulate the abundance of target mRNAs by guiding their cleavage at the sequence complementary region. Previous reports have shown that changes of several nucleotides within a miRNA 21-nt sequence do not affect its biogenesis and maturation [1],[2]. This finding raises the possibility to redesign the miRNA sequence to target specific transcripts, originally not under miRNA control. Such artificial miRNAs have been produced in dicotyledonous [3]–[5] and monocotyledonous plants [6] using different pre-miRNAs as backbones. We have successfully demonstrated that redesigned artificial miRNAs (amiRNAs) are biologically active and can be used to confer specific virus resistance in transgenic plants [4]. The pre-miR159a precursor was used to generate two amiRNA159s (amiR159-P69 and amiR159-HC-Pro) with sequence complementary to the RNA genome of two plant viruses, Turnip yellow mosaic virus (TYMV) and Turnip mosaic virus (TuMV), respectively. The amiR-P69 was designed to target sequences encoding the P69 suppressor of TYMV whilst amiR159-HC-Pro would target sequences for the HC-Pro silencing suppressor of TuMV. Transgenic lines carrying both 35S-pre-amiR159-P69 and 35S-pre-amiR159-HC-Pro transgenes can express the appropriate amiRNA at high levels and showed specific resistance to either TYMV and TuMV, depending on the expression of the cognate amiRNA [4]. Specific resistance to TuMV was also seen with plants expressing amiR159-TuCP directed against the TuMV coat protein (CP) gene [4]. In animal systems, RNA interference (RNAi), a gene-silencing mechanism similar to that of miRNA, has been used in clinical trials as antiviral therapeutics to inhibit replication of several human pathogenic viruses (reviewed in [7],[8]). As demonstrated for HIV-1, a major problem of RNAi-mediated antiviral therapies is the emergence of resistant virus variants, which differ from the wild type virus by having fixed point mutations in the target sequence leading to imperfect matching; these mutant viruses are not properly processed by the enzymatic silencing machinery [9]–[13]. Some mismatches within the target sequence are tolerated by the RNAi machinery whereas other mismatches, such as those in the central region (position 9 to 11) of the target sequence, compromise RNAi-guided antiviral therapies [14],[15]. However, all these studies suffer from the drawback of having a superimposition of two different selective forces: on the one hand, purifying selection acts at the protein level (i.e., the necessity of maintaining a functional protein) and, on the other hand, diversifying selection acts at the RNA sequence level favoring mutant genomes capable of evading RNA silencing. Mallory et al (2004) have used an in vitro wheat germ system to assay for critical positions within a miRNA target site needed for efficient plant mRNA cleavage [16]. Analysis of scanning mutants revealed that mismatches at the center and the 3′ end of the miRNA are more tolerated compared to mismatches at the 5′ region [16]. Recently, the molecular mechanism of RISC-mediated RNA cleavage has been investigated by in vitro reconstitution assays using human RISC [17]–[19]. It was found that the accessibility of RNA target site correlates directly with the RNA cleavage efficiency, indicating that RISC is unable to unfold structured RNA. In the course of target recognition, RISC transiently contacts single-stranded RNA nonspecifically and promotes siRNA-target RNA annealing. Furthermore, the 5′ portion of the siRNA within RISC creates a thermodynamic threshold that determines the stable association of RISC and the target RNA. Furthermore, in addition to this clear position-effect, overall desestabilization of the double strand structure has little effect on RNAi activity until an energy threshold is reached, beyond of which a negative correlation exist between stability and RNAi-mediated inhibition [15]. Here, we first investigated whether the 21-nt of an amiRNA target site is both necessary and sufficient for amiRNA-mediated specific resistance. Second, we were interested in identifying critical positions within the target site for this resistance. Third, we have explored the patterns of sequence polymorphism of viral sequences that evolve under the only selective pressure of amiRNA-mediated silencing. To address these issues, we established a heterologous-virus resistance system using a TuMV-GFP viral vector to carry a non-essential 21-nt sequence of the P69 gene targeted by amiR159-P69. This heterologous-virus system allows us to modify any nucleotide within the 21-nt target site without altering virus coding sequences and thus without affecting replication and activity. In other words, this heterologous system allows separating the selective pressure imposed by protein functionality from the selective pressure imposed at the sequence level by RNA silencing. The 21 scanning mutant viruses were inoculated on amiR159-P69 plants and the proportion of transgenic plants that became infected was used to determine the importance of the mutated nucleotide position within the amiRNA target site. We have previously demonstrated that a 21-nt amiRNA, with sequence complementary to a viral sequence, can mediate cleavage of target viral RNA and confer resistance on transgenic plants [4]. However, it was not known whether the 21-nt viral target site, complementary to the amiRNA sequence, was sufficient for specific resistance. To this end, we constructed a green fluorescence protein (GFP) gene carrying a 21-nt sequence (5′-CCACAAGACAAUCGAGACUUU-3′) of the TYMV P69 gene at its 3′-end and inserted the GFP-P6921nt fusion gene in between the NIb and CP genes to generate a TuMV-GP69 chimeric virus (Fig. 1). As a control, we mutated 4 nts (position 9 to 12 from the 3′-end; underline) of the target 21-nt sequence (5′-CCACAAGACCUGAGAGACUUU-3′) to give GFP-P6921ntm, which was inserted in the same position of the viral genome to generate the TuMV-GP69m chimeric virus (Figure 1D). The presence of two TuMV NIa protease cleavage sites (CVYHQA) at both the N- and C-termini of the GFP-P6921nt fusion protein allows the release of GFP plus a 7 amino acid C-terminal extension (PQDNRDF) from the TuMV-GFP viral polyprotein (Figure 1C and 1D). Virus infection can be easily confirmed and followed by monitoring GFP signals from infected tissues. We have previously shown that transgenic Arabidopsis thaliana plants expressing amiR-P69 can specifically target the P69 gene of TYMV and displayed specific resistance to TYMV [4] although these plants remained susceptible to heterologous virus (TuMV-GFP) infection (Figure 2A, top second panel). Figure 2 shows that insertion of the 21-nt sequence of the TYMV P69 gene into TuMV-GFP, which was targeted by amiR159-P69, rendered these amiR159-P69 plants resistant to TuMV-GP69 (Figure 2A, top third panel). Control experiments showed that the amiR159-P69 plants remained sensitive to TuMV-GP69m (Figure 2A, top fourth panel), which carried 4 mutations in the central region of the 21-nt site of the P69 gene. Systemic leaves of amiR159-P69 plants displayed GFP fluorescence when inoculated with TuMV-GFP or TuMV-GP69m (Figure 2B, top second and fourth panels), but no GFP signal was detected upon TuMV-GP69 inoculations (Figure 2B, top third panel). Plants expressing amiR159-HC-Pro were resistant to chimeric TuMV-GFP, TuMV-GP69 and TuMV-GP69m and no GFP was detected on systemic leaves of inoculated plants (Figure 2B, bottom panels). These results were expected since all these 3 chimeric viruses contained the HC-Pro gene targeted by amiR159-HC-Pro [4]. Next, we established the heterologous virus resistance system in N. benthamiana and tested amiRNA-mediated resistance efficiency. Figure 3A shows amiR159-P69 expression levels in 4 independent transgenic N. benthamiana lines (#1, 2, 3, and 4). Progeny plants of these lines were challenged with TuMV-GFP, TuMV-GP69 or TuMV-GP69m. The GFP signal produced by infection with TuMV-GFP or TuMV-GP69m can be detected at 4 dpi (early stage of symptom development) (Figure 3B). As expected, at 7 dpi, transgenic N. benthamiana plants expressing amiR159-P69 were resistant to TuMV-GP69 but susceptible to TuMV-GFP and TuMV-GP69m (Figure 3C). Plants that were sensitive to virus infection showed severe wilting symptoms (Figure 3C). These results, which were very similar to those obtained with A. thaliana transgenic plants, provided further confirmation that the targeted 21-nt site is necessary and sufficient for specific amiRNA-mediated specific resistance. In addition, the results also suggested that pre-amiR159-P69, which is a modified form of the Arabidopsis miR159 precursor, can be processed by N. benthamiana plants to produce functional amiR159-P69 to confer virus resistance. As the 4 nt mutation on the central positions of the target sequence (position 8 to 12) compromised specific resistance, we decided to further investigate nucleotide positions within this 21-nt sequence that are critical for amiRNA-mediated resistance. Note that the sequence can be systematically altered without affecting essential viral gene functions because the amiR159-P69 target sequence is non-essential to the TuMV-GP69 chimeric virus. Accordingly, we generated a series of mutants by making all possible synonymous scanning substitutions within the 21-nt sequence of P69 in the background of TuMV-GP69 (Figure 4A). Each A of the P69 viral sequence that pairs to a U of amiR159-P69 was changed to a C to create a C∶U mismatch; each C and G of the viral sequence was changed to an A to create A∶G and A∶C mismatches; and each U of the viral sequence was changed to a C to create C∶A mismatches. A total of 21 mutant viruses with single nt substitution from the 1st to the 21st position of the target site were used to challenge non-transgenic WT and amiR159-P69 N. benthamiana plants. The proportion of inoculated amiR159-P69 plants that showed visible symptoms after inoculation, i.e. pathogenicity, was used as a measure of the importance of the mutated nucleotide within the 21-nt target site in amiR159-P69–mediated specific resistance. We used TuMV-GFP, TuMV-GP69, and TuMV-GP69m as controls. Whereas WT plants were susceptible to TuMV-GP69 no symptoms developed in amiR159-P69 plants even at 10 dpi (Table 1, and Figure 4B bottom third panel). By contrast, TuMV-GFP and TuMV-GP69m elicited 100% infection on WT as well as amiR159-P69 transgenic tobacco plants, and these infected plants displayed symptoms at 5 dpi (Table 1, and Figure 4B bottom second and fourth panel). WT plants were 100% susceptible to all 21 scanning mutant viruses and symptoms appeared at 5 dpi (data not shown), indicating that the single nt substitutions on the target 21-nt sequence did not affect mutant virus replication nor movement. On the other hand, these mutant viruses showed variable pathogenicity after inoculation on amiR159-P69 plants (Table 1, and Fig 4B top panels & bottom first panel). Fifteen mutants showed pathogenicity values that were significantly greater than zero (Table 1). For these pathogenic mutants, the percentage of infected plants ranged from 8.33% (m7) to 92.86% (m9). Mutants were classified according to their pathogenicity using a 2-step cluster analysis. The minimum number of clusters into which the mutants can be significantly partitioned was three (Bayesian weight 98.44%; Kruskal-Wallis test: H = 17.739, 2 d.f., P<0.001). Figure 4C assigns mutants to the different clusters. The first cluster (green bars in Fig. 4C), is characterized by positions causing low pathogenicity, with an average value of 6.91±0.59%, suggesting that these sites are non critical for resistance. These low pathogenicity mutants are scatter along the entire 21-nt region. The second cluster contains mutants of intermediate pathogenicity (yellow bars in Figure 4C), with an average value of 36.36±4.71%, suggesting that these positions are moderately important for amiRNA-mediated resistance. Most of these medium effect mutants are located between nucleotides 10 and 18, with the exception of m2, which is located at the 3′ end of the target sequence. Finally, the third cluster contains those mutants with a greater likelihood of resistance breakdown (red bars in Figure 4C). On average, these large effect mutants have 81.85±4.14% pathogenicity, highlighting their importance in amiRNA-mediated resistance. These important sites mostly congregate on the 3′ third of the target sequence, plus m9 and m12 which are located in the center of the sequence. In good agreement with the pathogenicity data, symptoms elicited by these mutants were generally delayed in comparison with TuMV-GFP (Table 1). For the 9 small-effect mutants, the median delay in symptom development was two days, for the medium effect mutants, two days, and for the large-effect mutants, only one day. We sought to gain deeper insights into the question of why different substitutions within the 21-nt target sequence overcame the amiRNA-mediated resistance to different degrees. To address this issue, we recovered viral RNAs from symptomatic leaves of amiR159-P69 plants and analyzed the 21-nt target sequence on the viruses by RT-PCR and sequencing. Several possibilities could account for the resistance breakdown. (1) The scanning mutation could affect amiR159-P69-mediated cleavage to different degrees in amiR159-P69 plants. (2) The scanning mutant virus could undergo rapid evolution accumulating additional mutation(s) within the target site to further increase the number of mismatches and consequently the ability to replicate in presence of the amiR159-P69. (3) Since the 21-nt sequence and the GFP gene are non-essential for virus survival, the surviving mutant virus could undergo in-frame deletions in this region that would render the virus unrecognizable by the amiR159-P69. To discriminate amongst these possibilities, we designed two primers (PTuNIb-8671 and MTuCP-8982) to amplify an 1136 bp DNA fragment including a partial NIb gene, GFP gene, the target 21-nt sequence, and a partial CP gene (Fig. 1B). This primer set can be used to check for any possible deletion within the GFP-21nt sequence. The recovery of lower molecular mass PCR fragment(s) would indicate deletion of this region. In addition, we designed another primer set (PXFP-532 and MTuCP-8982) to amplify a 482-bp fragment that included a partial GFP gene and a partial CP gene (Fig. 1B). This 482-bp DNA fragment can be used to analyze sequences surrounding and within the 21-nt target site. Figure 5A top panel shows that several virus sequences recovered from symptomatic amiR159-P69 plants contained deletion of the NIb-CP gene, such as TuMV-GP69m2 (lanes 3, 4 and 7) and TuMV-GP69m3 (lanes 10 and 13). In addition, several viruses, such as TuMV-GP69m5-13, -15, and -19 contained partially-deleted 21-nt sequence (Fig. 5B). Moreover, TuMV-GP69m5-15 also accumulated two additional mutations on positions 6 and 8 (Fig. 5B). We selected viruses with no deletion on the NIb-CP gene region and sequenced the GFP-CP gene regions (Figure 5A, bottom panel). Our results showed that virus sequences recovered from symptomatic amiRNA plants contained additional mutations within the 21-nt target site (Table 2). These scanning mutant viruses have 1–3 additional mutation(s) on the 21-nt target site and most of additional mutations introduced additional mismatches (Table 2). Only 11 out of the 21 positions showed additional mutations. These 11 positions are 3, 4, 6, 8, and 10–16 (Figure 5C). Positions 8, 14 and 16 were of little importance for amiRNA mediated resistance (see above), whereas all other 8 positions had either a moderate or a large effect on the likelihood of escaping the amiRNA-mediated resistance. All together, our results indicate that up to 2 mutations on critical positions within the 21-nt sequence can overcome specific resistance. Interestingly, we found that 40 out of 55 observed additional mutations were transitions. Over 50% of the additional mutations in positions 3, 4, 6, 8, 11, and 15 were transition mutations. For example, there were 100% U→C or C→U transitions in position 3 and 4. In position 15, the G→A transition represented 88.89% of all observed mutations at this particular site. Moreover, there were 50 to 66.67% of G→A transitions at positions 6, 8, and 11. This result is not surprising, since it is well known that virus coding regions show an excess of transitions over transversions [20],[21]. Three reasons can account for this bias: (i) the underlying mechanisms of mutation render transitions easier than transversions, (ii) the redundancy of the genetic code is expected to make the average effect of a transition smaller than the average effect of a transversion, and (iii) RNA editing by deaminase-like enzymes have been shown to induce transition mutations in single-stranded regions of certain viral genomes [22],[23]. Our results show that transitions rather than transversions also mainly accumulate in viral sequences, such as that of the target of amiR159-P69, which are not under the selective constrain imposed by being a coding sequence. Furthermore, not all transitions are equally represented in Table 2, since G→A (17/40) and C→U (14/40) are significantly over represented (χ2 = 12.600, 3 d.f., P = 0.006). This bias amongst transitions is expected if the viral RNA was edited by cytidine deaminase enzymes. Here, we have developed a heterologous-virus resistance system to study and identify critical positions of amiRNA target site for amiRNA-mediated resistance. The amiR159-P69 transgenic plant were resistant to TYMV, but not to TuMV (a heterologous-virus), because there was no sequence homology with amiR159-P69 on the TuMV viral genome [4]. However, the chimeric heterologous-virus TuMV-GP69 carrying the 21-nt sequence of P69 gene cannot infect amiR159-P69 plants because of amiR159-P69-mediated cleavage. By contrast, the TuMV-GP69m virus with mutations on the central region within this sequence is sufficient to prevent amiRNA-mediated cleavage on the viral RNA and compromise specific virus resistance. These results indicated that the 21-nt target site is portable and is necessary and sufficient to confer virus resistance. Because of the genome organization and proteolytic processing strategy of potyvirus, TuMV can express GFP when a cDNA for this protein is inserted in-frame between the NIb and CP genes. The encoded GFP protein contains two NIa proteinase cleavage sites (CVYHQ/A) at the N- and the C-terminus such that GFP can be released from the viral polyprotein by proteolytic processing. In addition, the additional 21-nt target site that encodes seven amino acids is also nonessential for TuMV. Therefore, any modification on the GFP gene and the 21-nt target site would not affect the chimeric TuMV as evidenced by its ability to infect plants and stably replicate. Using an in vivo assay we identified critical positions on the 21-nt target sequence for RISC-amiRNA-mediated cleavage. Scanning mutations on the 21-nt target site of the challenging chimeric virus showed different degree of resistance breakdown on amiR159-P69 transgenic plants. Based on the proportion of amiR159-P69 plants that become susceptible, we defined critical, moderately critical and non-critical positions on the 21-nt sequence. Positions 3–6, 9, and 12, are found to be critical for resistance because chimeric virus with mutations at these sites were pathogenic, on average, on ∼82% of amiR159-P69 plants. Positions 2, 10, 11, 13, 15, and 18 are classified to be moderately critical mutations giving average pathogenicity of ∼36% in transgenic plants. The remaining positions are classified as non-critical for resistance since mutants at these sites were only pathogenic in less than 7% of inoculated plants. In summary, most critical positions are localized on sequences complementary to the 5′ portion of the amiRNA whereas the moderate critical positions are mainly localized on the central region of the target site. The exception being position 18, which is complementary to the 18th nucleotide on 3′ portion of the amiRNA, and was also moderately important for amiRNA-mediated resistance. These results are consistent with those obtained with in vitro miRNA-mediated cleavage using a wheat germ system [16]. All together, results suggest that the 5′ portion of the miRNA is more important in governing the specificity of miR165/166 regulation [16]. Furthermore, the “two-state model” for RISC-mediated target recognition also proposes that the specific interaction between RISC and the substrate is initiated via the 5′ portion of siRNA, as the 3′ portion is less favorably structured to undergo base pairing before the initial recognition of a mRNA target [17]. Representative results showed that several virus sequences recovered from symptomatic amiRNA plants contained deletions and additional mutations within the 21-nt target site. This observation is consistent with the hypothesis that mutations in certain critical positions within the target site reduced amiRNA-mediated cleavage efficiency (Figure 6A). The reduced RISC activity allowed certain mutant viruses to escape amiRNA-mediated cleavage, allowing them to replicate and complete an infectious cycle (Figure 6B). During subsequent virus replication, additional mutations or deletions of the target sequence would be positively selected because they would escape from the amiRNA-mediated specific resistance more efficiently (Figure 6B). Indeed, the effect of miRNA-mediated cleavage was drastically alleviated in transgenic plants expressing the silencing suppressor P1/HC-Pro. Chimeric Plum pox virus (PPV) carrying an endogenous miRNA target site can also overcome the resistance by deletion and mutation on the 21-nt target sequence [24]. Finally, deletions, in addition to point mutations, are also a very common pathway taken by HIV-1 to escape from RNAi-based therapy in cell culture experiments [10],[13]. In general, these deletions have a major impact on the local RNA secondary structure, creating new hairpin structures not accessible to the siRNAs [13]. In some deletion mutant viruses recovered from breakdown plants, the entire GFP gene, along with small portions of the NIb gene C-terminus or the CP gene N-terminus, has been deleted from the viral genome (data not shown). These results suggest that TuMV can tolerate small deletions in the NIb or CP gene. In addition, deletions in between the GFP gene and the target site (Fig. 5B) may be triggered by polymerase-jumps on repeat sequence (ACAA). Widespread plant miRNA-directed translational repression as an important miRNA-mediated regulatory mechanism in plants has recently been reported [25]. Imperfect pairing with central mismatches in small RNA-target hybrids promotes translational repression because it excludes slicing [25]. This observation suggests the possibility that imperfect pairing between the amiRNA and mutant target sequences might lead to translational repression rather than viral RNA cleavage. In contrast to the catalytic effects of amiRNA-mediated viral RNA cleavage, translational repression requires stoichiometric amounts of amiRNAs and therefore is not as efficient. Inefficient translation inhibition might allow residual virus replication and progeny virus can still escape the repression by fixing changes in the target sequence. In this study, we have provided evidence that the 21-nt target site is necessary and sufficient for amiRNA specific resistance and we have also identified several positions on the target site that are critical for this resistance. These results are clearly important for future design of amiRNA-mediated virus resistance. Highly conserved regions on viral genomes should be selected as target sites to minimize the likelihood of fixation of mutations responsible for resistance breakdown, because these mutations might affect viral protein function and hence have a negative impact on virus fitness and survival. Furthermore, several amiRNAs targeting different conserved regions on a viral genome could be co-expressed in transgenic plants to minimize the chances of resistance breakdown. Finally, the heterologous-viral system described here also can be used for viral evolution studies in the future. As we have highlighted several times here, the amiR159-P69 target sequence inserted in the TuMV-GFP genome is functionally neutral. This has allowed us to separate selective pressures acting on the protein level from those acting on the RNA level. Consequently, the patterns of molecular evolution should be different. Not surprisingly, and in agreement with previous data obtained with other viruses, we have observed that most of the mutations fixed within the target were transition mutations [21]. We consider it striking that 77.50% of these fixed transitions were of the type G→A and C→U. These transitions are from the particular type induced by cellular cytidine deaminases (CDAs) involved in innate immune responses to viral infection [26], a phenomenon particularly well described for HIV-1 and other retroviruses [23],[27] but never before on an RNA virus. This observation prompted us to hypothesize that as an antiviral strategy plants may have an RNA-editing system that induces hypermutagenesis in viral genomes. A thaliana contains a family of nine paralogous genes that are annotated as CDAs owing to their homology to CDA1 [28]. These nine genes are good candidates to explore whether their gene products possess cytidine deaminase activity and whether they are indeed involved in plant antiviral defense. Two amiRNA transgenic A. thaliana lines, amiR159-P69 and amiR159-HC-Pro, were used in this study [4]. Plants of N. benthamiana were transformed with Agrobacterium tumefaciens containing the pBA-amiR159-P69 plasmid by standard methods. T2 transgenic N. benthamiana (a mixture of homozygotes and hemizygotes) were analyzed for transgene and miRNA levels and 4 independent lines (#1, 2, 3, and 4) with high amiR159-P69 expression levels were selected for virus challenge experiments. Seeds were surface-sterilized and chilled at 4°C for 2 d before being placed on Murashige and Skoog (MS) medium with/without antibiotics or sowed directly on Florobella potting compost/sand mix (3∶1). Plants were maintained in a growth room (16 h light/8 h darkness, 20 to 25°C). Ten µg of total RNA was resolved in a 15% polyacryamide/1× TBE (8.9 mM Tris, 8.9 mM boric acid, 20 mM EDTA)/8M urea gel and blotted to a Hybond-N+ membrane (Amersham). DNA oligonucleotides with the exact reverse-complementary sequence to miRNAs were end-labeled with 32P-γ-ATP and T4 polynucleotide kinase (New England Biolabs) to generate high specific activity probes. Hybridization was carried out using the ULTRAHyb-Oligo solution according to the manufacturer's directions (Ambion) and signals were detected by autoradiography. In each case, the probe contained the exact antisense sequence of the expected miRNA to be detected. The TuMV infectious clone (p35STuMV-GFP) comprises of a 35S promoter and the full-length cDNA of TuMV-GFP. The GFP gene was inserted between NIb and CP genes. This chimeric TuMV-GFP virus was used as a surrogate wild type virus and as a backbone for construction of various chimeric recombinant viruses described here. We fused the 21-nt sequence (5′-CCACAAGACAAUCGAGACUUU-3′) of the TYMV P69 gene targeted by amiR159-P69 to the 3′ end of the GFP gene. The GFP-P69 fusion sequence was then inserted in between the NIb and CP genes to generate the p35STuMV-GP69 infectious clone. As a control, the central 4 nts (underlined) of the 21-nt target sequence (5′-CCACAAGACCUGAGAGACUUU-3′) was mutated to give GFP-P69m which was also inserted in the same position of the virus to generate p35STuMV-GP69m. As the 21-nt target sequence is in a non-essential region of the TuMV-GP69 it can be altered without affecting essential viral gene function. We performed serial single nt mutagenesis from the 1st-nt to the 21st-nt of the target site on the TuMV-GP69 infectious clone by PCR mutagenesis and the resulting series of scanning mutants were confirmed by sequencing. A total of 21 single-nt substitution recombinant viruses were generated. Based on the mutation position, the recombinant viruses were named TuMV-GP69mX, in which X refers to the mutation position. For example, the mutant with substitution on the 1st-nt of the target site was named TuMV-GP69m1. To evaluate the efficiency of amiRNA-mediated specific resistance toward wild type and mutant viruses, we have established a standard protocol for virus challenge inoculation and quantitative evaluation of pathogenicity. Our overall aim was to reduce the time for virus maintenance and propagation in host plants so as to minimize possible virus evolution. All recombinant viruses were propagated from DNA infectious clones. Aliquots of 20 µL, containing 1 µg of DNA in sterilized water, were mechanically applied onto carborundum-dusted leaves of Chenopodium quinoa Willd with a sterilized glass spatula. Seven days post-inoculation (dpi), local lesions were obtained on inoculated leaves. Viruses were then isolated from single lesions and transferred to N. benthamiana for amplification. Four dpi leaves of N. benthamiana with viral infection symptoms were used as the source of inoculum to challenge WT and amiR159-P69 N. benthamiana plants for evaluation of virus pathogenicity (i.e., frequency of break-down). Twenty amiR159-P69 plants were used for each experiment, and the experiments were repeated 3 times. Resistance efficiency of amiR159-P69 plants challenged with recombinant viruses were compared with those obtained with control viruses, including TuMV-GFP, TuMV-GP69 and TuMV-GP69m. Pathogenicity was evaluated between two and four times for each one of the 21 TuMV-GP69mX recombinant viruses. Count data from experiments that were statistically homogeneous were pooled into a single set, whereas experiments that behaved as outliers were removed from the dataset. In plants displaying symptoms, it was important for us to verify the sequence of the 21-nt target site to ensure that no other mutations had occurred to confound our results. To this end, total RNA was extracted from infected leaf tissues using the Trizol reagent (Invitrogen). One µg total RNA was used for reverse-transcriptional polymerase-chain reaction (RT-PCR) with PTuNIb-8671 (5′-GAACCAGCTCAAGAGGATCT-3′) and MTuCP-8982 (5′-GCCACTCTCTGCTCGTATCTTGGCACGCGC-3′) for amplification of the viral region between the partial NIb and the CP genes (Fig. 1B). The PCR fragments then were analyzed by sequencing. The pathogenicity of different recombinant viruses was estimated as the frequency of infected plants out of the total number of inoculated plants. The LaPlace's point estimator for the Binomial frequency parameter was used instead of the commonly used maximum likelihood estimator [29]. The LaPlace method provides a more robust estimate of the Binomial parameter for small sample sizes [30]. Binomial 95% confidence intervals (CI) were also computed. TuMV-GP69mX recombinant viruses were classified into groups of similar pathogenicity using the two-step cluster analysis [31]. In brief, this method classifies data in groups that minimize the within-group variance whilst maximizing the among-groups variance. The method starts with the simplest model (i.e., all viruses are equally pathogenic) and computes its likelihood; then, it classifies the mutants into two clusters and computes the likelihood of this model; finally, it does the same for three clusters, four clusters and up to 21 clusters (i.e., each site behaves in a different way and no classification is possible). For each model, Schwarz's Bayesian information criterion (BIC) was used as a measure of the goodness-of-fit and the model with the lowest BIC was considered to be the best one [32].

10.1371/journal.pntd.0007703

Chikungunya virus requires cellular chloride channels for efficient genome replication

Chikungunya virus (CHIKV) is a re-emerging, pathogenic alphavirus that is transmitted to humans by Aedes spp. mosquitoes—causing fever and debilitating joint pain, with frequent long-term health implications and high morbidity. The CHIKV lifecycle is poorly understood and specific antiviral therapeutics or vaccines are lacking. In this study, we investigated the role of host-cell chloride (Cl-) channels on CHIKV replication.We demonstrate that specific pharmacological Cl- channel inhibitors significantly inhibit CHIKV replication in a dose-dependent manner, suggesting that Cl-channels are pro-viral factors in human cells. Further analysis of the effect of the inhibitors on CHIKV attachment, entry, viral protein expression and replicon replication demonstrated that Cl- channels are specifically required for efficient CHIKV genome replication. This was conserved in mosquito cells, where CHIKV replication and genome copy number was significantly reduced following Cl- channel inhibition. siRNA silencing identified chloride intracellular channels 1 and 4 (CLIC1 and CLIC4, respectively) as required for efficient CHIKV replication and protein affinity chromatography showed low levels of CLIC1 in complex with CHIKV nsP3, an essential component of the viral replication machinery. In summary, for the first time we demonstrate that efficient replication of the CHIKV genome depends on cellular Cl- channels, in both human and mosquito cells and identifies CLIC1 and CLIC4 as agonists of CHIKV replication in human cells. We observe a modest interaction, either direct or indirect, between CLIC1 and nsP3 and hypothesize that CLIC1 may play a role in the formation/maintenance of CHIKV replication complexes. These findings advance our molecular understanding of CHIKV replication and identify potential druggable targets for the treatment and prevention of CHIKV mediated disease.

Chikungunya virus (CHIKV) is a mosquito-borne virus that infects humans and causes chikungunya fever—characterized by fever, rash and chronic arthralgia. Treatment of chikungunya fever is limited to the alleviation of the symptoms and no vaccine is available to prevent infection. Consequently, new anti-CHIKV therapies or targets are urgently required. Since its initial isolation in Tanzania in 1952, CHIKV has spread widely across tropical/sub-tropical and more temperate regions. It has been responsible for epidemics in >70 different regions of the world, resulting in high morbidity and financial burden. Despite this, the CHIKV lifecycle is poorly understood. Here, we demonstrate for the first time that CHIKV requires host-cell ion channels that mediate chloride ion (Cl-) flux through the membranes of infected cells, to complete its lifecycle. Specifically, using pharmacological compounds, we show that Cl- channels are required for efficient replication of the virus genome, identify two specific channels (CLIC1 and CLIC4) required for replication and demonstrate that CLIC1 may interact with viral non-structural protein 3 (nsP3)—an essential component of the CHIKV replicase complex. These findings advance our understanding of CHIKV replication and identify potential druggable targets for the treatment and prevention of CHIKV mediated disease.

Chikungunya virus (CHIKV) is a mosquito-borne virus of the Alphavirus genus in the Togaviridae family. It was first isolated during an outbreak in Tanzania in 1952 [1], since which its geographic range has expanded globally to include almost 40 countries [2]. Following transmission, CHIKV replicates in the fibroblasts of the dermis and disseminates through the blood-stream to several tissues including muscle, joints and the liver [3]. CHIKV causes chikungunya fever which is characterized by high fever, maculopapular rash, myalgia and debilitating arthralgia [4]. In some cases more severe symptoms occur—including encephalitis, encephalopathy, myocarditis, hepatitis and Guillain Barré syndrome [5, 6], however chikungunya fever-associated death is rare [7]. Due to the chronic debilitating symptoms and sequela that persist for up to 3 years [8], CHIKV has a major impact on morbidity and loss of economic productivity within at-risk populations [9]. Treatment of CHIKV-associated disease is limited to the relief of symptoms with no licensed vaccines or direct acting antivirals currently available. CHIKV is a small, enveloped virus, with a single-stranded positive-sense RNA genome of ~12 kilobases. The genome possesses a type-0 5’ 7-methyl-GpppA cap, a 3’ poly(A) tail and two open reading frames (ORFs). The first ORF (ORF-1) encodes the non-structural polyprotein P1234 that is processed to yield four mature non-structural proteins (nsP1-4). The second ORF (ORF-2) encodes the structural polyprotein that is processed into the capsid protein, E3, E2, 6K, and E1. Virus attachment onto mammalian cells is mediated by the CHIKV E2 protein and the cell adhesion molecule Mxra8 [10]. Internalization is achieved via clathrin-mediated endocytosis, although clathrin-independent pathways have also been identified [10, 11]. Following CHIKV trafficking to early endosomes, the E1 glycoprotein facilitates endosomal fusion and capsid release into the cytoplasm [11, 12]. Replication of the viral genome is not well studied but through analogy to other alphaviruses proteolytic cleavage of P1234 in cis by the protease function of nsP2 releases the RNA-dependent RNA polymerase nsP4, initiating the synthesis of minus-strand RNA. Subsequent proteolytic cleavage of the remaining P123 polyprotein initiates synthesis of genomic and sub-genomic RNAs from the minus-strand template [13]. Viral RNA replication occurs in membrane-bound replication complexes, termed spherules, located at the plasma membrane [13, 14], facilitating an optimal environment for replication and protection of dsRNA intermediates from host cell detection. The structural polyprotein is translated from sub-genomic RNA and is co- and post- translationally cleaved by viral and host proteases. Virus assembly and budding takes place at the plasma membrane. The regulation of ionic homeostasis, mediated through cellular ion channels, has emerged as a requirement for a number of virus infections [15]. It has previously been shown that two pore domain (K2P) potassium channels play a role during the trafficking of Bunyamwera virus (family Peribunyaviridae) in endosomes [16, 17] and that Hazara virus (family Nairoviridae) similarly requires endosomal potassium to permit release from endosomes [18]. Ebola virus (family Filoviridae) is known to require endosomal calcium channels for entry into its host cells [19], while entry of Influenza virus (family Orthomyxoviridae) depends on binding of the hemagglutinin to the voltage-dependent calcium channel Cav1.2 [20]. Hepatitis C virus (family Flaviviridae) replication is inhibited by Cl- channel blockers [21] whilst the Cl- channel CLC6 was identified as a pro-viral factor in a genome-wide loss of function screen during CHIKV infection [22]. In this study, we used a panel of ion channel modulating compounds to demonstrate that CHIKV requires Cl- channel activity during its lifecycle, in both mammalian and mosquito cells, for efficient replication of the viral genome. Through RNAi silencing, two Cl- intracellular channels, CLIC1 and CLIC4, were identified as pro-viral factors for CHIKV replication, with CLIC1 potentially interacting with CHIKV nsP3—further implicating this channel as a significant host cell factor during CHIKV infection. These findings expand our understanding of CHIKV pathogenesis and reveal Cl- channels as a potential host cell target for the development of much needed CHIKV antivirals. Huh7 cells (hepatocytes derived from human hepatocellular carcinoma) and BHK-21 cells (fibroblasts derived from Syrian golden hamster kidney) were a gift from M. Harris (University of Leeds, UK). C6/36 cells (Aedes albopictus larva) were a gift from S. Jacobs (The Pirbright Institute, UK). All cell lines tested negative for mycoplasma. Mammalian cells were maintained in Dulbecco’s modified Eagle’s medium (DMEM, Sigma) supplemented with 10% fetal bovine serum, 100 units/ml penicillin, 100 μg/ml streptomycin and non-essential amino acids (Lonza) in a humidified incubator at 37°C and 5% CO2. Invertebrate cells were maintained at 28°C in Leibovitz’s media (Gibco) supplemented with Tryptose Phosphate Broth (Thermo Fisher Scientific) and 100 units/ml penicillin, 100 μg/ml streptomycin. The CHIKV ICRES and CHIKV-Fluc replicon (SGR), in which the ORF-2 region encoding the structural proteins has been replaced with sequence encoding firely luciferase, cDNA clones were previously described [23]. Both are based on the isolate LR2006 OPY1, representing the East Central South African genotype. CHIKV TST-nsP3 was generated by synthesizing a 129 bp DNA fragment containing the twin-strep-tag (TST) encoding sequence flanked by two SpeI restriction sites (ThermoFisher Scientific). This fragment was cloned into the CHIKV-Fluc SGR using a unique SpeI site in region corresponding to the hypervariable domain of nsP3 (position 5222). The TST-nsP3 fragment was then excised and cloned into the CHIKV ICRES plasmids using the unique KfII and the AgeI restriction sites. All CHIKV plasmids were linearized using NotI and in vitro transcribed using the mMESSAGE mMACHINE SP6 Kit (Ambion). 1 μg of RNA was electroporated into 1.2 × 106 BHK-21 cells using a single square wave pulse (260 V, 25 ms). Cells were seeded into a 75cm2 flask and incubated for 48 hrs at 37°C. Supernatant was harvested, clarified by centrifugation for 5 min at room temperature (RT), aliquoted and stored at -80°C. The CHIKV titer was determined by standard plaque assay on BHK-21 monolayers and expressed as plaque forming units/ml (PFU/ml). Huh7 cells were seeded into 96 well plates at 10 000 cells/well (20 000 cells/well for C6/36 cells) one day prior to treatment with increasing doses of Ribavirin (Fluorochem); 4,4'-Diisothiocyanato-2.2'-stilbenedisulfonic acid (DIDS, Sigma); 9-Anthracenecarboxylic acid (9-ACA, Sigma); indanyloxyacetic acid-94 (IAA-94, Cayman Chemical), 5-Nitro-2-(3-phenylpropylamino)benzoic acid (NPPB, Santa Cruz Biotechnology) and dimethyl sulfoxide (DMSO, Fisher Chemical). Compound stocks were dissolved in DMSO and stored at -20°C in aliquots until dilution into complete DMEM/Leibovitz’s for addition onto cells. After 6 hrs and 24 hrs incubation respectively, media/compound was removed and cells incubated in Opti-MEM (Gibco) plus 1 mg/ml 3-(4,5-Dimethylthiazol-2-yl)-2,5-Diphenyltetrazolium Bromide (MTT) for 30 min at 37°C. After removal of Opti-MEM/MTT, cells were lysed in DMSO and absorbance determined at 570 nm on an Infinite F50 microplate reader (Tecan). Absorbance of samples was normalized to untreated control cells. Three independent experiments were conducted, each repeat consisting of three wells/dose. A two-way ANOVA was employed and a Sidak’s or Dunnett’s multiple comparisons test was performed, comparing each compound dose to the average of untreated samples. The MNTD was defined as the highest compound concentration that does not show a significant reduction in normalized absorbance to the untreated samples. Huh7 cells were treated with MNTDs of Ribavirin, DIDS, 9-ACA, IAA-94, NNPB and DMSO (control) as follows. Huh7 cells were seeded into 12 well plates at 100 000 cells/well. The next day, CHIKV at MOI 2 (2xPFU as determined on BHK-21 monolayers/Huh7 cell number) and the respective inhibitor were diluted in 150 μl complete DMEM and adsorbed to the cells for 1 hr at 37°C. CHIKV/compound was removed, cells washed with PBS and incubated for 12 hrs in DMEM/compound. Supernatant was then harvested the titer of released CHIKV determined by plaque assay. To determine the virucidal activity of the compounds, the inhibitors and CHIKV were diluted into 150 μl with complete DMEM and incubated at 37°C for 1 hr prior to adsorption to the cells. Treatment with DEPC was conducted by diluting CHIKV and DEPC (2 mM final concentration) in PBS and incubating at RT in the dark. CHIKV and compounds were adsorbed to cells for 1 hr at 37°C and were then removed and cells incubated for 12 hpi in complete DMEM. The supernatants were harvested and CHIKV titer determined by standard plaque assay. The impact of the compounds on CHIKV attachment to the cell surface was investigated by performing CHIKV adsorption in the presence of the inhibitors at 4°C for 1 hr to prevent uptake. Cells were then rigorously washed with cold PBS and returned to 37°C in complete DMEM. At 24 hpi supernatants were harvested and CHIKV titer determined by standard plaque assay at 24 hpi. The first two hours of the CHIKV lifecycle were investigated by treating the cells with the inhibitors in 1 ml DMEM 1 hr prior to absorbing the virus in the presence of the inhibitors. After removal of the virus, cells were incubated one further hour in 1 ml complete DMEM plus inhibitors before these were removed by washing in PBS. Cells were then kept in complete DMEM until the titer of the released virus was determined at 12 hpi. C6/36 cells were seeded into 12 well plates at 200 000 cells/well. The next day, cells were infected with CHIKV at MOI 2 (2xPFU as determined on BHK-21 monolayers/C6/36 cell number) and treated with the MNTD of compounds as described for Huh7 cells. The supernatants were harvested and CHIKV titer determined at 24 hpi. All experiments were performed in three independent repeats, each consisting of 2 wells/condition. A one-way ANOVA was employed and a Dunnett’s multiple comparisons test was performed, comparing treated samples to untreated sample. Huh7 cells were seeded into 6 well plates at 300 000 cells/well. The next day, cells were infected with CHIKV (MOI 0.1, 2, 10) in the presence of the MNTD of the inhibitors as described above. After removal of the virus, cells were incubated in complete DMEM/inhibitor until cells were washed in PBS, detached by trypsinization, and lysed in lysis buffer (25 mM Tris•HCl pH 7.4, 150 mM NaCl, 1% NP-40, 1 mM EDTA, 5% glycerol) at 24 hpi. Cell lysate was cleared by centrifugation at >16 0000 x g for 10 mins at 4°C and total protein concentration determined by Pierce BCA Protein Assay (Thermo Scientific). Equal amounts of protein were resolved by SDS-PAGE and proteins transferred to PVFD membranes using semi-dry transfer. Western blots were probed with antibodies against nsP1 (1:1000, rabbit polyclonal, in-house)[24], nsP3 (1:1000, rabbit polyclonal, in-house)[24, 25], capsid (1:1000, rabbit polyclonal, in-house), CLIC1 (1:1000, mouse monoclonal, Abcam, ab77214), CLIC4 (1:200, mouse monoclonal, Santa Cruz Biotechnology, sc-135739) and the housekeeping protein actin (clone AC-15, mouse monoclonal, Sigma) overnight at 4°C diluted in Odyssey Blocking Buffer (Li-Cor). After washing in PBS, the western blot was incubated for 1 hr at RT with the respective secondary antibodies (IRDye 800CW Donkey anti-Mouse; IRDye 680LT Donkey anti-Rabbit; Li-Cor). The membrane was dried an imaged using the Odyssey Fc Imaging System (Li-Cor). Huh7 cells were seeded into 12 well plates at 100 000 cells/well. The next day, cells were washed and incubated in live cell imaging solution (140 mM NaCl, 2.5 mM KCl, 1.8 mM CaCl2, 1 mM MgCl2, 20 mM HEPES pH 7.4, 20 mM Glucose, 1% BSA) supplemented with the MNTD of inhibitory compounds for 50 mins at 37°C. Cells were incubated on ice for 10 mins and Alexa Fluor 488 EGF complex (Invitrogen E13345) was added to a final concentration of 0.8 μg/ml. After 45 mins, cells were fixed in 4% Formaldehyde/PBS and imaged using the IncuCyte ZOOM system (Essen Bioscience). The default software parameters for a 12 well plate (Corning) with a 10× objective was used for imaging 4 fields of view/well. A processing definition was established to automate identification of Alexa Fluor 488 positive objects. The green object count/well was extrapolated by the IncuCyte ZOOM software. Two wells/conditions were analyzed in three biological repeats. A one-way ANOVA was employed and Dunnett’s multiple comparisons test was performed comparing each sample to the untreated sample. A modest, but significant decrease in Alexa Fluor 488 EGF uptake upon treatment with the carrier control (DMSO) was controlled for by additionally comparing the samples to DMSO-treated samples in the Dunnett’s multiple comparison test. The CHIKV-Fluc SGR contains a Firefly Luciferase (Fluc) reporter gene which is expressed from the sub-genomic promotor [23]; therefore, detection of Fluc activity is indicative of CHIKV genome replication. The plasmid was linearized using NotI and in vitro transcribed using the mMESSAGE mMACHINE SP6 Kit (Ambion). 400 ng CHIKV-Fluc SGR RNA was co-transfected with 100 ng Renilla Luciferase (Rluc) encoding RNA using 1 μl Lipofectamine 2000 (Thermo Fisher) into Huh7 cells seeded the previous day into 24 well plates at 50 000 cells/well. During the time of transfection, cells were incubated with Opti-MEM/inhibitor. Cells were lysed in passive lysis buffer (Promega) at 6 hrs post-transfection. The dual luciferase assay was performed according to the manufacturer’s protocol (Promega). The Fluc signal was adjusted to the Rluc signal as follows: (average Rluc signal (untreated control)/ Rluc signal (sample X)) x Fluc signal (sample X). Three independent repeats were performed, with each repeat consisting of 2 wells/conditions, which were analyzed as 2 technical repeats each. A one-way ANOVA was employed and Dunnett’s multiple comparisons test was performed comparing each sample to the untreated sample. Huh7 cells were infected with CHIKV and treated with inhibitory compounds as described above. At 6 hpi, total RNA was extracted from cells using TRI Reagent Solution (Applied Biosystems) according to the manufacturer’s instructions. Strand-specific qPCR (ssqPCR) was performed according to the protocol described by Plaskon and colleagues [26]. Briefly, 500 ng of RNA were reverse-transcribed with gene specific primers (S1 Table) using the SCRIPT cDNA Synthesis Kit (Jena Bioscience) according to the manufacturer’s protocol. 100ng of strand-specific cDNA was used as template for the quantitative PCR performed with the qPCRBIO SyGreen Blue Mix Lo-ROX (PCR Biosystems) with gene specific primers (S1 Table) amplifying a 94 bp region of the CHIKV nsP1 encoding sequence using the following PCR program: 95°C for 2 mins, 40 x (95°C for 5 sec, 60°C for 30 sec), dissociation curve 60°C-95°C as pre-defined by the Mx3005P thermal cycler (Agilent technologies). In vitro transcribed CHIKV ICRES RNA was reverse transcribed and a cDNA dilution series employed as a standard to quantify copy numbers in the respective samples. All experiments were performed in four independent repeats, each consisting of 2 wells/condition. A one-way ANOVA was employed and Dunnett’s multiple comparisons test was performed comparing each sample to the untreated control sample. C6/36 cells were infected with CHIKV (MOI 2) and treated with inhibitory compounds as described above. At 24 hpi, total RNA was extracted from cells using TRI Reagent Solution (Applied Biosystems) according to the manufacturer’s instructions. RNA integrity was confirmed by denaturing ageraose gel electrophoresis before reverse transcription of 1 μg of RNA using the High-Capacity RNA-to-cDNA Kit (Applied Biosystems) according to the manufacturer’s protocol. Quantitative PCR was performed using the qPCRBIO SyGreen Blue Mix Lo-ROX (PCR Biosystems) with primers amplifying a 78 bp region of the CHIKV nsP1 encoding sequence (fwd primer: 5’CCGACTCAACCATCCTGGAT’3, rev primer: 5’GGCAGACGCAGTGGTACTTCCT’3), 100ng of cDNA template and a PCR program as described above. In vitro transcribed CHIKV ICRES RNA was reverse transcribed and a cDNA dilution series employed as a standard to quantify copy numbers in the respective samples. All experiments were performed in three independent repeats, each consisting of 2 wells/condition. A one-way ANOVA was employed and Dunnett’s multiple comparisons test was performed, comparing each sample to the untreated control sample. Huh7 cells were seeded into 12 well plates at 100 000 cells/well. The next day, cells were transfected with 75 pmol/well CLIC1, CLIC4 and control-B (= scrambled) siRNA (Santa Cruz Biotechnology), respectively, with each siRNA representing a pool of three 19–25 nt siRNAs. Transfection was conducted according to the manufacturer’s protocol (Santa Cruz Biotechnology sc-29528) with a siRNA to transfection reagent ratio of 1:1. At 24 hrs post transfection, the transfection was repeated to achieve a clearly discernable reduction in protein expression level. At 48 hrs post initial transfection, cells were infected with CHIKV (MOI 10) as described above and a sample harvested for analysis on western blot. Supernatant was harvested at 24 hpi and standard plaque assays performed to determine viral titer. Cells were lysed and analyzed on western blots as described. For double knock down of CLIC1 and CLIC4, a total of 150 pmol siRNA/well was transfected into the cells. Three independent repeats were performed and CHIKV titers of CLIC1/CLIC4 knock down samples were expressed as percentage of scrambled siRNA control sample per repeat. A one-way ANOVA was employed and Dunnett’s multiple comparisons test was performed, comparing each sample to the scrambled siRNA control sample. Huh7 cells were infected with CHIKV ICRES TST-nsP3 (MOI 10) as described above. At 24 hpi, cells were lysed in lysis buffer (2x, pH 7.2, 20mM PIPES, 240mM KCl, 60mM NaCl, 10mM MgCl2, 2% Triton x-10, 20% Glycerol) and total protein concentration of clarified cell lysate determined by BCA assay (Pierce). Strep-Tactin Sepharose 50% suspension (iba, 2-1201-002) was washed with wash buffer (100 mM Tris-Cl, 150 mM NaCl, 1 mM EDTA; pH 8) before equilibrating with lysis buffer. 1mg/ml total protein was incubated with the sepharose o/n at 4°C. Sepharose was washed once with lysis buffer followed by wash buffer (100 mM Tris/HCl, pH 8.0; X mM NaCl; 1 mM EDTA) with increasing and then decreasing stringency (150 mM, 275 mM, 500 mM, 150 mM NaCl). Sepharose was boiled in standard SDS sample buffer and bound proteins analyzed on western blots as described above. We first assessed well-characterized Cl- channel blockers—including diisothiocyanostilbene-2,20-disulfonic acid (DIDS), 9-anthracene carboxylic acid (9-ACA), indyanyloxyacetic acid 94 (IAA-94) and 5-nitro-2-3-phenylpropylamino benzoic acid (NPPB) for their effects on CHIKV infection. For these assays, maximal non-toxic doses (MNTD) for human hepatoma cells (Huh7) were determined for each compound by MTT assay (S1 Fig). Next, Huh7 cells were infected with CHIKV (MOI 2) in the presence of the MNTD of each compound; at 12 hpi titers of released virus were determined. Treatment with NPPB resulted in an 18-fold reduction in CHIKV progeny compared to untreated cells, while treatment with DIDS and 9-ACA led to an 8-fold decrease in CHIKV titer (p ≤ 0.05) (Fig 1A). IAA-94 had no detectable effects. Ribavirin, shown previously to be active against CHIKV [27, 28], was included as a positive control and significantly inhibited production of CHIKV virions (56-fold decrease). To confirm these findings, the expression of CHIKV nsP1 and capsid protein in the presence each Cl- channel inhibitor were determined by western blot analysis. Fig 1B shows reduced expression of nsP1 and capsid in DIDS, 9-ACA and NPPB-treated cells compared to untreated/DMSO-treated cells and only in the presence of IAA-94 were nsP1/capsid levels unaffected. DIDS, 9-ACA, NPPB and Ribavirin were found to inhibit CHIKV replication in a dose dependent manner (Fig 1Ci–1Civ). This data further implies that functional Cl- channels are required during the CHIKV lifecycle. The inhibitory effect on viral protein expression was consistent in cells infected with greater and smaller MOIs (S2 Fig). To exclude any direct virucidal activity of the compounds, CHIKV virions were incubated with the MNTD of each compound in vitro at 37°C for one hour prior to their dilution in media and infection onto untreated cells (Fig 1D). DEPC was included in this assay as a positive control, as it has previously been shown to modify histidine’s on CHIKV glycoproteins, preventing viral fusion [29]. Treatment with DIDS, 9-ACA, NPPB and Ribavirin had no significant effects on the CHIKV titer following direct virion treatment, implying that the compounds do not inactivate the viral particle. Taken together, these data provide the first reported requirement for DIDS-, 9-ACA- and NPPB-sensitive cellular Cl- channels during the CHIKV lifecycle. We next sought to identify the stage of the CHIKV lifecycle that is sensitive to Cl- channel inhibition. The CHIKV lifecycle begins with virus attachment to host cells mediated by the CHIKV E2 protein. To analyze cell attachment, CHIKV (MOI 2) was adsorbed onto cells at 4°C (to prevent virus internalization) for 1 hr in the presence of either DIDS, 9-ACA, NPPB or Ribavirin. Both virus and compound were then removed by rigorous washing and cells were warmed to 37°C to permit the internalization of attached virus particles. Virus infection efficiency was lower in this assay compared to infections carried out at physiological temperature. Thus, cell supernatants were assayed for infectious CHIKV progeny at 24 hpi (Fig 2A), rather than 12 hpi as in other assays. Treatment with 9-ACA, NPPB and Ribavirin had no influence on CHIKV titers, indicating that these compounds do not impact CHIKV E2 mediated cell surface attachment. DIDS had a modest but non-significant effect on virus binding, implying it may partially impact this lifecycle stage. The first CHIKV particles fuse with endosomal membranes as early as 2 mins post-infection, making CHIKV entry a rapid process (12). Indeed it has been reported that almost all CHIKV particles escape the endosomal system within the initial 22 mins of infection. Following endosomal fusion, the viral capsid is released, genome is translated and replication complexes begin to form. We investigated the requirement of Cl- channels during this early lifecycle stage by adsorbing CHIKV onto pre-treated cells for 1 hr at 37°C (MOI 2). Virus and compound were then removed and non-cell associated viruses were washed from cells. Cells were treated with each compound for a further 1 hr and after their removal, CHIKV titers were assessed at 12 hpi. As shown in Fig 2B, 9-ACA, NPPB and Ribavirin had no effect on CHIKV infection when treatment was limited to the early lifecycle stages. DIDS-treated cells, however, displayed reduced CHIKV titers (p ≤ 0.05), consistent with its modest effects on virus attachment. Some of the inhibitory effect associated with DIDS-treatment appeared to be non-specific and associated with general inhibition of clathrin-mediated endocytosis, since cells accumulated lower levels of Alexa Fluo 488 EGF following DIDS-treatment (Fig 2C). NPPB-treatment in this assay led to a small increase in EGF uptake by an as yet undescribed mechanism. Taken together, these data indicate that neither 9-ACA or NPPB inhibit CHIKV at the stages of virus attachment or internalization, whilst DIDS inhibits CHIKV early attachment and internalization. We then investigated the effects of the Cl- channel inhibitors on efficient CHIKV RNA replication using a CHIKV-Fluc SGR (Fig 3A) and Fluc luminescence as a proxy of CHIKV RNA synthesis. For these assays, cells were briefly pre-treated with the MNTD of DIDS, 9-ACA, NPPB or Ribavirin before co-transfection with the CHIKV-Fluc SGR and a 5’capped Rluc mRNA in trans, to control for differences in transfection efficiency. Cells were assessed for CHIKV replication by the measurement of luciferase activity 6 hrs post-transfection. NPPB-treatment significantly inhibited CHIKV-Fluc SGR replication 16-fold (p ≤ 0.0001) while DIDS-treatment led to a 9-fold decrease in CHIKV-Fluc SGR replication. As expected, ribavirin also significantly lowered CHIKV-Fluc SGR replication 37-fold (p ≤ 0.0001). 9-ACA treatment produced only a comparatively modest ~3-fold reduction (p ≤ 0.01) in CHIKV-Fluc SGR replication (Fig 3B). To confirm these findings, we quantified CHIKV genomic (Fig 3C) and intermediate complimentary minus-strand (Fig 3D) RNA copy number (Fig 3C) by ssqPCR. For these assays, cells were infected (MOI 2) in the presence of each Cl- channel inhibitory compound and total RNA extracted at 6 hpi. Significantly lower levels of the CHIKV genomic RNA were observed in cells treated with DIDS (p ≤ 0.05), 9-ACA (p ≤ 0.05), NPPB (p ≤ 0.05) or Ribavirin (p ≤ 0.01) (Fig 3C), consistent with the SGR data (Fig 3B). Similarly, copy numbers of the intermediate complimentary minus-strand RNA were significantly reduced with DIDS- (p ≤ 0.001), 9-ACA- (p ≤ 0.01), NPPB- (p ≤ 0.001) and Ribavirin- (p ≤ 0.0001) treatment. From this data, we inferred that the efficiency of CHIKV genome replication is dependent on the function of Cl− channels that are sensitive to DIDS, 9-ACA and NPPB–via either a direct role in replication of the CHIKV genome or indirectly through upregulation of ORF-1 translation. All of the Cl- channel compounds used in this study inhibit a broad range of cellular Cl- channels. To date, ≥40 Cl- channels have been identified that fall into six classes; 1) Cl- intracellular channels (CLICs), 2) voltage-gated Cl- channels (CLCs), 3) cystic fibrosis transmembrane conductance regulator (CFTR), 4) calcium-activated Cl- channels (CaCCs), 5) ligand-gated Cl- channels and 6) volume-regulated Cl- channels (VRAC) [30, 31]. We focused on CLIC1 and CLIC4 as candidate channels, due to their known sensitivity to DIDS, 9-ACA and NPPB and involvement in other virus infections such as hepatitis C virus and Merkel cell polyomavirus [21, 32]. Initially, we confirmed by western blot that CHIKV infection did not induce expression of CLIC1 or CLIC4 (S3 Fig). CLIC1 and CLIC4 were then silenced using siRNA (Fig 4Ai), cells were then infected with CHIKV (MOI 10) and released virus titers determined at 24 hpi. Compared to control siRNA transfected cells, in those with reduced CLIC1 or CLIC4 expression levels, replication of CHIKV was significantly inhibited (p ≤ 0.05 and p ≤ 0.001, respectively) (Fig 4B), confirming a role for both CLIC1 and CLIC4 in efficient CHIKV replication. Co-silencing of both CLIC1 and CLIC4 (Fig 4Aii) did not result in synergistic inhibition of CHIKV replication (Fig 4B), implying that both channels are required for the same or a related stage in the virus lifecycle. Given the inhibitory effects of chloride channel inhibition on CHIKV genome replication (Fig 3), we reasoned that the channels may form part of the virus replicase complex. To assess this, cells were infected with CHIKV twin-strep-tag nsP3 (CHIKV TST-nsP3) (MOI 10), which encodes a 28 amino acid TST-tag within the hypervariable domain of nsP3. At 24 hpi, cells were lysed and TST/strep-tactin affinity pull-downs performed—whereby the TST-tag of nsP3 selectively binds to Strep-Tactin Sepharose. Following elution, proteins interacting with CHIKV nsP3 were identified by western blot analysis. Negative controls for this analyses included cells infected with wild-type CHIKV (i.e. expressing untagged nsP3) and uninfected cells. We observed that untagged nsP3, capsid, and to lesser extent CLIC1, bound non-specifically to the Sepharose resin, albeit to low levels (Fig 4C). However, compared to both controls, enriched levels of CLIC1 were evident in complex with TST-nsP3—implying low levels of direct/indirect interaction between CLIC1 and nsP3. CLIC4 was not identified as an nsP3 interacting partner, suggesting an independent function outwith the CHIKV replicase complex. Previously published immunoprecipitation studies demonstrated an interaction between alphavirus capsid protein and nsP3 [33], which was included as a positive control to validate the pull-downs (Fig 4C). Taken together, these data suggest that CLIC1 interacts either directly or indirectly with nsP3, an essential component of the CHIKV replication machinery. As an arbovirus, CHIKV infects both humans and mosquitos. We thus investigated if the dependency of CHIKV on cellular Cl- channels is limited to human cells or extends to those of its mosquito host. To assess this, C6/36 cells derived from the larva of Ae. albopictus were infected with CHIKV in the presence of the MNTD (S1B Fig) of NPPB, the Cl- channel inhibitor that exhibited the most significant inhibitory effects on CHIKV replication in mammalian cells (Figs 1 and 3). A 357-fold reduction in CHIKV titer, following infection (MOI 2) of NPPB-treated C6/36 cells, was observed at 24 hpi (p ≤ 0.001), implying that Cl- channels are required during the CHIKV lifecycle in mosquito host cells. Ribavirin-treatment, used as a positive control, led to a 39-fold decrease in CHIKV titer (p ≤ 0.001) (Fig 5A). The effects of the other Cl- channel inhibitors (DIDS, 9-ACA and IAA-94) tested in mammalian cells, were also investigated in infections of C6/36 cells (MOI 2) (S4 Fig). Only DIDS-treatment led to a significant reduction in CHIKV replication at 24 hpi. As this could have been due to a similar non-specific effect as observed in Fig 2.—DIDS, 9-ACA and IAA-94 were not pursued further in C6/36 cell infection studies. By analogy to the data from mammalian cells, the effects of NPPB on viral genome replication were investigated using the CHIKV-Fluc SGR. Though the cytotoxic effect of CHIKV-Fluc SGR transfection in combination with NPPB-treatment prevented replicon analysis, CHIKV RNA copy numbers could still be determined by reverse transcription qPCR 24 hpi. Viral RNA copy numbers were reduced 137-fold in NPPB-treated cells (Fig 5B) and 15-fold in Ribavirin-treated cells (p ≤ 0.0001). Taken together, these data demonstrate that NPPB-sensitive Cl- channels are involved in the CHIKV lifecycle in both human and mosquito host cells. In this study we demonstrate that cellular Cl- channels have a significant pro-viral role during CHIKV infection. Our data shows that Cl- channel modulators inhibit efficient CHIKV RNA synthesis and that the intracellular Cl- channels CLIC1 and CLIC4 are specifically required for CHIKV replication. CLIC1 was found in complex with CHIKV nsP3, albeit to low levels, suggesting a direct involvement in CHIKV replication complex formation or nsP3-mediated viral functions. For the first time, these data highlight the requirement of cellular ion channels during the CHIKV lifecycle. Cl- channels at the cell membrane regulate an array of cellular processes including cell-volume control, fluid transport and cell excitability. Intracellularly, Cl- transport across organelle membranes regulates endosome, lysosome and Golgi acidification. We observed inhibitory effects for three out of four assessed Cl- channel inhibitors on CHIKV replication in mammalian cells (Fig 1A). The tested compounds do not display virucidal activity (Fig 1D) nor influence virus binding or the early stages of the viral lifecycle (Fig 2). Intriguingly, 9-ACA- and NPPB-sensitive Cl- channels were required specifically for efficient RNA replication—as evidenced by the effect of the inhibitors on the CHIKV-Fluc SGR and copy number of the CHIKV genomic and intermediate minus-strand RNA (Fig 3). NPPB-sensitive Cl- channels appear to have a conserved function in mosquito cells, as application of the inhibitor to infected C6/36 cells led to a reduction in CHIKV replication (Fig 5). To identify specific Cl- channels required for CHIKV replication, we focused on two CLIC family members, namely CLIC1 and CLIC4. CLIC proteins are small proteins (236–253 aa, with the exception of CLIC5B and CLIC6), harboring an N-terminal GST-like domain and a C-terminal alpha-helical domain (reviewed in [34]). CLICs are metamorphic proteins that reversibly alternate between soluble cytoplasmic and membrane-associated forms, by rearrangement of the GST-like fold under oxidative conditions. Consequently, CLICs are multifunctional, exerting GST-like enzymatic functions as well as functions in membrane trafficking, endosomal sorting and functioning as Cl- channels. In Merkel cell polyomavirus infected cells, both CLIC1 and CLIC4, mediate small T antigen induced cell motility via their Cl- channel activity [32]. In this study, CLIC1 and CLIC4 silencing inhibited CHIKV replication and, given that these effects mirrored those of the Cl- modulators, we reasoned that this effect was likely due to Cl- channel activity. Interestingly, we identified CLIC1 as a potential interacting partner of nsP3—an essential component of the CHIKV replication complex. We thus speculate that CLIC1 is in its membrane-inserted conformation and interacts with nsP3 and/or other ns-protein(s) bound to nsP3 as part of the replicase, to ensure optimal replication conditions within this membrane imbedded complex. These functions may align with the known roles of intracellular Cl- function, namely the regulation of organelle pH to maintain complex stability and genome integrity. In addition, isolation of CHIKV replication complexes has shown that all proteins needed for viral genome replication are present in the membrane fraction, and no soluble proteins are required [35], further supporting a role for membrane associated CLIC1. The double knock down of CLIC1 and CLIC4 did not have a synergistic effect on inhibiting CHIKV replication (Fig 4B) and CLIC4 was not observed to directly interact with nsP3. These results imply that, although required for efficient CHIKV replication, CLIC4 functions through an alternative mechanism or different stage in a pathway to CLIC1. Using a combination of broad acting inhibitors and siRNA screens, the involvement of several Cl- channel/transporter proteins in the hepatitis C virus lifecycle has previously been reported [21], with CLC2, CLC3, CLC5, CLC7 being specifically required for genome replication. A Cl- channel from the same family, CLC6, has been identified as a pro-viral factor in a genome-wide loss of function screen performed with a CHIKV reporter virus [22]. This raises possibility that further Cl- channels are involved in the replication of the CHIKV genome or other stages of the virus lifecycle. For example, it is conceivable that Cl- channels not targeted by the compounds used in this study, or other ion channels, may play a role in post-entry events, leading up to genome replication—i.e. trafficking or uncoating as observed for e.g. Bunyamwera virus [16, 17]. Anion channels have been shown to be important in other viral systems. Cl- channels were identified as part of the Semliki Forest virus, that is closely related to CHIKV, replication complex by quantitative proteomics [36], supporting our CLIC1/nsP3 findings. In addition, the replication of Tomato bushy stunt virus (family Tombusviridae) was shown to depend on Cl—proton exchanger function [37]. Gef1p Cl-silencing, a homologue of the mammalian CLC proteins in yeast, inhibited replication through its downstream effects on Cu2+ homoeostasis, inhibiting the functionality of the viral replicase. Direct interaction of the voltage-dependent anion channel 1 with VP1 and VP3 of infectious bursal disease virus (family Birnavirdae) was shown to stabilize the ribonucleoprotein complex, allowing full activity of the viral polymerase [38]. It may be possible that CLIC1, as an nsP3 interacting protein, is required for CHIKV genome replication for a similar mechanism. Interestingly, CLIC-dependent Cl- efflux has been shown to act during NLRP3 inflammasome activation and signaling [39] and it has recently been shown that the NLRP3 inflammasome is activated in CHIKV infected humans [40]. A small molecule inhibitor of the inflammasome abrogated inflammatory pathology in mice without influencing the CHIKV titer. This could imply that CLIC Cl- channel function is also involved in the inflammatory response to CHIKV infection, in addition to its role in efficient CHIKV genome replication. Inhibitors specific to CLIC1 and CLIC4 may hold potential as both CHIKV antivirals and inhibitors of the CHIKV inflammatory response. Specific inhibitors would exclude possible off-target effects caused by the currently available broad acting Cl- channel inhibitors. In conclusion, for the first time this study identifies Cl- channels as essential host cell factors for efficient CHIKV replication and in conjunction with other recent studies highlights the significance of ion channel modulation as a druggable target to inhibit virus infection. Notably, we show that the requirement for cellular Cl- channels is conserved between human and mosquito host cells and it is likely that other Cl- channels are required at various stages of the CHIKV lifecycle. In general, Cl- channels may represent a potential target for the development of antivirals acting against a broad variety of (arbo-)viruses and encouragingly Cl- channel inhibitors are in clinical or pre-clinical use [30], potentially facilitatatig the development of Cl- channel specific compounds for future anti-CHIKV strategies.

10.1371/journal.pcbi.1004494

Predicting the Effect of Mutations on Protein-Protein Binding Interactions through Structure-Based Interface Profiles

The formation of protein-protein complexes is essential for proteins to perform their physiological functions in the cell. Mutations that prevent the proper formation of the correct complexes can have serious consequences for the associated cellular processes. Since experimental determination of protein-protein binding affinity remains difficult when performed on a large scale, computational methods for predicting the consequences of mutations on binding affinity are highly desirable. We show that a scoring function based on interface structure profiles collected from analogous protein-protein interactions in the PDB is a powerful predictor of protein binding affinity changes upon mutation. As a standalone feature, the differences between the interface profile score of the mutant and wild-type proteins has an accuracy equivalent to the best all-atom potentials, despite being two orders of magnitude faster once the profile has been constructed. Due to its unique sensitivity in collecting the evolutionary profiles of analogous binding interactions and the high speed of calculation, the interface profile score has additional advantages as a complementary feature to combine with physics-based potentials for improving the accuracy of composite scoring approaches. By incorporating the sequence-derived and residue-level coarse-grained potentials with the interface structure profile score, a composite model was constructed through the random forest training, which generates a Pearson correlation coefficient >0.8 between the predicted and observed binding free-energy changes upon mutation. This accuracy is comparable to, or outperforms in most cases, the current best methods, but does not require high-resolution full-atomic models of the mutant structures. The binding interface profiling approach should find useful application in human-disease mutation recognition and protein interface design studies.

Few proteins carry out their tasks in isolation. Instead, proteins combine with each other in complicated ways that can be affected by either the natural genetic variation that occurs among people or by disease causing mutations such as those that occur in cancer or in genetic disorders. To understand how these mutations affect our health, it is necessary to understand how mutations can affect the strength of the interactions that bind proteins together. This is a difficult task to do in a laboratory on a large scale and scientists are increasingly turning to computational methods to predict these effects in advance. We show that by looking at the multiple alignments of similar protein-protein complex structures at the interface regions, new constraints based on the evolution of the three dimensional structures of proteins can be made to predict which mutations are compatible with two proteins interacting and which are not.

The formation of protein-protein complexes plays an essential role in the regulation of various biological processes. Mutations play fundamental roles in evolution by introducing diversity into genomes that can either be selectively advantageous or cause a change in protein affinity that can result in malfunction of the protein interaction network [1, 2]. The Human Genome Project has yielded a wealth of data concerning natural human genetic variation that remains to be fully utilized. For example, it is well known that different people with the same condition often respond differently to the same treatment. A treatment that is effective in one population may have no effect or even be deleterious in another population. Knowledge of how individual subpopulations respond to drugs therefore remains a major bottleneck within the drug discovery process. Understanding how this natural variation within the human genome impacts the protein interaction network is expected to yield insight into this process, provided that the impact of a mutation on the formation of a protein complex can be reliably predicted. The rational design or modification of the affinity and specificity of protein-protein interactions is another challenging issue that has stimulated considerable efforts, as it presents many promising applications, notably for both industrial and therapeutic purposes [3]. Most of these efforts involve the prediction of the effect of a mutation upon the Gibbs free energy change of protein-protein binding (ΔΔG) on a large scale. Quantitatively, ΔΔG values for protein interactions may be measured experimentally by a variety of biophysical techniques [4, 5]. However, these methods are, with few exceptions, inherently low-throughput due to the need to express and purify each individual mutant protein before measurement. Alternatively, deep mutational scanning can be coupled with functional selection to analyze the effect of a large number of mutations on protein binding at specific sites within a protein [6, 7]. Deep sequencing is a very powerful method that has generated impressive insights into residue-specific contributions to protein binding. However, this method is still in its infancy and routine application is still difficult. As a result, scientists have increasingly turned to computational methods to predict ΔΔG values. For a rigid protein, the ΔΔG of folding or protein binding can be determined relatively accurately from a full-atomic description of the protein structure or complex, using either potentials based on molecular mechanics that attempt to quantify the interactions in physically meaningful terms [8], statistical potentials based on the likelihood of similar interactions and local conformations occurring in the PDB [9], or some combination of the two. However, this approach ignores the structural changes that can occur upon mutation, which can alleviate clashes and position residues in conformations more favorable for interaction. Accordingly, much research has gone into the incorporation of flexibility into energetic calculations [8, 10–12]. However, the method is computationally expensive for large datasets to the extent that it becomes prohibitive for genome-wide studies or even scanning mutations on a single protein. In addition, in many cases, a more exact physical representation of the molecular structure and interactions have proved to be less accurate than simpler models due to the inherent inaccuracy of each term in the force-field. As such, alternative methods have been proposed that either use reduced representations of the protein structure or simplified interaction schemes (for example, the use of Cβ and contact potentials) [13, 14] or omit the atomic details of the structure entirely and use machine-learning to predict ΔΔG values from sequence conservation or from gross structural features such as solvent accessibility and secondary structure. The accuracy of a machine learning method ultimately depends on the quality of the feature set and the experimental data available to train the method. If the training set is representative, completely covering all relevant types of interactions and not significantly biased towards specific interactions, it is possible to use machine learning to accurately predict the effect of a mutation using features that are only weakly predictive on their own. If the training set is not representative, then a model formed from only weak predictors is usually not generalizable [15]. The effect of mutations on protein stability has been heavily studied experimentally and non-redundant datasets have been constructed that are believed to be representative of all classes of possible interactions. By contrast, information on the effects of mutations on protein complex formation is much more limited with the data heavily focused on only a few interaction types [16]. For this reason, constructing a machine learning method for the prediction of ΔΔG values for protein complex formation is more difficult than constructing a machine learning method for stability predictions [9]. As a result, the resulting methods generally have a lower accuracy compared to protein stability predictions [9]. Furthermore, the models are usually less generalizable and often show large drops in accuracy when tested on new data not in the training set. This limitation can be overcome if new and more accurate predictors are available for ΔΔG prediction. Because physics based features often share the same limitations, attempts have been made to predict ΔΔG using alternate scoring methods. Based on their success in the prediction of ΔΔG values for protein stability [17, 18], sequence based features have been suggested as predictors of protein-protein interaction ΔΔG values [19]. Protein binding affinity is under evolutionary pressure and we expect residues that contribute strongly to binding energetics to be more strongly conserved than residues which have minimal impact on binding. The conservation of binding residues plays an important role in many “hot spot” prediction programs which seek to identify sites on the interface which strongly influence binding [20]. Taking this approach further, it is likely that the observed distribution of amino acids at a site within the interface reflects at some level the amino acid energetic preferences for binding. Other things being equal, the probability of finding an amino acid which unfavorably impacts binding at an interface site will be less than finding a more favorable amino acid—provided that affinity, and not some other property, is the driving force for selection. However, in many cases there are other driving forces for selection besides protein-protein binding affinity such as binding specificity [6, 21, 22], foldability [23], or protection against aggregation [24]. In addition, closely related sequences bear the imprint of their evolutionary relationship independent of any functional relationship [25]. The limited time of divergence from a common evolutionary ancestor creates a phylogenetic signal that can complicate analysis as not all possible mutations are effectively sampled during the divergence time [26]. Both effects can be reduced by considering structurally similar interfaces rather than closely evolutionarily related proteins. Structurally similar interfaces are expected to serve similar roles regardless of their evolutionary relationship; an effect that can be seen by the existence of highly similar interfaces in proteins that are otherwise structurally dissimilar and evolutionarily distant [27]. Using this approach, we show an interface binding profile score, called BindProf, formed from an aligned ensemble of structurally similar interfaces has accuracy as a standalone feature similar to, or in most cases, better than many composite all-atom potentials. Unlike the all-atom energies, it can be calculated very rapidly once the profile is constructed. The on-line server of the BindProf program is freely available at http://zhanglab.ccmb.med.umich.edu/BindProf. To predict the free-energy change of protein-protein interactions, ΔΔG, BindProf adopts a multi-scale approach shown in Fig 1 using a variety of features at different levels of structural resolution using machine learning with sequence and structure based features to learn the correct weighting between terms using a regression tree classifier. A unique feature of BindProf is the inclusion of a structural profile score reflecting the likelihood of a given sequence being found in the ensemble of structurally similar protein-protein complexes. Since function follows structure more closely than sequence, we expect the structural profile score to more accurately reflect ΔΔG changes than sequence conservation. Such an expectation has been borne out in our protein design program EvoDesign [28, 29], where the structural profile score was found to be the dominant factor in a multi-scale approach that resulted in the majority of tested sequences experimentally folding to the designated structures. Since the most distinctive feature of our approach is the use of structurally similar interfaces of protein complexes in the PDB to score the effect of a mutation, we first consider the most accurate way to predict ΔΔG of protein binding using only this information. The amino acid distribution of structurally similar complexes can be analyzed quantitatively by the use of structural profile scores. Similar to a position specific scoring matrix, a structural profile score F(p, a) reflects the log odds likelihood of an amino acid (A) being found at a particular position (p) in an aligned ensemble of structurally similar proteins [30] F(p,A)=∑a=120g(p,a)M(A,a) (1) where g(p, a) is the Henikoff weighted frequency of the amino acid a appearing at the pth position in a multiple sequence alignment (MSA) with exactly redundant interface sequences removed; M(A, a) is the BLOSUM substitution matrix with a varying for 20 amino acids, which is used to account for missing structures in the PDB. Experimental ΔΔG values are therefore hypothesized to be proportional to the mutant profile score defined as the difference between the profile scores of the wild type (WT) and mutant (Mut) amino acids at position p in the interface: ΔΔGcalc=∑a=120g(p,a)M(AWT,a)−∑a=120g(p,a)M(AMut,a) (2) The profile therefore depends on both the cutoff level for defining a similar complex and the measure of similarity used. The definition of “similar” is less straightforward in regards to interfaces than it is with overall protein structure. Similarity of protein structures can be defined by a normalized, length independent measure of structural difference, TM-score, which has been shown to have a close relationship with fold classification [31, 32]. For interfaces, a straightforward definition is to use the normal procedure for the structural comparison of proteins but to only consider interface residues in the comparison [33]. A similar interface in this case is defined as having a high TM-score when only residues at a given cutoff distance (4 Å) from the other chain are considered for alignment and scoring (iTM-score, see definition in Methods) [33]. A more stringent comparison (Iscore) can be made by considering not only backbone alignment but also contact patterns at the interface to more clearly distinguish closely related proteins [33]. Finally, even close structural matches can result in significantly different binding energetics if there is a mismatch of interaction types at the interface. For example, the mutation of hydrophobic to a charged residue can result in a severe loss of affinity if the mutation is located within a hydrophobic pocket. Accordingly, the alignment can be modified to take into account physicochemical similarity during alignment using a pharmacophore classification of residues to identify residue similarity (PCscore) [34]. In Fig 2 we show the correlation between ΔΔG values calculated by the mutant interface profile scores (Eq 2) and experimental ΔΔG values of protein-protein interactions from the SKEMPI database [16] as a function of the alignment methods and cutoff values. Each method shows the expected general rise and fall in the accuracy at extreme values as the cutoff is made either too strict or too loose. Too loose cutoffs degrade the accuracy of the profile score as structurally unrelated complexes are included in the profile and the specific information from structurally related complexes is lost. Too strict cutoffs, on the other hand, include too few sequences to construct an accurate profile that reflects all the actual allowable possibilities at the interface. While all similarity measures show low accuracy asymptotically at very high and low cutoff values, a simple unimodal distribution of accuracy with the cutoff value is only observed for the profile score formed from structural alignment of the monomeric protein. In this case, the accuracy of the profile score reflects the underlying bimodal distribution of the TM-Score, which has a sharp division near TM-Score cutoff values of 0.5 separating similar folds from unrelated structures [32]. Since TM-Scores of 0.5 and above correspond with high probability to similar folds while a TM-Score below this value indicates essentially no relationship between structures [32], the monomeric profile score is only accurate above a TM-Score 0.5. However, the actual correlation with the experimental ΔΔG values is modest and the profile scores from all interface alignment methods yield a significantly better correlation for nearly the entire range of cutoff values. The relationship between cutoff value and ΔΔG prediction for the interface alignment methods (iTM-score, Iscore, and PCscore) is more complex reflecting a more complex underlying distribution. In each case, the accuracy of ΔΔG prediction is at least bimodal with the cutoff value. Like the monomeric structure profile, the accuracy rises at strict cutoff values. As the cutoff is reduced it levels off as an adequate representation of closely related complexes is built. However, unlike the monomeric structure profile, the accuracy rises again at lower cutoff values, eventually reaching a higher accuracy than can be achieved by profiles constructed from closely related complexes. Closer inspection of the actual origin of this effect is the inclusion of sequences at lower cutoff values that can be aligned accurately to a region within the interface but with relatively poor overall global alignment. From the viewpoint of applications which rely on global properties like the recognition of convergently evolved similar interfaces for function annotation [35–38], these sequences are less useful as they reflect similarity in only a small region of the interface. However, on a physical level, binding interactions are fundamentally local properties. In the interior of a protein, amino acids are tightly packed and a mutation at one site can cause a rearrangement of the protein core [39]. At the interface, however, packing is less tight and a considerable fraction is exposed to solvent even in the protein complex [40]. The difference in packing gives a conformational freedom at the interface that is not present in the interior which can retard the propagation of packing defects throughout the interface after a mutation [41]. With this in mind, the relative inaccuracy of profiles based on PCscore alignment at predicting ΔΔG values can be explained, despite the fact that PCscore is the only method that attempts to incorporate physicochemical similarity into the alignment procedure. Because PCscore penalizes amino acid mismatches more severely, more sequences with good local matches but poor global similarity are missed. Taken individually, sequences with higher interface similarity should be more predictive of ΔΔG then sequences with lower interface similarity. However, the accuracy of the interface profile score is highly dependent on the number of sequences that can be aligned at the site of the mutation. A representative example is shown in Fig 3A. At a high interface similarity cutoff (IScore = 0.25), the accuracy of the profile score rises steeply until about 15 sequences can be aligned at the position, mirroring a similar result for protein stability [28, 29]. At low interface similarity (IScore = 0.2), the number of sequences is less predictive of the accuracy of the profile score, likely because a sufficient number of sequences can be found for all positions except those at the extreme edge of the interface (see below). We therefore considered an adaptive procedure to form a more accurate profile. The sequences are first sorted by descending interface similarity. All sequences with an interface similarity above a strict cutoff are added to the profile and up to n sequences are added until the second, looser cutoff is reached. Fig 3C shows the improvement in ΔΔG prediction from Iscore alignment as a function of n for the optimal high and low interface similarity cutoff values (IScore = 0.25 and 0.19). n reaches a shallow maximum around 80 sequences. The adaptive profile shows a significant improvement over the profile formed from a high similarity cutoff and a smaller improvement over the profile formed from a high similarity cutoff. To assess the potential of interface profile scores for either standalone ΔΔG prediction or as a feature in machine learning based score combinations, we compared the accuracy of interface profile scores formed from high, low, and adaptive profiles by Iscore alignment to a diverse set of multi-scale potential terms. Although iTM-score profiles are slightly more accurate than Iscore profiles at predicting ΔΔG (Fig 1), we chose Iscore profiles for comparison because an additional feature calculated from the profile, the fraction of conserved contacts, can be used to predict the accuracy of the profile score for machine learning. The tested set of potentials includes: the all-atom empirical potential FoldX [42, 43], a composite statistical and physics based potential from Rosetta (Talaris 2013) [44], residue and all-atom docking potentials (PIE [45] and PISA [46], respectively), all atom and Cβ based statistical potentials (DCOMPLEX [47] and RF_CB [48], respectively), a shape complementarity score [49], changes in the total, polar, and hydrophobic solvent accessible surface area (SASA), the difference in hydrogen bond counts across the interface in the structures of the WT and mutant complexes, the volume difference between WT and mutant residues, and pharmacophore count differences of hydrophobic, and aromatic and hydrogen bonding forming residues between the WT and mutant complexes [50]. The Pearson’s correlation coefficient c between predicted and experimental ΔΔG values is shown in Fig 4 for the adaptive interface profile score and the multi-scale potentials described above. When all mutations are considered, the adaptive interface profile score is more accurate at predicting ΔΔG than all the other potentials considered except for FoldX. However, the difference in c between FoldX and the adaptive interface profile score is not statistically significant when using a two-tailed Fischer r-to-z transformation (p-value = 0.32). The difference in c between the adaptive interface profile score and the all-atomic docking potential PISA is also statistically insignificant (p-value = 0.2). The adaptive interface profile score is superior in accuracy to all other potentials tested at high statistical significance (p-value<0.001). From Fig 4, FoldX appears the most accurate single method in terms of Pearson correlation coefficient c although it is statistically indistinguishable with BindProf and PISA. However, this value could be biased somewhat by the fact that the side-chains of the mutant have been reconstructed using the FoldX force field. A mismatch between the force field used to optimize the side-chain rotamers and the scoring potential can result in a degradation of the performance. In our early trials, the Talaris2013 Rosetta force field generally showed similar performance to FoldX values if the side-chains were reconstructed using the Talaris2013 forcefield. We note that although the BindProf score compares favorably with other individual potentials, the Pearson correlation coefficient c is still relatively low (below 0.5). However, one of the key features of BindProf is that it works on a fundamentally different basis then the other methods that are currently in use. This complementarity should be of important help for improving the overall recognition accuracy of multiscale potentials when combined with other sources of potentials as demonstrated below. In many applications it is desirable to know the accuracy of ΔΔG prediction across different categories of experimental ΔΔG values. For example, the accuracy of predicting destabilizing mutations is significantly less important in protein design than the accuracy of predicting favorable mutations, as strongly destabilizing mutants are rejected during the design process. Any inaccuracy in prediction therefore only matters to the extent they are misclassified as favorable or neutral mutations. On the other hand, favorable mutations should be enriched during the design process and accurate ΔΔG prediction is essential for these mutations. We therefore recalculated the Pearson’s correlation coefficient c between experimental and calculated ΔΔG values restricting the dataset to the entries with experimental ΔΔG values within the appropriate range. Interface profile scores show exceptional performance relative to other predictors (c = 0.5) at predicting favorable mutations (ΔΔG values ≤0 kcal/mol, 27% of the total, see Fig 5B). This is an important result as finding favorable mutations is a very important target for many applications, such as protein design to build more tightly binding interfaces, which have so far proven difficult to predict by physics based methods [12, 51]. The most predictive feature in most categories, FoldX, performs poorly here (c = 0.28 compared to c = 0.46 for destabilizing mutations), similar to previous observations which also included a degree of backbone flexibility by incorporating a short relaxation before the calculation of FoldX energies [51]. Likewise, other features like shape complementarity and the statistical potentials DCOMPLEX and RF_CB that normally perform well also perform poorly in this category. This effect is even more magnified when only strongly favorable mutations (ΔΔG values ≤ -1 kcal/mol, 8% of the total) are considered (Fig 5D). Interface profiles are less accurate in predictions of unfavorable mutations (ΔΔG values ≥0 kcal/mol, 75% of the total in Fig 5E), likely because the statistics of unfavorable mutations are based on a lower number of frequency counts within the profile [17]. Full atomic physical potentials (FoldX and the Rosetta’s Talaris2013 score function) and docking potentials (PISA and PIE) do well in this category. Shape complementarity is also predictive of unfavorable mutations (c = 0.31) while it is not predictive of favorable mutations (c = -0.13). All methods were inaccurate in determining the subtle differences between neutral mutations (ΔΔG values between 1 and -1 kcal/mol, 46% of the total, Fig 5G). Fortunately, inaccuracies within this range are usually of less consequence since a mutation with a ΔΔG value between 1 and -1 is often tolerated with little impact on a protein’s function. However, the cumulative impact can be significant when multiple mutations are considered such as in protein design applications. Since all the methods are inaccurate within this range and only a small fraction of mutations are actually favorable, reverting mutations with predicted ΔΔG values >1 kcal/mol back to WT may be a successful strategy for loss of affinity in design proteins through the accumulation of many small errors. We next sought to see if the accuracy of interface profile scores could be predicted from the characteristics of the profile. Interface residues play different roles in protein-protein interactions and display both different conservation patterns and different types of interactions depending on their relative position within the interface [40]. Since the accuracy of both the interface profile scores and the sequence and physics based scores are expected to be sensitive to these changes, it is of interest to compare the accuracy of different methods based on the different types of interface residues. This requires that a standard classification of the roles that different residues play in binding be made, which is difficult if only their geometric position within the interface is considered. Instead, one of the most natural classification of interface residues for binding energetics is determined by comparing the relative solvent accessible area of the residue in the monomeric protein (rASA) to the relative solvent accessible area in the protein complex (rASAc) (Fig 6). Following Levy [40], the “core” residues are defined as residues which are exposed in the monomeric protein (rASA>25%) but buried in the protein complex (rASAc <25%). Core residues are typically hydrophobic with a composition strongly divergent from the composition of the remainder of the protein surface [52]. Core residues supply the bulk of the energy driving association by hydrophobic interactions [53]. The hydrophobic interactions within the complex cause the core region to become tightly packed upon complex association with little room for conformational variability. For these reasons, the core residues are strongly conserved during evolution [53, 54], and mutations in this region are usually more strongly unfavorable when compared to mutations at the periphery of the interface (see Figs 7 and S1). “Rim” residues surround the core residues and are also exposed in the monomeric protein. But unlike the core residues, the rim residues become only partially (0–25% rASAc) buried upon complex formation. The rim residues have a composition more similar to the surface of the protein away from the interface [52]. Rim residues are frequently charged and often engage in hydrogen bonding or salt bridges with the binding partner [53]. The rim residues help to alleviate protein aggregation by charge repulsion and can contribute to binding specificity by forming specific polar contacts with the binding partner. In some cases, the rim residues also tune the strength of binding, stopping the formation of an excessively stable complex which prevents the formation of other complexes within the interaction network. Most of the favorable mutations are found within this region, with the most common favorable mutation being a charge reversal which alleviates an unfavorable electrostatic interaction within the complex. Rim residues show much less sequence conservation than the core residues. Because of their role in the fine tuning of protein interactions and because the rim of the interface is less tightly packed [41] than the core residues, these residues are much less evolutionarily conserved. “Support” residues are partially buried in the monomeric protein, and fully buried in the complex. As such, they are usually hydrophobic and located in the center of the interface near the core residues. However, because the change in surface area upon complex formation for support residues is less than core residues they are less important energetically and are subject to more sequence variation than the core residues. The final two categories of “surface” (rASAc >25% and rASA <25%) and “interior” (rASAc <25% and rASA <25%) consist of residues that make no contacts with the binding partner. Mutations within these regions only influence complex formation indirectly by influencing conformational changes, by destabilizing protein folding [23, 55], or by long-range electrostatic interactions and alteration of the hydrogen-bonding network [56]. Consequently, they generally have a minimal impact on the energetics of complex formation (Fig 7). Since this classification by changes in rASA upon complex formation also indirectly reports on the position of the mutation within the interface, it is expected that the performance of the interface profile score will vary as well. The interface profile score is most accurate for the core residues (Fig 9) which are generally located at the center of the interface (Fig 6). The alignment is significantly more accurate in this region compared to the rest of the interface, especially when the cutoff is restricted to only highly similar complexes (Fig 8). The relative advantage of the interface profile score over methods is decreased when non-core residues are considered. The all-atom physics based potentials Talaris2013 and FoldX were also less accurate in predicting the ΔΔG of mutations outside the core residues, most likely because electrostatic and hydrogen bonding interactions are significantly more difficult to predict by physics-based methods than interactions primarily based on hydrophobic contacts [57]. Instead, the docking potentials PIE and PISA are the most accurate methods for the RIM regions. PIE and PISA are statistical potentials based on the difference in distance distributions between native and incorrectly docked complexes at the residue (PIE) or atomic level (PISA). By contrast, some of the sequence-based features increased in accuracy in the Rim relative to the Core region such as the change in the count of the number of hydrogen-bond donors and acceptors and the number of aromatic residues. Finally, ΔΔG within the interior and surface regions is correlated with the change in hydrophobic and polar interfacial SASA after mutation. Although the correlation is modest here (Fig 9E and 9F), this is an important result as other features performed poorly for these regions. The results above suggest: These features motivated us to combine the interface profile score with other scoring functions which the profile score is complementary to increase the mutation residue recognition. One common approach of the automated feature combination is machine-learning techniques which use features that are weakly predicting on their own but can be combined to give an optimal prediction of ΔΔG. We first examined whether a technique can be constructed using only the information within the interface profiles. We constructed a 13 feature set by considering 3 interface profile scores using profiles made from high and low interface similarity cutoffs (Iscore = 0.19 and Iscore = 0.25) and the adaptive interface profile along with 10 additional features reflecting the quality of the high and low interface similarity profiles. These cutoff levels were selected on the basis of validation on a separate testing dataset comprised of 20% of the data not used in validating the final result. For the high and low interface similarity profiles we calculated additional features, including The first two features report on the relative quality of the alignment of the structural profile; whether the ensemble of aligned structures actually resembles the protein complex under question or not. The last three features measure the information content within the profile and reflect whether the profile is sufficiently diverse to fully reconstruct the mutational landscape of the interaction. A random forest algorithm was then used to predict ΔΔG with these features using repeated 10 fold cross-validation (Fig 10A). Using only the features derived from the interface profile scores, it was possible to get a correlation coefficient of c = 0.71±0.07 (Fig 10A) on the 10 fold cross-validated set. This level of accuracy compares favorably to the accuracy of other state-of-the-art methods [8, 14, 50, 51], despite being two orders of magnitude faster than the molecular dynamics based energy minimization methods [8, 51] and having far fewer terms than other machine learning based models [14, 50]. A true direct comparison, however, is difficult because of the different datasets used in training and different methods of cross-validation for various methods. In particular, our dataset considers both single and multiple site mutations but is only trained on dimeric complexes. A true test at the statistical significance level would require retraining each method with the specific dataset used here. Furthermore, small differences in accuracy in machine learning based methods using large amounts of features may not translate to real differences in accuracy outside of the SKEMPI dataset [16]. Nevertheless, it is possible to conclude that the structural interface profile-based method by itself can give an accuracy comparable to state of the art methods. Among the top performing methods, the Beatmusic method [9] using a combination of 13 statistical potentials weighed by solvent accessibility achieves a correlation coefficient of 0.4 on a non-redundant, single mutation set of the SKEMPI database and 0.68 after the removal of outliers. The residue level contact potential of Moal and Fernandez-Recio [14] achieves a similar performance of c = 0.68 when tested against the SKEMPI subset used here. The interface profile scores and profile-based features can be incorporated with the other potentials to give an even more accurate method. We consider two additional methods, using tenfold cross-validation to confirm the results. The first method uses all the 13 profile features above and the Cβ potentials PIE and RF_CB (Fig 10B). This method has the advantage that the side-chains do not need to be calculated for each position which is the most time-consuming part of the calculation. This method has even greater accuracy than the profile only method (c = 0.80±0.04). Although the dominant feature in terms of determining relative error is the Cβ statistical potential RF_CB, the most important term in terms of node purity is the low interface similarity profile score and the other profile based features are also important features in the approach both in terms of relative error and node purity (Fig 10B right side). If all the terms are considered, the accuracy increases only slightly (c = 0.83±0.05) above the residue-level potential model (Fig 10C left side). In this model, the interface profile scores are still dominant terms (Fig 10C right side). The standard cross-validation normally used to validate the accuracy of machine learning assumes the validation set is a non-biased subset that is representative of the actual population. In reality, the SKEMPI database is a non-representative sample of the actual protein-protein complexes. To test this bias, we performed an additional, stricter cross-validation by holding out all mutants of the proteins being tested during training [50]. This leave one out approach to cross-validation is more realistic than the standard validation process as information on mutants for the specific protein being tested is normally not available and therefore should not be included in the validation procedure. This procedure also has the effect of testing the influence of protein specific information on the model procedure and therefore serves as an indication of the overall generalizability of the model. The results of this procedure performed for the potential including all terms (Fig 10C) is shown in Fig 11 for the 24 proteins that have more than 10 mutants. The standard error of ΔΔG prediction is reported here rather than the correlation coefficient c as the range of ΔΔG values varies substantially among different proteins. For example, the experimental ΔΔG values for three of the proteins (1GC1, 1E22, and 1A22, left side of Fig 11) are mostly near zero (mean |ΔΔG|<0.5), indicating neutral mutations that have little effect on protein binding. The standard error of prediction is therefore more informative in this case as c becomes less meaningful when the values are distributed only within a narrow range. As can be seen from Fig 11, the impact of leaving out the tested protein during training does not have a substantial impact on prediction—the mean standard error across the set increases only slightly from 1.11 kcal/mol to 1.33 kcal/mol. Such minor decrease in accuracy is smaller than the decrease seen with many other machine learning methods. For example, the accuracy of the mCSM method drops from an original cross validated standard error of 1.02 kcal/mol to 1.55 kcal/mol using a similar leave one protein out approach [50]. Overall, this accuracy is still comparable to or higher than most of the much more computationally intensive molecular dynamics based methods explicitly considering conformational flexibility [51, 58, 59]. Like all mutation prediction models, the final machine-learning model has limitations. Many of the limitations are general and apply to any method that attempts to predict ΔΔG values for affinity changes by a structure-based approach. First, the model is trained only to predict ΔΔG values for dimeric complexes where mutations occur only on the side of the interface for individual complexes. While the method can be extended relatively easily to predict mutations for trimers and other types of oligomeric complexes, removing the restriction to search for linked mutations on both sides of the interface simultaneously is more difficult. Many of the terms such as the profile scores, the associated confidence measures of the profile scores, and the pharmacophore counts are strictly linearly additive with respect to the number of mutations. This assumption, which is generally not true for mutations affecting protein stability, is backed by large-scale binding selection mutagenesis experiments showing that the enrichment ratio of double mutants is strongly predicted by the enrichment ratios of the respective single mutations [60]. In these experiments, only one protein is mutated at a time corresponding to mutations on one side of the interface only. When both sides of the interface are mutated, specific interactions such as the formation of a salt-bridge across the interface can cause strong non-linearity when double mutations are compared to the sum of the respective single mutations [61]. However, for most applications one-sided mutations are of the most interest since the binding partner can be assumed to have the WT sequence since mutations are generally rare. Finally, training and testing was performed on the SKEMPI database [16]. This database includes entries for all complexes for which a ΔΔG value and structure are available. The database does not evenly represent the universe of actual protein complexes and some protein complexes and mutation types are heavily represented while others are underrepresented. Exploring other more comprehensive datasets should help further improve BindProf. Protein-protein interactions are critical for nearly every process in the cell and deleterious mutations hindering these interactions can have severe consequences for the associated cellular function. A variety of efforts from personalized medicine to understand viral evolution require knowing how specific mutations effect the protein-protein interactions. Conversely, designing proteins with improved binding or altered specificity requires that the impact of mutations on the native interface be understood. Currently this information is not available experimentally on the proteome-wide scale necessary for these tasks. Towards this end, considerable effort has been devoted towards developing methods to predict the impact of mutations on binding affinity. Most of these approaches rely on physics based methods that attempt to faithfully model on the atomic level the interactions determining protein-protein binding affinity. However, a major obstacle of such approaches is the need for the reconstruction of the full-atomic model for every mutant complex, which limits the accuracy of the approach (since the position of the side-chains is difficult to model) and reduces the computational speed and the range of applications (since rebuilding the full-atomic model is generally the most time-consuming step). In this work, we developed a novel approach, BindProf, aiming to overcome some of these limitations by introducing an interface structure profile based scoring function built on the multiple sequence alignments of analogous protein-protein interactions collected from the PDB. Interface profile scores constructed in this manner can be used as either as a predictor of the Gibbs free energy change of protein-protein binding (ΔΔG) in their own right or combined with other features in a machine learning approach. Considered as a standalone feature, the adaptive interface profile score created by BindProf has an accuracy similar to the best all-atom potentials (Fig 4). However, unlike physics based potentials, the profile scores can be used to score thousands of mutations across a protein-protein interface very quickly (approximately 20 msec per mutation as opposed to an average of 115 seconds, for instance, for building and scoring a full atom complex by FoldX) as once the profiles are constructed the scoring of individual mutants is reduced to a very fast table lookup. In addition, the accuracy of the interface profile score can be inferred from the location of the mutation within the interface and from the characteristics of the structures used to create the profile (Fig 9). This is an advantage over current physics-based methods in which the accuracy is difficult to infer ahead of time. As such, profile scores play prominent roles in composite scoring approaches where they are combined with other features predictive of their accuracy such as the average RMSD for the aligned residues and the sequence entropy within the profile at the mutation position (Fig 10). We therefore expect that interface profiles may play important roles in future composite scoring approaches. The effectiveness of interface profile scoring in predicting binding affinity changes has implications beyond the prediction of ΔΔG values for protein affinity changes. First, the fact that such a method can be constructed at all is independent confirmation of the results of Gao and Skolnick [62] that the existing PDB library is densely connected and approaching completeness with respect to the interface structural space, even if it is not yet complete with respect to the fold space of all possible quaternary structures. If the interface structural space of the PDB library was sparsely connected with few known structural neighbors for each complex, the profile would consist of only a few sequences and the structural profile would not be predictive of ΔΔG values. This effect can be inferred from Fig 2 when only high cutoff values are considered. Second, the degree of correlation between ΔΔG and the interface profile score bears some relationship to the degree that evolution has selected for protein binding affinity at the interface rather than other factors, although the exact relationship is obscured by the limited amount of experimental data available. As more experimental ΔΔG values are measured, profile scoring may help establish the exact role of binding affinity in evolutionary fitness. Overall, the creation of a novel evolutionary based approach with specific characteristics (including high complementarity with physics based scores, high accuracy in finding favorable mutations, low computational cost on a per mutant basis, and a relative insensitivity to side-chain conformation) should find an important application in many biomedical studies including protein design and disease-associated mutation analyses. Experimental ΔΔG values were derived from the SKEMPI database that consists of experimental protein affinity changes upon mutation for protein-protein complexes in which a crystal structure of the WT complex are available [16]. A subset of the database was used for testing of the interface profile scoring and multi-level machine learning. First, the selection was restricted to mutations occurring at one side of the interface to match the normal biological situation in which mutations are relatively rare and it is expected that at least one chain in the complex is WT. Since the interface profile score is fundamentally a property between two protein pairs, only dimeric complexes were selected for analysis from this set, although the method can be extended for the analysis of higher oligomeric complexes. Finally, the SKEMPI database contains multiple entries for a single mutation for 186 entries in this set. These redundant entries were averaged with outlier replicants with ΔΔG values one standard deviation above the mean disregarded. The final dataset contains 1725 entries for 130 complexes. Both single site point mutations and multiple point mutations are considered. For random forest machine learning, three separate training, testing, and validation datasets were constructed. The training set (60% of the data) was used to construct the model, while the testing set (15% of the data) was used to tune the number of variables attempted in each split. The final model was evaluated by 10 fold cross-validation repeated three times on the validation set (25% of the data). Crystal structures were first downloaded from the PDB and stripped of water and all non-protein ligands. A short optimization of the structure of the WT protein complex was then performed to eliminate small clashes and other undesirable features by the RepairPDB function within FoldX [43]. Structures of the mutant complex were then generated from the optimized WT structures by the BuildModel function within FoldX. The temperature for FoldX model building and energy scoring is set to the experimental temperature when known, otherwise it is set to 298 K [16]. For all the sequence and physics based energies except the docking functions PIE [45], PISA [46], and DCOMPLEX [47] and the all atomic energy functions Talaris 2013 [44] and FoldX [42, 43, 63] energies were calculated separately for the mutant and WT complex structures and for both monomeric structures. The predicted ΔΔG values are then equal to: ΔΔGWT→Mut=[EWT(complex)−EWT(monomers)]−[EMut(complex)−EMut(monomers)] (3) where E is the relevant energy function. For the docking functions PIE, PISA, DCOMPLEX, FoldX and the Rosetta Energy function Talaris2013, this calculation is performed internally and ΔΔG is directly proportional to the difference between the energies of the two complexes: ΔΔGWT→Mut=EWT(complex)−EMut(complex) (4) Changes in SASA upon mutation and number of hydrogen bonds across the interface were calculated by the Interface Analyzer in Rosetta [64]. Interface structural alignment was performed using the COTH complex library of non-redundant dimeric structures. To create this library, higher order complexes in DOCKGROUND [65] are first split into all possible combinations of pairwise dimers. This is repeated for all the alternative binding modes contained within the pdb file. All dimers with either chain having less than ten interface residues are removed. The remaining structures are then filtered based on sequence and structure similarity of the complete complex to other complexes in the library. If a dimer shares at least 70% sequence identity and a TM-score at least 0.8 obtained from MM-align [66] to another structure in the complex library, it is removed from the database. The current library contains ~55000 protein-protein complexes. Interface alignment was performed by either Ialign [33] or PCalign [34] program. The iTM-score and Iscore values are calculated by Ialign and PCscore returned by PCalign. The equation for the interface similarity metric iTM-score is a direct analogue of the scoring matrix for TM-score [31] except that only residues within a cutoff depth of 4 Å are considered for the alignment, i.e. iTM-score=1LQ∑i=1Na11+di2/d02 (5) where LQ is the total number of residues in the interface, Na is the number of aligned residues, di is the distance between the Cα atoms of residues at ith aligned residue pair, and d0 is an empirical scaling factor dependent on LQ to ensure the length invariance of the final score [31]. The Iscore is defined similarly except for the addition of a contact overlap factor fi reflecting the fraction of conserved contacts, i.e. Iscore=1LQ∑i=1Nafi1+di2/d02. (6) Here fi = (ci/ai + ci/bi)/2, where ai and bi are the numbers of interfacial contacts of ith aligned residue pair for the template and query complex, respectively, and ci is the number of overlapped contacts. A contact is defined as being overlapped if the residues forming these contacts are aligned in the two pairs of chains. The PCscore is defined analogously to the Iscore with the addition of chemical similarity measure Ii of ith residue pair: PCscore=fcLQ∑i=1Na11+0.25(1−Ii)+di2/42 (7) where fc is the ratio of common contacts between two sets of aligned interfacial residues. Ii equals to 1 if the ith pair of aligned residues are in the same chemical type, or 0 otherwise. To define the chemical equivalency, the amino acids are split into non-overlapping groups of positively charged (K, R), negatively charged (E, D), mixed hydrogen bond donor/acceptors (N, Q, S, T), aromatic (F, W), hydrophobic(C, A, I, L, M, P, V, G) and mixed donor/acceptor or aromatic (H, Y).

10.1371/journal.pbio.1002197

Impaired Mitochondrial Energy Production Causes Light-Induced Photoreceptor Degeneration Independent of Oxidative Stress

Two insults often underlie a variety of eye diseases including glaucoma, optic atrophy, and retinal degeneration—defects in mitochondrial function and aberrant Rhodopsin trafficking. Although mitochondrial defects are often associated with oxidative stress, they have not been linked to Rhodopsin trafficking. In an unbiased forward genetic screen designed to isolate mutations that cause photoreceptor degeneration, we identified mutations in a nuclear-encoded mitochondrial gene, ppr, a homolog of human LRPPRC. We found that ppr is required for protection against light-induced degeneration. Its function is essential to maintain membrane depolarization of the photoreceptors upon repetitive light exposure, and an impaired phototransduction cascade in ppr mutants results in excessive Rhodopsin1 endocytosis. Moreover, loss of ppr results in a reduction in mitochondrial RNAs, reduced electron transport chain activity, and reduced ATP levels. Oxidative stress, however, is not induced. We propose that the reduced ATP level in ppr mutants underlies the phototransduction defect, leading to increased Rhodopsin1 endocytosis during light exposure, causing photoreceptor degeneration independent of oxidative stress. This hypothesis is bolstered by characterization of two other genes isolated in the screen, pyruvate dehydrogenase and citrate synthase. Their loss also causes a light-induced degeneration, excessive Rhodopsin1 endocytosis and reduced ATP without concurrent oxidative stress, unlike many other mutations in mitochondrial genes that are associated with elevated oxidative stress and light-independent photoreceptor demise.

Mitochondrial dysfunction is associated with a number of metabolic and neurological diseases such as Leigh syndrome and progressive blindness. Increased oxidative stress, which is often associated with mitochondrial dysfunction, is thought to be a common cause of disease progression. Here, we identified nuclear genes that encode mitochondrial proteins, whose loss causes the demise of photoreceptor neurons. Contrary to the common idea that this degeneration is triggered by elevated levels of oxidative stress, we find no change in the levels of oxidative stress. We show that activating photoreceptor neurons with light significantly increases energy production, and that this process is required to sustain their activity. Mitochondrial dysfunction impairs this capacity and leads to a premature termination of the light response. This in turn impairs the cycling of the light-sensitive receptor Rhodopsin in photoreceptors, and Rhodopsin accumulates in the cell inducing toxicity. This distinct mechanism of degeneration suggests that different mitochondrial diseases may follow different paths of disease progression and would hence respond differently to treatments.

The causes of progressive dysfunction or death of photoreceptors (PRs) is genetically heterogeneous in humans [1]. PR degeneration is a complex process influenced by numerous genes and environmental factors. Although prolonged exposure to sunlight is one of the major causes of retinal degeneration, more than 200 genes have been associated with retinal diseases in humans [2,3]. Genes associated with retinal diseases affect a variety of cellular processes including phototransduction, cellular stress, metabolism, catabolism, and mitochondrial function [1,3,4]. PR activity is a highly energy-dependent process [5,6], and mitochondrial dysfunction has been implicated in glaucoma, optic atrophy, Leber hereditary optic neuropathy (LHON), and retinitis pigmentosa [1,7,8]. A widely accepted view postulates that increased reactive oxygen species (ROS) levels, resulting from mitochondrial dysfunction, is a major cause of retinal degeneration in human and mouse [9]. According to this model, light triggers mitochondrial activity, leading to increased production of ROS and cellular damage. In Drosophila, the function of several mitochondrial genes has been assessed in PRs. These include Succinate dehydrogenase A (SdhA), a subunit of mitochondrial Complex II [10], Sicily, a protein required for Complex I assembly in mitochondria [11], Opa1, a protein required for inner mitochondrial membrane fusion [12], Aats-met, a mitochondrial methionyl-tRNA synthetase [13], and NnaD, a mitochondrial zinc carboxypeptidase [14]. Consistent with previously published data in mammals [9], all of the mutants in which ROS was assessed have been associated with elevated ROS, suggesting that increased oxidative stress promotes PR degeneration [10–13]. Genetic screens in Drosophila have identified mutations in numerous genes that cause PR degeneration and that are also conserved in human. These mutants can be categorized into two broad groups: those that cause light- and activity-dependent PR degeneration and those that cause light- and activity-independent degeneration. The majority of mutations in genes that are primarily implicated in the phototransduction pathway typically cause light-dependent PR degeneration either due to aberrant Rhodopsin1 (Rh1) trafficking or Ca2+-mediated excitotoxicity [15–17]. However, mutations that constitutively activate the phototransduction pathway, leading to excessive Ca2+ influx, cause light-independent PR degeneration, e.g., loss of function of rdgA [18] or in trpP365, which encodes a constitutively active TRP (Transient Receptor Potential) channel [19]. Light-independent PR degeneration has also been documented for a single fly mitochondrial mutant. The authors showed that the demise of neurons is due to oxidative stress because of the loss of SdhA in mitochondria [10]. Since light dependence has not been tested for the other mutations causing mitochondrial dysfunctions, it is not obvious which mutations cause which type of neurodegeneration, nor what the nature of the insults are that underlie these neurodegenerations. In this study, we show that mutations that impair mitochondrial ATP production without a concurrent increase in oxidative stress exhibit light-dependent PR degeneration. In contrast, mutations that affect ATP production as well as oxidative stress exhibit light-independent PR degeneration that can be exacerbated by light exposure. Furthermore, the observed light-induced PR degeneration in mutants affecting mitochondrial ATP synthesis stems from defects in the phototransduction cascade leading to aberrant endocytosis and delay in the degradation of Rh1. To identify genes required for the maintenance of neurons in the visual system, we performed an unbiased mosaic genetic screen on the X chromosome. We induced large homozygous mutant clones of essential genes in the eyes using the ey-FLP system and screened for age-dependent defects in electroretinograms (ERGs) [20,21]. ERG recordings are induced by light and exhibit “on” and “off” transients (arrow and arrowhead in Fig 1A), indicative of synaptic communication between the PR neurons and postsynaptic cells. They also exhibit a corneal negative response, the amplitude of which signifies the depolarization of PR neurons (dashed line) (Fig 1A). One of the isolated complementation groups, named ppr, pentatricopeptide repeat containing protein (see below), displayed a dramatic reduction in ERG amplitude as well as a loss of “on” and “off” transients in five-wk-old but not 2–3-d-old animals, suggesting a progressive PR degeneration (Fig 1A). The causative mutations of the five alleles of this complementation group were mapped to CG14786 (ppr), an uncharacterized gene in Drosophila (Fig 1B and 1C and S1 Fig). All alleles carry a premature stop codon (Fig 1B and 1C). Two rescue transgenes, a 20 kb P[acman] BAC (P/ΦC31 artificial chromosome for manipulation) CH322-75O21 genomic fragment that contains CG14786 [22] and a 5 kb genomic fragment of CG14786 (Fig 1C), rescue the pupal lethality associated with the loss of ppr. Moreover, ppr mutants (pprA, W150Stop) carrying the genomic rescue transgene (P[acman] BAC CH322-75O21, S1A Fig) show normal ERG amplitudes in aged animals (Fig 1A). The human homolog of ppr is LRPPRC, a mitochondrial protein (Fig 1D) whose loss causes Leigh syndrome [23]. Similar to LRPPRC and other pentatricopeptide proteins, the Ppr protein contains multiple PPR repeats (Fig 1E and S1B Fig; hence ppr). The Ppr protein has a putative amino terminal mitochondrial targeting sequence, as predicted by Mitoprot [24] (Fig 1E). To assess the subcellular localization of the protein, we created transgenic lines carrying a 5 kb genomic rescue transgene in which ppr is tagged at the C-terminus with Green Fluorescent Protein (GFP) (Fig 1C). This construct rescues the lethality of pprA and pprE, is ubiquitously expressed, and the protein colocalizes with a mitochondrial protein, ATP5A (Fig 1F and S1C and S1D Fig). In summary, we identified mutations in a fly homolog of LRPPRC, a protein that is localized to mitochondria and whose loss causes a progressive decline of PR function. To determine whether the progressive age-dependent decay in ERG amplitudes is light-dependent, we raised the flies in constant darkness or a 12 h light/dark cycle for five weeks. The ERG amplitudes of mutant PRs are not affected when the flies are raised in the dark, whereas flies maintained under a 12 h light/dark cycle exhibit severely diminished ERG amplitudes (Fig 2A and 2B). Moreover, the ERG amplitude is dramatically reduced in one-week-old ppr mutant flies if they are maintained under constant light (Fig 2C). Hence, the progressive defect in ERG loss in ppr mutants is induced by light. To assess the morphological features of ppr mutant PRs upon aging and light exposure, we examined cross-sections of the retina by light and Transmission Electron Microscopy (TEM). In the fly eye, PR cells are organized in ~800 ommatidia, and each ommatidium contains eight PR cells (R1–R8). Cross-sections across the retinal PRs reveal the dense microvillar structures of the rhabdomere (Fig 2D arrows), a stack of membranes that are highly enriched in Rh1 and are required for phototransduction [15]. Retina of control, young ppr mutants and ppr mutants reared in the dark for three weeks show very similar morphological features (Fig 2D–2F and S2A–S2F Fig). Moreover, the morphology of PRs of control flies maintained on a 12 h light/dark cycle for three weeks are comparable to young flies (Fig 2G and S2G Fig), whereas the morphology of ppr mutants is highly aberrant (Fig 2H and 2I and S2H and S2I Fig). This phenotype is fully rescued by a genomic rescue transgene (Fig 2J and 2K). Degeneration occurs in PRs R1–R6, which all express Rh1 [25] (blue arrow in Fig 2D), whereas R7 and R8 are spared (red arrows, Fig 2H and 2I; compare to red arrow in 2D). Hence, the residual ERG amplitude in ppr ERG traces may be contributed by R7 and R8. In summary, a light-induced mechanism causes degeneration of PRs in ppr mutants. Although both young and old ppr mutants raised in the dark display normal ERGs, light-dependent PR degeneration typically indicates a defect in the phototransduction cascade [15,16]. To establish if the ppr mutants display defects in the phototransduction cascade, we recorded ERGs upon repetitive pulses of light [26–29]. Flies were kept in the dark for 3–4 min prior to the ERG recordings and stimulated with 10–15 cycles consisting of white light for 1 sec followed by a 1.5 sec dark period (Fig 3A). In ppr mutant eyes, there is a rapid run-down of the amplitude and “on” and “off” responses, whereas control flies show only a very modest reduction in amplitude (Fig 3A). Since this phenotype is activity-dependent, it prompted us to assess the inactivity period (darkness) needed to recover normal ERG amplitude in ppr mutant eyes. We exposed flies to light for 30 sec to maximally reduce the stimulation response in ppr mutant eyes (Fig 3B). Upon a 5–120 sec rest period, we measured the recovery of the ERG amplitude and observed a full recovery to the light response upon a two minute rest period in the dark (Fig 3B). These results demonstrate that Ppr function is required to maintain PR activity. Phototransduction in Drosophila PRs (Fig 3C) is initiated with the conversion of Rh1 to active meta-Rh1 (MRh1) by blue light (~480 nm) [15,29,35,36]. MRh1 triggers a G-protein cascade that activates Phospholipase C (PLC, encoded by norpA), causing hydrolysis of the membrane phospholipid, phosphatidylinositol 4,5-bisphosphate [PI(4,5)P2] (Fig 3C). Hydrolysis of PI(4,5)P2 activates the light-sensitive TRP channel causing a Ca2+ influx, which is essential to depolarize the PRs [29,37–41]. The observed transient depolarization phenotype upon repetitive stimulation, as observed in ppr mutant PRs (Fig 3A and 3B), could be due to impaired PLC activity [27] and/or the inability to quickly regenerate PI(4,5)P2 [28,42], resulting in diminished TRP activity. However, we did not find any evidence for a loss of PLC activity in ppr mutants. Indeed, PLC loss typically impairs PI(4,5)P2 hydrolysis [43] and causes a delay in repolarization [44], neither of which was observed in ppr mutants (Fig 3A and S3A–S3C Fig). To assess if the expression of other proteins required for the phototransduction pathway are affected in ppr mutant PRs, we performed western blots of many key players in the process [15,16,35]. As shown in Fig 4A, none of the proteins tested display altered expression levels in ppr mutant eyes (2–3-d-old, reared in the dark). Hence, ppr does not seem to affect proteins known to be required for the light transduction pathway. In addition, there are no hints of morphological changes in these PRs prior to testing (Fig 4B and 4C). These data indicate that PLC activity is not impaired and that most known players are present. Upon photoisomerization of Rh1 to MRh1 by a photon of blue light, the latter is quickly inactivated by Arrestin2 (Arr2) binding (Fig 3C and S4E Fig) [45,46]. Subsequently, MRh1 is reisomerized to Rh1 by a photon of orange light (~580 nm), leading to the release of Arr2 [32,45]. The mechanism of Rh1 recycling requires Ca2+ influx through TRP channels [32–34,45,47–49]. A small fraction of Rh1/Arr2 complex is endocytosed and degraded [50]. A reduced Ca2+ influx results in increased levels of the Arr2/Rh1 complex, causing excessive endocytosis of Rh1, which is toxic to cells as it stresses the endolysosomal system [47,50–54]. The inability to maintain a sustained light response in ppr mutant eyes (Fig 3A) suggests an impaired Ca2+ influx in PRs [37–40], which in turn may affect the Rh1 cycle and hence lead to an increased internalization of Arr2-bound Rh1 upon exposure to light. Inducing a constitutive Ca2+ influx, however, severely impairs the function and affects the morphology of ppr mutant PRs, even in newly eclosed flies (S3D and S3E Fig), possibly due to synergizing effects of Ca2+ toxicity and mitochondrial stress. To assess whether Rh1 internalization is affected, we performed whole mount antibody staining for Rh1 [47,55]. As shown in Fig 4D–4G, ppr mutants show no defect in the dark but exhibit Rh1 accumulation when exposed to light. Note that although Rh1 is found throughout the rhabdomeres when sections are performed (S4A–S4D Fig), in whole mount preparations Rh1 is detected on the outer rim as the antibodies cannot penetrate the membrane stack. However, the whole mount protocol reveals internalized Rh1 much better (Fig 4D–4G) than stained sections (S4D Fig) [47,55–57]. In addition, brief exposures to blue light followed by orange light cause very similar accumulations of Rh1 in the cytoplasm, indicating a defect in Rh1 cycling in ppr mutant PRs (S4E Fig). Since increased cytoplasmic Rh1 is known to cause degeneration of PRs [52], our data suggest that Rh1 mediates degeneration of ppr mutant PRs. To determine if increased Rh1 internalization in ppr mutant PRs is associated with a defect in Arr2 dynamics, we tested Arr2 translocation to rhabdomeres upon blue light exposure and its release following orange light exposure. To detect Arr2, we expressed Arr2::GFP under the control of the Rh1 promoter (Green or Gray, Fig 4H–4J), which is active in R1–R6 PRs [31]. We generated small ppr mutant mitotic clones with ey-FLP in otherwise heterozygous retina. The mutant ppr PRs (dotted circles, Fig 4H–4J) can be distinguished from wild-type PRs by the absence of RFP (shown in red). Upon blue light exposure, Arr2 translocates to the rhabdomere membranes (binding to Rh1). The subsequent exposure to orange light relocates Arr2 to the cytoplasm as it is released from Rh1 [31,45,52]. In wild-type and ppr mutant clones, Arr2::GFP levels are low in rhabdomeres when flies are kept in the dark (Fig 4H). However, upon exposure to ~1.5 min of blue light, Arr::GFP levels are increased in wild-type as well as ppr mutant rhabdomeres (Fig 4I). To assess the release of Arr2::GFP from rhabdomeres, we kept flies in blue light for 30 min followed by a 60 min exposure to orange light prior to fixation. As shown in Fig 4J and 4K, we observe a higher level of GFP florescence in ppr mutant rhabdomeres than in wild-type rhabdomeres, indicating the slow release of Arr2 in ppr mutant rhabdomeres. Moreover, light-induced internalized Rh1 in ppr mutant PRs colocalizes with Arr::GFP punctae when compared to wild-type PRs (S4F Fig). Hence, our data indicate that impaired dynamics of Arr2 release from Rh1 in ppr mutant PRs leads to increased internalization of Rh1 and toxicity. To test if excessive Rh1 internalization causes PR degeneration in ppr mutant eyes, we examined whether reducing Rh1 suppresses the light-dependent degeneration of ppr PRs. Maturation of Rh1 requires the binding of the chromophore, 11-cis 3-hydroxyretinal, to the opsin moiety. In the absence of the chromophore, opsin is not exported to the rhabdomere but is instead degraded [17,58,59]. In flies, the major source of the chromophore is derived from dietary β-carotene/vitamin A [59]. Indeed, Rh1 levels can be reduced to less than 3% by raising flies in vitamin A-deficient food, and this reduction has been shown to suppress Rh1-mediated PR degeneration [47,56]. Interestingly, under constant light or dark conditions, the ERG amplitude in flies deprived of β-carotene is comparable to those raised on normal food (Fig 5A and S5A Fig). ppr mutants raised in constant light for seven days on normal food display ERG amplitude that is ~ 20% of control (Fig 5A and 5B), whereas ppr mutants raised on vitamin A-deficient food display an ERG amplitude that is ~60% of control. Hence, removal of most Rh1 in PRs (S5A Fig) strongly suppresses the neurodegenerative phenotypes associated with the loss of ppr. To assess whether depriving flies of vitamin A suppresses the morphological alterations of ppr mutant PRs induced by light exposure, we performed TEM of the retina in flies reared in a 12 h light/dark cycle for three weeks. As previously shown [52], rhabdomeres of flies deprived of vitamin A are small, since Rh1 is an important structural component of rhabdomeres (Fig 5D). When raised on normal food, the morphology of ppr mutant rhabdomeres is severely affected (Fig 5C), but PRs of mutant flies deprived of vitamin A are indistinguishable from controls, albeit reduced in size in both cases (Fig 5D and S5B Fig). Combined with the Arr2 data, these results indicate that increased Rh1 internalization is a major cause of PR degeneration in the absence of ppr. We have previously shown that the retromer complex alleviates endolysosomal stress in PRs by preventing Rh1 from entering the endolysosomal pathway. Hence, overexpression of subunits of the retromer promotes its activity and suppresses Rh1-induced endolysosomal trafficking defects in some mutants [56]. Similarly, overexpression of vps35 in PRs suppresses or delays the neurodegenerative defects in ppr mutants (S5C Fig). These data provide further support that aberrant Rh1 internalization/degradation is a major cause of PR degeneration in the absence of ppr. Since Ppr is a mitochondrial protein, and the phototransduction process consumes a significant amount of ATP [5,60,61], we sought to assess whether loss of ppr compromises ATP production. LRPPRC, the human homolog of ppr, and its homologs are required for polyadenylation and stability of mitochondrial RNA (mtRNA) and translation [62,63]. Mitochondrial DNA is transcribed as two long polycistronic precursor RNAs [64]. The precursor RNAs are then processed to create smaller mtRNAs, which are stabilized by the addition of a polyA tail [64,65]. To assess whether ppr is required for mtRNA stability, we quantified the mtRNA levels for 14 transcripts by RT-qPCR and normalized this data to mitochondrial precursor RNA. As shown in Fig 6A, except for Complex I, all mtRNA levels are significantly reduced in mitochondria of ppr mutant larvae, in agreement with a role of Ppr proteins in mtRNA stability [62,63,66–69]. Moreover, the mtDNA content, normalized to nuclear DNA, in ppr mutants is about four times higher than in control larvae (Fig 6B), suggesting that the loss of mtRNA may induce a compensatory increase in mitochondrial biogenesis. Indeed, when we normalize the mtRNA levels with nuclear RNA (RP49), we found an increase in mitochondrial precursor RNA levels (right of Fig 6C) consistent with an increase in mitochondrial biogenesis (Fig 6B). However, normalization of processed mtRNA with nuclear RNA reveals that the mtRNA of ND5, CoI, and CoII are up-regulated and Cyt-b is down-regulated, whereas others are unchanged (Fig 6C). Hence, the overall mtRNA levels in a cell are not dramatically altered. These data suggest the presence of a compensatory response, which induces mitochondrial biogenesis in ppr mutant and can, in part, counterbalance the reduced mtRNA stability per mitochondrion. Given that mtRNAs encode 13 different proteins that are all components of the mitochondrial electron transport chain (ETC) complex (I, III, IV, and V) [13,64], we sought to determine enzymatic activities of individual ETC components from whole cell lysates (Fig 6D) or isolated mitochondria (S1 Data). We also measured Citrate synthase (CS) activity to normalize ETC complex activity. We observed significant decreases in the activities of Complex I, Complex II, and Complex IV in ppr mutant larvae (Fig 6D). The decreased activity of Complex II is striking, as its subunits are encoded in the nucleus [70]. Nevertheless, these data are consistent with the reduced Complex II activity that was observed in LRPPRC knockout mice [63]. Finally, the defects in ETC activity are rescued by a wild-type genomic copy of ppr, showing that the loss of ppr is indeed responsible for these phenotypes. To assess mitochondrial energy production, we measured the rate of oxygen consumption of intact mitochondria in vitro by polarography. In the presence of the Complex I-specific oxidizable substrates malate and glutamate, ppr mutant mitochondria exhibit a significant defect in state III (ADP-stimulated O2 consumption rate), resulting in a decreased respiratory control ratio (RCR), defined as the ratio of state III to state IV (ADP-limiting O2 consumption rate) (Fig 6E). The observed partial deficiencies of several ETC complexes in ppr mutants, combined with the defective respiration of isolated ppr mutant mitochondria (manifesting as reduced state III rate and RCR), are indicative of a reduced efficiency of oxidative phosphorylation (OXPHOS), or in other words, reduced OXPHOS-dependent ATP production [71]. We therefore measured steady state levels of ATP and observed reduced ATP levels in ppr mutant larvae when compared to control animals (Fig 6F). Together, these results provide compelling evidence that ppr regulates mitochondrial RNA levels and thereby affects OXPHOS and ATP levels. PRs are known to consume up to 10% of total ATP in blowflies [60,72]. ATP consumption increases 5-fold above baseline in Drosophila PR in the presence of light [5,6]. Thus, we tested ATP levels in ppr mutant eyes exposed to light for 1 h (Fig 6G) and found that the ATP deficit in mutant PRs is about twice as high (40%) as in the third instar larvae (20%). Furthermore, we investigated the change in ATP levels following light exposure in control and ppr mutant heads. Interestingly, there is a significant increase (48%) in ATP levels in wild type controls upon a 1 h light exposure but only a subtle increase (13%) in ATP level in ppr mutants (Fig 6H). In summary, there is impaired ATP production in the eye of ppr mutants. It has been shown that mitochondrial activity is triggered by Ca2+ influx in neurons [72–75]. We therefore measured changes in ATP following light exposure in mutants that have an impaired Ca2+ influx (trp [38] and norpA/PLC [37,76]). Indeed, these mutants fail to increase ATP levels following light exposure (Fig 6H), suggesting that a Ca2+ influx is required to activate mitochondrial ATP production. These data indicate the presence of a feedback mechanism required for ATP generation to ensure the continuity of the phototransduction process. Besides ppr, the Drosophila genome contains a single other gene that contains PPR motifs, bicoid stability factor (bsf) [66]. RNAi-mediated knockdown of bsf also affects mtRNA stability [66]. Hence, ppr and bsf may be partially redundant. We identified an allele of bsf (bsfSH1181; Fig 7A) that appears to be a null allele, as no Bsf protein was detected in western blots (Fig 7B). Ubiquitous expression of bsf cDNA rescues the pupal lethality associated with bsfSH118. Finally, Bsf also colocalizes with Ppr::GFP (Fig 7C). These data permitted us to compare and contrast the mitochondrial phenotype of ppr mutants to bsf mutants. When we assessed mtRNA levels, as shown in Fig 7D and 7E, bsf mutants show a similar phenotypic profile to ppr mutants although typically more severe. In addition, bsf mutants also show defects in the ETC activity (Fig 7F). Similar to ppr mutants, CII activity is reduced in bsf mutants. However, unlike ppr, bsf mutants display a severe reduction in CIII activity. These data suggest that Ppr and Bsf may play partially redundant functions. To test this hypothesis, we created double mutants. As mentioned before, ppr and bsf mutants cause pupal lethality. However, “ppr–bsf” double mutants die as embryos (Fig 7G), suggesting that Ppr and Bsf are partially redundant. Mitochondrial defects have been shown to cause elevated ROS levels and retinal degeneration in mammals and flies [9–13,77]. We therefore tested if ROS levels are elevated in ppr mutants by staining with dihydroethidium (DHE), a dye which detects superoxide radicals [78,79]. As shown in Fig 8A, ppr mutant clones in eye imaginal discs, marked by loss of GFP, do not show differences in fluorescence intensity when compared to neighboring wild-type tissue. We also performed DHE staining in adult eyes exposed to 24 h constant light. As shown in Fig 8A, the level of DHE staining in mutant eye is similar to control eye (Fig 8B–8C). We also assessed ROS levels by assaying mitochondrial aconitase activity. The native activity of this enzyme is extremely sensitive to elevated ROS [80] and a highly reliable readout in Drosophila [11,77]. As shown in Fig 8D, aconitase activity in mutant animals is comparable to control, suggesting that ROS levels are not affected in ppr mutants. Furthermore, we overexpressed human copper-zinc superoxide dismutase (hSOD1), a potent suppressor of neurodegeneration induced by ROS in flies [12,77,81], in ppr mutant PRs. However, we did not observe a suppression of the degenerative phenotype (Fig 8E), again implying that PR degeneration in ppr mutants is not induced by oxidative stress. Based on our findings, loss of ppr causes reduced ATP production but does not alter steady state ROS levels. However, ppr deficiency causes a severe loss of ERG responses and Rh1 accumulation upon repetitive light stimulation as well as a progressive light-induced PR degeneration. In the genetic screen that permitted the isolation of ppr, we identified mutations in numerous genes whose proteins are targeted to mitochondria [20]. To assess if mutations in genes that have been shown to affect ATP production display similar phenotypes, we evaluated an embryonic lethal allele of knockdown (knd16A) [21], which encodes a homolog of CS, and CG7010 (pdha21A, G170E) [20], which encodes the E1 subunit of Pyruvate dehydrogenase. Loss of CS impairs the tricarboxylic acid (TCA) cycle and hence NADH and ATP production [82,83], whereas Pyruvate dehydrogenase converts pyruvate to acetyl-CoA and mediates entry of glycolytic products into the TCA cycle [84]. Mutant clones in the eyes of knd and pdha show normal primary ERG amplitudes in young flies and flies aged in complete darkness, similar to ppr mutant PRs (Fig 9A). However, a seven-day exposure to light nearly abolishes ERG amplitudes in these mutants, whereas wild type control PRs are barely affected. Hence, loss of kdn or pdha causes a severe light-induced degeneration. In addition, both mutants fail to sustain the ERG amplitude upon repetitive stimulation in young animals (Fig 9B), similar to the phenotypes associated with the loss of ppr (Fig 3A). These observations suggest that perturbations of oxidative metabolism leading to loss of ATP production in both mutants underlie these phenotypes. Given that loss of ppr induces Rh1 accumulation upon light exposure, we tested Rh1-localization in knd and pdha mutant PR of flies raised in the dark. Rh1 localization is indistinguishable from controls in 2-d-old flies (Fig 9C, 9E, and 9G). Similar to ppr mutant PRs (Fig 4G), an ~24 h exposure to constant light leads to a substantial increase in cytoplasmic Rh1 in knd and pdha mutants when compared to controls (Fig 9D, 9F, and 9H). Finally, as shown in S6A and S6B Fig, mutant clones of knd and pdha in eye discs do not show any change in DHE staining, suggesting that loss of these enzymes does not affect ROS production. In summary, the key phenotypes associated with the loss of ppr in the eye are very similar to those of knd and pdha, suggesting a common underlying pathology. Since increased ROS is commonly associated with mitochondrial dysfunction and causes retinal degeneration in flies and humans [9–11,13,77], we tested whether mutations that severely affect the ETC and exhibit a severe increase in ROS levels cause both a light-dependent and light-independent degeneration. Mutations in sicily show a severe reduction in Complex I activity, a reduction in ATP levels (S7A Fig), and a significant increase in ROS production and PR degeneration [11,77]. In dark-reared young flies, ERG amplitudes recorded from sicily mutants are comparable to controls (S7B Fig). When raised in the dark for seven days, sicily mutant eyes exhibit a ~50% reduction in ERG amplitude, whereas the ERG amplitudes of sicily mutant eyes is reduced by ~80% when the flies are kept in constant light for seven days (S7B Fig). These findings suggest that both light-independent and light-dependent mechanisms cause degeneration in sicily mutant PRs. As noted in ppr, kdn, and pdha mutants (Figs 3A and 9B), sicily mutants also show a loss of ERG amplitude upon repetitive stimulation in young animals (S7C Fig). Upon exposure to light for 7 d, we observe an increase in Rh1 levels in the cytoplasm of sicily mutant PRs (S7E Fig) when compared to controls (Fig 9D). Finally, Rh1 localization in dark-reared sicily mutant PRs is indistinguishable from controls (Fig 9C and 9D), similar to what we observed in ppr, kdn, and pdha mutant PRs (Figs 4 and 9). These results indicate that in sicily mutants, increased ROS levels [11,77] promote a degeneration that is exacerbated by Rh1 accumulation upon light exposure. In a forward genetic screen designed to identify mutations in essential genes that cause neuronal degeneration, we identified mutations in numerous nuclear genes that encode mitochondrial proteins. One of these genes corresponds to ppr, a homolog of human LRPPRC (Fig 1). Interestingly, ppr mutant PRs do not degenerate in the dark, in contrast to other mitochondrial mutants such as sicily [11] and sdhA [10], suggesting that a different mechanism underlies the degeneration in ppr mutants. Intriguingly, unlike many other mitochondrial mutants [10–13,77], loss of ppr does not affect ROS levels but impairs ATP production (Figs 6G and 6H and 8A–8D), suggesting that a reduced ATP production underlies the light-dependent degeneration. This hypothesis is supported by the identification and characterization of mutations in two other genes encoding Pyruvate dehydrogenase and CS, which play an important role in the TCA cycle. Both are critical to sustain mitochondrial ATP production [82,83,85–87]. These results, however, do not rule out the possibility that the ratios of other metabolites will be altered because of the different mitochondrial defects, and that these alterations contribute to degeneration. Nevertheless, our results indicate that mutations that affect mitochondrial ATP production without altering ROS levels do not cause PR degeneration in the absence of neuronal activity. This is in sharp contrast with other mitochondrial mutations like sdhA that display increased ROS [10]. Hence, reduced neuronal activity in this subgroup of mitochondrial mutants has neuroprotective effects. In a French Canadian population, mutations in human LRPPRC have been associated with Leigh Syndrome, an autosomal recessive neurodegenerative disorder with onset in infancy [23]. LRPPRC is a key regulator of mtRNA polyadenylation and stability as well as translation [62,63,69], and loss of LRPPRC causes a decrease in mtRNA abundance, defects in translation, ETC activity, and mitochondrial ATP production. In agreement with the phenotypes associated with loss of LRPPRC, we observe a reduction in mtRNA stability in mitochondria of ppr mutants (Fig 6A). We also show that ppr and bsf, the two fly homologs of LRPPRC, play partially redundant roles. We find that CIII activity, which is not affected in ppr mutants (Fig 6D), is significantly lower in bsf than in ppr mutants (Fig 7F). In contrast, CIV activity is significantly down-regulated in ppr mutants (Fig 6D) when compared to bsf mutants (Fig 7F). Surprisingly, both ppr and bsf mutants display a decreased activity of the nuclear-encoded CII (Figs 6D and 7F), a phenotype also observed in LRPPRC knockout mice [63]. We do not know the cause for this reduced CII activity. Reduced CII activity in ppr and bsf mutants as well as in LRPPRC knockout mice may be related to the increase in mitochondrial DNA and transcription, as observed in mouse knockouts for Mterf3 and Tfb1m [88,89]. Finally, we show that mitochondria isolated from ppr mutants show reduced ADP-stimulated oxygen consumption (Fig 6E), suggesting a defect in OXPHOS leading to reduced mitochondrial ATP production (Fig 6F–6H). In summary, features associated with the loss of ppr in flies are similar to what has been described in human cell and mouse experiments [62,63,68,69,90]. Phototransduction is a high ATP-consuming process, and eyes have been estimated to consume 10% of total ATP produced in blowflies [5,60,72]. Moreover, neurons primarily rely on mitochondrial OXPHOS for ATP production [72,91,92]. We show that ATP synthesis increases upon exposure to light in controls suggesting the need for a constant energy supply during phototransduction (Fig 6H). Ca2+ has been shown to activate ATP synthesis in mitochondria, and we observe that blocking Ca2+ influx in PRs also inhibits light-induced ATP production (Fig 6H). Hence, the failure to maintain PR activity during repetitive light exposure in young ppr animals (Fig 3A) may result from reduced mitochondrial activity. In ppr mutant PRs, a defect in Rh1 cycling (S4E Fig), due to reduced Ca2+ influx as predicted by reduced ERG amplitude (Fig 3A), induces excessive internalization of Rh1 (Fig 4G and S4E Fig). Excessive Rh1 internalization is known to overload the endolysosomal system, resulting in neurodegeneration upon prolonged light exposure [47]. Indeed, reducing Rh1 by reducing vitamin A uptake strongly suppresses the PR degeneration associated with ppr mutants (Fig 5A–5D). The observation that the overexpression of the retromer complex protein Vps35, which recycles internalized Rh1 and protects PRs from degeneration [56], partially rescues light-induced ERG phenotypes in ppr (S5C Fig) provides further support that excessive Rh1 mediates degeneration of ppr mutant PRs. Mitochondrial dysfunction is one of the leading causes of neurodegeneration [93]. However, mitochondrial disease-associated phenotypes differ significantly depending on the gene that is affected and the nature of the mutations [94]. Comparing the phenotypes observed in previously characterized Drosophila mitochondrial genes allows us to start subdividing them into more discrete phenotypic groups that can be correlated with the observed physiological defects. For example, sdhA mutants exhibit PR degeneration in the dark and an increase in ROS, yet the ATP levels remain normal [10]. In contrast, ppr, kdn, and pydh mutants exhibit reduced ATP levels [82,83,85–87,90], unaltered ROS levels and their PR only degenerate when exposed to light (Figs 2 and 9A). Mutations that cause reduced ATP production and increased ROS levels may show an intermediate phenotype. sicily mutants show a severe CI deficiency, severely increased ROS levels [11,77] and reduced ATP levels (S7A Fig). Indeed, sicily mutants exhibit a light-independent PR degeneration that is accelerated by light exposure (S7B Fig). Consistent with phenotypes observed in ppr mutants, sicily mutant PRs fail to sustain ERG amplitude upon repetitive light exposure (S7C Fig) and accumulate Rh1 when exposed to light (S7D Fig). These observations suggest that reduced mitochondrial ATP production exacerbates the phenotype induced by excessive ROS production through Rh1-mediated toxicity. In summary, our data suggest that the mechanisms that underlie the neurodegenerative phenotypes in a number of mitochondrial mutants are due to differences in key parameters like ATP production and ROS levels. Obviously, other mechanisms are also likely to play a role in mitochondrial dysfunction-associated neurodegeneration. For ERG recordings, flies were immobilized on a glass slide with glue. A sharp glass-recording electrode, filled with 100 mM NaCl was placed on the surface of the eye, and another sharp glass reference electrode was inserted in the thorax. Field potential recordings were performed after three to four minutes of darkness. The PR response was digitized and recorded using AXON-pCLAMP8.1. To record ERGs from a single stimulation, ~1 sec of light flashed using a halogen lamp (~1,700 Lux). To record ERGs from repeated stimulations, repeated cycles of ~1 sec of light followed by ~1.5 sec of darkness was used. See [104] for a detailed method of ERG recording in Drosophila. Fly heads were dissected and fixed overnight at 4°C in 4% paraformaldehyde, 2% glutaraldehyde, 0.1 M sodium cacodylate (pH 7.2), postfixed in 1% OsO4 for 1 h, dehydrated in ethanol and propylene oxide, and then embedded in Embed-812 resin (Electron Microscopy Sciences). One micron-thick sections were stained with toluidine blue and imaged with a Zeiss microscope (Axio Imager-Z2) equipped with an AxioCam MRm digital camera. Thin sections (~50 nm) were stained in 4% uranyl acetate and 2.5% lead nitrate, and TEM images were captured using a transmission electron microscope (model 1010, JEOL). Images were processed with ImageJ and Adobe Photoshop. See [105] for detailed methods. For immunostaining of larval tissue and adult testis, tissues were dissected in PBS (pH7.2), fixed in 3.7% formaldehyde in PBS for 20 min, and washed in 0.2% Triton X-100 in PBS (PBT). For whole mount staining of fly eyes, heads were prefixed in 4% formaldehyde in PBS for 30 min after removal of the proboscis. Fly eyes were then dissected from these heads, fixed for another 15 min, and washed in 0.3% Triton X-100 in PBS. Fixed samples were blocked in 1X PBS containing 5% normal goat serum and 0.2% Triton X-100 for 1 h (PBTS). Samples were incubated in primary antibody diluted in PBTS overnight at 4°C. For anti- PI(4,5)P2 staining, samples were incubated in primary antibody for two days at 4°C. Samples were washed in PBT, incubated in secondary antibody diluted in PBT for two hours at room temperature, and then washed in PBT prior to mounting. Primary antibodies were used at the following dilutions: Mouse monoclonal anti-Rh1 4C5, DSHB[106] 1:50, Rabbit anti-GFP (Invitrogen) 1:500, mouse anti-ATP synthase α subunit (Complex V; MitoSciences) 1:500, mouse anti- PI(4,5)P2 (Echelon) 1:100. Secondary antibodies conjugated to Cy3 (Jackson ImmunoResearch Laboratories, Inc.) or Alexa Fluor 488 (Invitrogen) were used at 1:500. Phalloidin conjugated with Alexa 488 or Alexa 647 (Invitrogen) 1:250 was added with secondary antibody. Samples were mounted in Vectashield (Vector Laboratories) before imaging with a confocal microscope. Heads from 1–2-d-old flies were used. Samples were processed as described in [57]. Primary antibodies were used at the following dilutions: rabbit Arr2 (1:2000) [46], rabbit RdgC (1:2000) [107], rabbit Trp (1:2000) [108], rabbit Inad (1:2000) [109], rabbit CalX (1:2000) [110], mouse Rh1(1:2000) DSHB [106], rabbit NinaC (1:1000) [111], rabbit PKC (1:1000) [112], mouse Actin (1:5000) (ICN Biomedicals) and Anti-Bsf (1:1000) [113]. All secondary antibodies conjugated to HRP (Jackson ImmunoResearch Laboratories, Inc.) were used at 1:10,000. Total RNA was isolated from control and pprA third instar larvae. Five micrograms of total RNA from each sample were reverse transcribed using Random Hexamer Primers and the High-Capacity cDNA Reverse Transcription Kit (Applied Biosystems). RT—qPCR analysis of the rp49, mitochondrial precursor and mature mitochondrial transcripts were performed in triplicates using 150ng of cDNA per reaction on a 7900HT Real-Time PCR System using ABI SYBR Green PCR Master Mix (Applied Biosystems). An initial activation step for 10 min at 95°C was followed by 40 cycles of 95°C for 10 s and 60°C for 30 s. The primer sequences used are provided in S1 Table. Data is presented as mean ± SD. Fold change was calculated as previously described [114], and statistical significance was determined using a two-tailed Student’s t test (p < 0.05). Method adopted from [115]. Drosophila whole DNA (genomic and mitochondrial) was purified from third instar larvae as the template for PCR. Template DNA was mixed with primers and green supermix reagent (iQ SYBR; Bio-Rad Laboratories). PCR was performed in a thermal cycler (iCycler; Bio-Rad Laboratories), and the data were collected and analyzed using the optical module (iQ5; Bio-Rad Laboratories) and related software following the manufacturer’s instructions. The following primer pairs were used to amplify a genomic DNA fragment corresponding to CG9277/β-Tubulin or a mitochondrial DNA fragment corresponding to CG34083/ND5, respectively: β-Tubulin forward, 5′-CCTTCCCACGTCTTCACTTC-3′; and β-Tubulin reverse, 5′-TTCTTGGCATCGAACATCTG-3′; and ND5 forward, 5′-GCAGAAACAGGTGTAGGAGCA-3′; and ND5 reverse, 5′-GCTGCTATAACTAAAAGAGCTCAGA-3′. Dissociation curves for the amplicons were generated after each run to confirm that the fluorescent signals were not attributable to nonspecific signals (primer-dimers). The mtDNA content (mtDNA/β-Tubulin ratio) was calculated using the formula: mtDNA content = 1/2ΔCt, where ΔCt = CtmtDNAΔ—Ctβ-Tubulin. Enzymatic activity assays were performed on larval whole cell extracts or isolated mitochondria from third instar larvae as previously described [13,116]. Polarography was performed on isolated mitochondria from third instar larvae as previously described [13,117]. Aconitase activity assays were performed in isolated mitochondria from third instar larvae as previously described [13]. ATP level for larvae, eyes, and heads were determined by ATP assay kit (Invitrogen) [118,119]. Flies were exposed to light (~1,800 Lux) for 1 h prior to detection of ATP levels in adult eyes and heads. Eyes were dissected in PBS, and heads were frozen on dry ice and separated on a metal plate kept on dry ice. Five third instar larvae, 20 eyes or 5 heads were dissected and homogenized in 50 μl of 100 mM Tris and 4 mM, EDTA, pH 7.8. These homogenates were snap-frozen in liquid nitrogen and then boiled for 3 min. Samples were then centrifuged, and the supernatant was diluted (1/50 for larvae and 1/2 for heads and eyes) in extraction buffer mixed with luminescent solution. Luminescence was measure on FLUOstar OPTIMA plate reader. DHE staining was performed as described previously [78]. Flies were exposed to 24 h light (1,800 Lux) prior to DHE staining in adult eyes. Percentage protein similarity was determined using BlastP (NCBI). Protein domains were analyzed by PROSITE [120].

10.1371/journal.pgen.1002197

Trans-eQTLs Reveal That Independent Genetic Variants Associated with a Complex Phenotype Converge on Intermediate Genes, with a Major Role for the HLA

For many complex traits, genetic variants have been found associated. However, it is still mostly unclear through which downstream mechanism these variants cause these phenotypes. Knowledge of these intermediate steps is crucial to understand pathogenesis, while also providing leads for potential pharmacological intervention. Here we relied upon natural human genetic variation to identify effects of these variants on trans-gene expression (expression quantitative trait locus mapping, eQTL) in whole peripheral blood from 1,469 unrelated individuals. We looked at 1,167 published trait- or disease-associated SNPs and observed trans-eQTL effects on 113 different genes, of which we replicated 46 in monocytes of 1,490 different individuals and 18 in a smaller dataset that comprised subcutaneous adipose, visceral adipose, liver tissue, and muscle tissue. HLA single-nucleotide polymorphisms (SNPs) were 10-fold enriched for trans-eQTLs: 48% of the trans-acting SNPs map within the HLA, including ulcerative colitis susceptibility variants that affect plausible candidate genes AOAH and TRBV18 in trans. We identified 18 pairs of unlinked SNPs associated with the same phenotype and affecting expression of the same trans-gene (21 times more than expected, P<10−16). This was particularly pronounced for mean platelet volume (MPV): Two independent SNPs significantly affect the well-known blood coagulation genes GP9 and F13A1 but also C19orf33, SAMD14, VCL, and GNG11. Several of these SNPs have a substantially higher effect on the downstream trans-genes than on the eventual phenotypes, supporting the concept that the effects of these SNPs on expression seems to be much less multifactorial. Therefore, these trans-eQTLs could well represent some of the intermediate genes that connect genetic variants with their eventual complex phenotypic outcomes.

Many genetic variants have been found associated with diseases. However, for many of these genetic variants, it remains unclear how they exert their effect on the eventual phenotype. We investigated genetic variants that are known to be associated with diseases and complex phenotypes and assessed whether these variants were also associated with gene expression levels in a set of 1,469 unrelated whole blood samples. For several diseases, such as type 1 diabetes and ulcerative colitis, we observed that genetic variants affect the expression of genes, not implicated before. For complex traits, such as mean platelet volume and mean corpuscular volume, we observed that independent genetic variants on different chromosomes influence the expression of exactly the same genes. For mean platelet volume, these genes include well-known blood coagulation genes but also genes with still unknown functions. These results indicate that, by systematically correlating genetic variation with gene expression levels, it is possible to identify downstream genes, which provide important avenues for further research.

For many complex traits and diseases, numerous associated single nucleotide polymorphisms (SNPs) have been identified through genome-wide association studies (GWAS)through genome-wide association studies (GWAS) [1]. For many of these identified variants it is still unclear through which mechanism the association between the SNP and the trait or disease phenotype is mediated. A complicating factor is that disease-associated variants might not be the real causal variants, but are in linkage disequilibrium (LD) with the true disease-causing variant, making it difficult to accurately implicate the correct gene for a locus in disease pathogenesis. Within the major histocompatibility locus (MHC) on 6p, many SNPs have been found to be associated with complex diseases such as celiac disease, inflammatory bowel disease, psoriasis, rheumatoid arthritis, diabetes mellitus, schizophrenia, lung cancer and follicular lymphoma [2]–[10]. An analysis of the Catalog of Published Genome-Wide Association Studies [1] revealed that out of 1,167 unique SNP associations with a reported p<5×10−7, 82 (7.0%) were located within the MHC (Fisher's Exact p<10−30). Except for celiac disease [11] it remains largely unclear how MHC variants increase disease susceptibility. However, common variants have been identified that might exert their function by altering gene expression rather than by altering protein structure [2], [12]–[16] (expression quantitative trait loci, eQTLs). Comprehensive eQTL mapping (or genetical genomics [17]) will enable us to assess for every known disease-associated variant if it significantly affects gene expression. Genetic variants that affect expression of genes that map in their vicinity (cis-eQTLs) can potentially pinpoint the true disease gene from an associated locus. In addition, genetic variants may also affect expression of genes that reside further away or are on different chromosomes (trans-eQTLs) [18]. These trans-eQTLs are especially interesting, since they allow us to identify downstream affected disease genes which were not implicated by GWAS studies at all, and thereby potentially having the ability to reveal previously unknown (disease) pathways. In this study we performed a comprehensive eQTL mapping to explore the downstream effects of SNPs on gene expression by analyzing genotype and expression data of 1,469 unrelated samples. In addition to a genome-wide analysis, we also performed a focused analysis for disease- and trait-associated SNPs and SNPs located within the HLA. We replicated the identified trans-eQTLs in a collection of monocyte expression data and expression data from subcutaneous adipose, visceral adipose, muscle and liver tissue. Principal component analysis (PCA) enabled us to remove non-genetic expression variation [19], [20], resulting in increased power to detect eQTLs. A stringent probe-mapping strategy was used to filter out false-positive cis-eQTLs due to primer-polymorphisms and false-positive trans-eQTLs due to cross-hybridizations. Furthermore, a permutation strategy was utilized that corrects for multiple-testing, while preventing potential confounders such as non-even distribution of SNP markers and expression probe markers across the genome, differences in minor allele frequency (MAF) between SNPs, linkage disequilibrium (LD) within the genotype data, and correlation between expression probes. Results of a genome-wide eQTL analysis on 289,044 common SNPs, present on the Illumina HumanHap300 platform in peripheral blood expression data of 1,469 unrelated individuals, are provided in Table 1, Table S1, Table S2, Figure S1 (controlling false discovery rate (FDR) at 0.05 using a permutation strategy). As reported before [21]–[25] we also observed that eQTLs are strongly enriched for trait-associated SNPs (SNPs associated with a trait or disease phenotype, as reported in the Catalog of Published Genome-Wide Association Studies [1]): We therefore concentrated on these variants and imputed (Impute v2.0 [26]) additional genotype data permitting us to test 1,167 trait-associated SNPs. After removing false-positive eQTLs due to primer-polymorphisms and cross-hybridization 472 (40.4%) of these SNPs were cis-eQTLs, affecting the expression of 679 different transcripts, representing 538 genes (Figure 1, Table 1, Figure S2, Table S3). 67 (5.7%) SNPs were trans-acting on 130 different transcripts, representing 113 genes (Table S4). Results on the number of detected eQTLs per complex trait are provided in Table S5 and Figure S3. For nearly all significant trans-eQTLs the effect was present in each of the seven individual patient and controls cohorts, making up the total dataset (Table S6). These trans-eQTLs provide valuable insight on previously unknown functional downstream consequences trait-associated SNPs have, e.g. rs2395185 is the strongest susceptibility variant for ulcerative colitis [27] (UC) but also the strongest SNP, trans-acting on Acyloxyacyl hydrolase (AOAH, p = 1.0×10−36), an enzyme that modulates host inflammatory responses to gram-negative bacterial invasion. It is known that deficiencies in response mechanisms against bacterial products like lipopolysaccharide, present on gram-negative bacterial cell walls, play an important role in UC disease pathogenesis [28]. Within the peripheral blood we observed that AOAH is significantly co-expressed with colony stimulating factor 1 receptor (CSF1R, r = 0.21) and major histocompatibility complex class II DR alpha (HLA-DRA, r = 0.19). Hyperstimulation of CSF1R has been implicated in UC [29], while HLA-DRA is one of the positional UC candidate genes mapping in very close proximity to rs2395185. Another UC HLA variant, rs9268877, was trans-acting on T cell receptor beta variable 18 (TRBV18), part of the TCRß locus at 7q34. It is known that TCRß mutant mice develop chronic colitis [30]. For type 1 diabetes (T1D) we observed that 59% (30/51) of the known and tested T1D associated SNPs are cis-acting (on in total 53 unique genes) and 17% (9/50) are trans-acting on 22 unique genes (Figure 2). Potentially interesting trans-genes include CCL2, CFB, CLN1, KRT19, OSR1 and RARRES1, all strongly co-expressed with each other. CCL2 and CFB are known immune response genes and have been implicated in T1D before [31]–[33]. For breast cancer we observed that rs3803662 [34] is trans-acting on origin recognition complex subunit 6 (ORC6L). This gene is involved in DNA replication and has been frequently used as part of prognostic profiles for predicting the clinical outcome in breast cancer [35], [36]. We observed a marked enrichment for SNPs within the MHC among the cis- and trans-acting trait-associated SNPs: 65 of 472 cis-acting SNPs (13.8%, EVD p<1.0×10−16) and 32 of 67 trans-acting SNPs (47.8%, EVD p<1.0×10−16) mapped within the MHC (Figure 3). These SNPs all map to the Human Leukocyte Antigens (HLA) locus: SNPs within the HLA class I region, class II region and class III region affect 20, 7 and 2 different genes in trans, respectively. While multiple associated SNPs have been identified for many complex diseases, it often remains unclear what the intermediate effects of these variants are that eventually lead to disease. It is reasonable to assume that for a particular phenotype the different associated SNPs eventually converge on the same downstream gene(s) or pathways. We identified 7 unique pairs of unlinked SNPs that are associated with the same phenotype and that also affect the same downstream genes in trans or cis (at FDR 0.05, Table 2, Figure 4a). In order to establish whether this was more than expected by chance, we repeated this analysis, while using a set of trans-eQTLs, equal in size to the set of real trans-eQTLs, most significant after having permuted the expression sample identifiers. We performed this procedure 100 times, and observed on average only 0.15 unique pairs of unlinked SNPs (range [0, 3], Figure 4b) that showed this convergence, which indicates that the observed number of converging pairs of SNPs is 47 times more than expected (EVD p<1.0×10−16) and implies a false-positive rate of 0.021. Due to this highly significant enrichment of converging pairs of SNPs and its low estimated false-positive rate, we also ran an analysis where we had relaxed the FDR for trans-eQTLs to 0.50 (Table S7). Here we observed 18 pairs of SNPs that converge on the same genes, whereas in the 100 subsequent permutations we observed this only on average for 0.84 SNP-pairs (range [0, 5], 21 times more expected by chance, EVD p<1.0×10−16, implying a false-positive rate of 0.047, Table 2, Figure 4b). Many of these converging downstream genes make biological sense: three independent loci, associated with hemoglobin protein levels [37]–[39] and ß thalassemia susceptibility [40], significantly affect hemoglobin gamma G (HBG2) gene expression levels (each with p<1.0×10−23, Figure 5). For mean corpuscular volume (MCV, Figure 5) two unlinked MCV SNPs [41], [42] also affect HBG2 gene expression levels in trans (at FDR 0.05), while other pairs of MCV SNPs converge on ESPN, VWCE, PDZK1IP1 and RAP1GAP. For mean platelet volume (MPV) we observed that MPV SNPs rs12485738 on 3p26 and rs11602954 on 11p15 affect several transcripts in trans. These two SNPs converge on GP9, F13A1, C19orf33, SAMD14, VCL and GNG11. As GP9 and F13A1 are known blood coagulation genes, C19orf33 is a potential candidate gene, involved in coagulation as well. This is substantiated by strong co-expression between GP9 and C19orf33 within peripheral blood (Pearson r = 0.45, p = 7.0×10−63) and the fact these SNPs independently also affect various other blood coagulation genes in trans (including CD151, GP1BB, ITGA2B, MMRN1, THBS1 and VWF, Figure 4). Many of these are specific to megakaryocytes that are platelet precursor cells [43]. As expected, the Gene Ontology term ‘blood coagulation’ is strongly overrepresented among all these trans-genes, Fisher's exact p = 1.0×10−10. We observed that MPV SNP rs12485738 (on 3p14.3) was also trans-acting on tropomyosin 1 (TPM1, 15q22.2, p = 9.7×10−9), a gene that is also regulated in cis by another MPV variant (rs11071720 on 15q22.2, p = 1.4×10−13). We observed this for two different expression probes that map within different locations of the TPM1 transcript (probes 5560246 and 610519), and note strong co-expression for these two TPM1 probes with 46 MPV trans-genes (Pearson r>0.19, p<1.0×10−11, including five known coagulation genes). Although several genes reside within the rs11071720 MPV locus, these observations strongly implicate TPM1 as the causal MPV gene. For both MPV and MCV we observed that the identified cis- and trans-eQTL probes generally were more strongly co-expressed in peripheral blood than expected (Figure S4, MPV co-expression Wilcoxon P<10−200, MCV co-expression Wilcoxon P = 0.009), substantiating the likelihood these genes reflect coherent biological sets. We repeated this co-expression analysis after we had regressed out all cis- and trans-eQTL effects, and observed that most of this co-expression was independent of the eQTL SNP-effect on the expression of these genes, which further substantiates that these genes are biologically related (MPV co-expression Wilcoxon P<1−200, MCV co-expression Wilcoxon P = 0.018). Although the observed convergence provides insight into downstream genes, it is not clear whether the MPV or MCV phenotypes are eventually caused through these trans-genes, or whether these trans-eQTLs emerged as a result of changes to the volume of the platelets or the erythrocytes. In order to gain insight into this, we analyzed the effect size of these SNP variants on both the expression levels and the phenotypes. While the effect sizes of these trait-associated SNPs on eventual phenotypes were usually small, their intermediate (molecular) effects was often substantially larger. This supports the notion that the effect on e.g. MPV and MCV is through these trans-genes, and suggests the presence of ‘phenotypic buffering’, shown previously in plants [44], in humans (Table 2, Figure 4b): the effects of the 18 converging pairs of SNPs on gene expression levels were often substantially higher than the originally reported effect sizes on the trait-phenotypes. For example, several MPV- and MCV-associated SNPs explain between 1.41% and 10.99% of trans-expression variation within the 1,469 unrelated samples, whereas these SNPs only explain between 0.24% and 1.12% of the MPV and MCV phenotype variation (and as such required over 13,000 samples [41], [42] for identification, Figure 4b). We analyzed peripheral blood which is a mixture of different hematopoetic cell types. In addition, we also assessed whether the identified trait-associated trans-eQTLs (detected at FDR 0.05) could be replicated in a single cell-type dataset. This is an important question, as it is potentially possible that the trans-acting SNP are able to alter the amount, volume or ratio of certain blood cell types, which might as a consequence result in an indirect net effect on the measured gene expression levels within the mix of the cells that comprise whole blood. We therefore analyzed monocyte expression data from 1,490 independent samples [45] and did not find evidence that this was a widespread phenomenon as we could replicate 46 out of the 130 different trans-eQTLs (each of these with a nominal p<1.0×10−5 in the monocyte data, Table S8). These replicated eQTLs include the genes AOAH, HBG2, GP9, F13A1, SAMD14, CD151, ITGA2B, MMRN1, THBS1, VWF and TPM1 mentioned above. Surprisingly we could also replicate the trans-eQTL effects on various blood-coagulation genes for mean platelet volume SNP rs12485738: One might argue that rs12485738 primarily increases platelet volume, resulting in a relatively higher volume of platelet-RNA when assessing total peripheral blood RNA. If this were to be the case, a measurable trans-effect is expected for platelet-specific (blood coagulation) genes in whole blood. Such an effect would then not actually be an expression-QTL, but rather a ‘cellular-QTL’. However, the trans-eQTLs for rs12485738 were also present in single cell-type monocyte datasets, indicating that the above concerns do not apply. Clearly, trans-eQTL effects can manifest themselves outside the primary cell-type, in which they are expected to operate. We also replicated 18 trait-associated trans-eQTLs (including AOAH, detected at FDR 0.05) in an independent dataset comprising four different non-blood tissues (subcutaneous adipose, visceral adipose, liver and muscle, Figure S5, Table S9 and S10). Since this dataset comprised only 90 samples, it is very encouraging that 18 trans-eQTL could be replicated. Here we investigated gene expression in peripheral blood from 1,469 individuals to identify cis- and trans-effects of common variants on gene expression levels. When comparing to other genetical genomics studies [12]–[14], [16], [18], [21]–[24], [45], [46] we observe an increasing percentage of genes that are cis- or trans-regulated (39% of 19,689 unique genes at FDR 0.05). When eQTL studies further increase the sample-sizes and thus increase statistical power, we expect that for the far majority of genes the expression levels are to some extent determined by genetic variation. GWA studies have identified many loci, but it is still often unclear what the affected gene in each locus is. Here we showed that 39% of trait-associated SNPs affect gene expression in cis which is helpful in pinpointing the most likely gene per susceptibility locus. However, GWAS do not immediately provide insight in the trans-effects of these susceptibility variants on downstream genes. Here we identified for 2.6% of all trait-associated SNPs trans-eQTL effects on in total 113 unique genes. While some of these trans-eQTLs are known to be involved in these phenotypes (such as HBG2 in hemoglobin protein levels and ß-Thallasemia), most of these genes have not been implicated before in these complex traits, and provide additional insight in the downstream mechanisms of these variants. Interestingly, 48% of trans-acting trait-associated SNPs map within the HLA, indicating the HLA has a prominent role in regulating peripheral blood gene expression. This might partly explain why the HLA has been found to be associated with so many different diseases. While we concentrated on peripheral blood, we could replicate 35% of the trans-eQTLs in monocytes. Particularly surprising was the observation that for SNPs, known to affect the volume of platelets or erythrocytes the identified trans-eQTL effects in whole blood were also present in these monocytes. Among these replicated genes are a considerable number of highly plausible trans-genes. For example, for mean platelet volume SNP rs12485738 we detected the same trans-eQTL effects on seven well-known blood coagulation genes (F13A1, GP1BB, GP9, ITGA2B, MMRN1, THBS1 and VWF) in both the peripheral blood data and the monocyte data. Interestingly, in both datasets, trans-effects for this SNP on another 31 genes were identified as well, which suggests these genes play a role in blood coagulation. It can thus be concluded that trans-eQTLs, identified in peripheral blood, generally apply to monocytes as well. We assumed these eQTLs might therefore also be present in other, non-blood tissues, as previously observed for rodents [47]–[49]. Indeed we could replicate some of these trans-eQTLs in a smaller dataset of four non-blood tissues. Importantly, as mentioned before [46], the allelic directions were nearly always identical to blood, which implies that trans-eQTLs, if also present in another tissue, work in the same way. Our observation that sets of independent SNPs, associated with the same complex phenotype sometimes also affect exactly the same trans-gene, further substantiates the validity of our findings. Based on the reported effect-sizes of these variants on these complex phenotypes, we have shown here that the individual effects of these SNPs on trans-gene expression can often be stronger. This suggests that these down-stream gene expression effects do not fully propagate to the eventual phenotype and are somehow buffered. This ‘phenotypic buffering’ has been observed before in plants [44] and suggests that additional compensatory mechanisms exist that control these complex phenotypes. However, we do realize that accurate estimates on this phenomenon requires the availability of both gene-expression and phenotype data for these traits. As we did not have these phenotypes for our samples, we relied upon estimates from literature. Future studies that have collected both genome-wide genotype, expression and phenotype data from the same individuals will permit answering the question what the extent of this phenotypic buffering is. We should emphasize that the number of converging pairs of SNPs that we identified must be a very strong underestimate, and as such the false-negative rate from this analysis is likely to be high: As we observed that on average 40.4% of the trait-associated SNPs affect gene expression levels in cis, we expect that many of these SNPs will exert effects on gene expression in trans. However, these effects are likely to be small and due to multiple testing issues our current study identified only a relatively small set of trans-eQTL effects. Likewise the number of detected converging pairs of SNPs is even smaller. However, as we observed this convergence for various pairs of SNPs, future genetical genomics studies using larger sample sizes will likely reveal many more pairs of converging SNPs, providing better insight in the downstream molecular mechanisms that are affected by these disorders. The convergence and phenotypic buffering we observed might also help uncover some of the missing heritability in complex disease. As there are probably many SNPs with low marginal phenotypic effects [50], GWAS currently lack power to detect these. However, the effect of these trait-associated SNPs on expression seems to be less multifactorial, leading to larger expression effects. These numerous expression disturbances will eventually converge to a phenotype, explaining the small phenotypic effect of individual trait-associated SNPs. Therefore, studying expression as intermediate phenotype will be important for disease association studies trying to account for the missing heritability of complex diseases. Disease SNPs, already found to be disease-associated and marked as eQTL, lead to a set of candidate downstream genes. Additional genetic variants that also affect the expression of these genes will therefore be powerful candidates for disease susceptibility. The peripheral blood genetical genomics study population contained 1,469 unrelated individuals from the United Kingdom and the Netherlands. Some of these are healthy controls while others are patient samples. The 49 ulcerative colitis (UC) cases in this study are part of the inflammatory bowel disease (IBD) cohort of the University Medical Centre Groningen. The 111 celiac disease samples were collected within the Barts and the London NHS Trust and the Oxford Radcliffe Hospitals NHS Trust. The 453 chronic obstructive pulmonary disease (COPD) samples were collected within the NELSON study. The 856 amyotrophic lateral sclerosis (ALS) cases and controls were collected in the University Medical Centre Utrecht. All samples were collected after informed consent and approved by local ethical review boards. Individual sample information is provided in Table S11. Peripheral blood (2.5 ml) for all samples was collected with the PAXgene system (PreAnalytix GmbH, UK). PAXgene vials were chosen to prevent density gradient centrifugation, immortalization or in vitro cell culture artifacts changing mRNA profiles. PAXgene tubes were mixed gently and incubated at room temperature for two hours. After collection, tubes were frozen at −20°C for at least 24 hours followed by storage at −80°C. RNA was isolated using the PAXgene Blood RNA isolation kit (PreAnalytix GmbH, UK). RNA was quantified using the Nanodrop (Nanodrop Technologies, USA). Total RNA integrity was analyzed using an Agilent Bioanalyzer (Agilent Technologies, USA). Peripheral blood samples were either genotyped using the Illumina (Illumina, San Diego, USA) HumanHap300, HumanHap370 or 610 Quad platform. Genotyping was performed according to standard protocols from Illumina. Although the different genotype oligonucleotide arrays differ, they share 294,757 SNPs, to which the analysis was confined. In addition, SNPs with a minor allele frequency of <5%, or a call-rate <95%, or deviating from Hardy-Weinberg equilibrium (exact p-value <0.001) were excluded, resulting in 289,044 SNPs for further analysis. Genotype calling for each SNP was performed by a previously described algorithm [51]. Anti-sense RNA was synthesized, amplified and purified using the Ambion Illumina TotalPrep Amplification Kit (Ambion, USA) following the manufacturers' protocol. Complementary RNA was either hybridized to Illumina HumanRef-8 v2 arrays (229 samples, further referred to as H8v2) or Illumina HumanHT-12 arrays (1,240 samples, further referred to as HT12) and scanned on the Illumina BeadArray Reader. Raw probe intensities were extracted using Illumina's BeadStudio Gene Expression module v3.2 (No background correction was applied, nor did we remove probes with low expression). The raw expression data of the 1,240 HT12 peripheral blood samples were combined with the raw expression data of 296 replication samples (described in detail in paragraph ‘Trans-eQTL replication dataset’). Both datasets (H8v2 and HT12) were quantile normalized separately to the median distribution and expression values were subsequently log2 transformed. Subsequently, the probes were centered to zero and linearly scaled such that each probe had a standard deviation of one. The HT12 and H8v2 arrays share a considerable number of probes with identical probe sequences. However, in a considerable number of occasions the two platforms use different probe identifiers for the same probe sequences. More importantly, although probe identifiers are often identical, they sometimes represent different probe sequences. In order to permit a meta-analysis incorporating data from both arrays, we decided on the following naming convention: if an H8v2 probe had the same sequence as an HT12 probe, the HT12 ‘ArrayAddressID’ probe identifier was used. If not, the original H8v2 probe identifier was used, but with the prefix “Human_RefSeq-8_v2-” to prevent any potential probe identifier ambiguity. A total of 52,061 unique probes were used for further analysis, representing 19,609 unique genes according to HUGO gene nomenclature. Various mapping strategies were used for the expression probes to get a mapping location that was as unambiguous as possible: if probes have been mapped incorrectly, or cross-hybridize to multiple genomic loci, it might be that an eQTL will be incorrectly deemed a trans-eQTL, while in fact it is a cis-eQTL or primer polymorphisms. We used Ensembl database version 52 (NCBI 36.3 assembly) to obtain, for each annotated gene, the transcript with the largest number of exons and included this main spliced transcript in our reference set. Second, we added one sequence per intron, extending intron boundaries 40 bp on each side to allow mapping of the 50 bp probe sequences that overlapping exon-intron junctions. Last, a version of the reference DNA genome with masked annotated transcripts was included. Probe sequences were mapped using NOVOALIGN V2.05.12 for all the sequences (main transcript, introns, and non standard exon-exon junctions) originating from the same transcript (parameters −t 150 −v 20 20 200 [>]( [ ̂_]*)_). For each probe it was determined whether it was mapping uniquely to one particular genomic locus, or, if multiple hits were present whether all these mappings resided in each other vicinity (<250 kb). Probes that did not map at all, or mapped to multiple different loci were excluded from further analyses. Using this approach, 43,202 of the 48,751 probes on the HT12 and 21,316 of the 22,185 probes on the H8v2 platform were eventually mapped to a single genomic location. In order to detect cis-eQTLs, analysis was confined to those probe-SNP combinations for which the distance from the probe transcript midpoint to SNP genomic location was ≤250 kb. For trans-eQTLs, analysis was confined to those probe-SNP combinations for which the distance from probe transcript midpoint to SNP genomic location was ≥5 Mb (to exclude the possibility of accidentally detecting cis-eQTLs due to long ranging linkage disequilibrium). Additionally, for the trans-eQTL analysis the effects of the significant cis-eQTLs were removed from the expression data by keeping the residual expression after linear regression. Association for cis- and trans-eQTL was tested with a non-parametric Spearman's rank correlation. For directly genotyped SNPs we coded genotypes as 0, 1 or 2, while for imputed SNPs we used SNP dosage values, ranging between 0 and 2. When a particular probe-SNP pair was present in both the HT12 and H8v2 datasets, an overall, joint p-value was calculated using a weighted (square root of the dataset sample number) Z-method. To correct for multiple testing, we controlled the false-discovery rate (FDR) at 0.05: the distribution of observed p-values was used to calculate the FDR, by comparison with the distribution obtained from permuting expression phenotypes relative to genotypes 100 times within the HT12 and H8v2 dataset for both the cis- and trans- analyses [52]. In order to increase the number of detectable cis- and trans-eQTLs we applied a principal component analysis (PCA) on the sample correlation matrix. We, among others [19], [20], argue that the dominant PCs, capturing the larger part of the total variation, will primarily capture sample differences in expression that reflect physiological or environmental variation as well as systematic experimental variation (e.g. batch and technical effects). Figure S6 shows for the 1,240 HT12 samples what per individual the PC scores are. It is evident there are, especially among the first PCs, strong batch effects are still present after proper quantile-quantile normalization. By removing the variation captured by these PCs, we expected that the residual expression is more strongly determined by genetic variants and the number of significantly detected cis- and trans-eQTLs will increase. An aspect to consider is that with the removal of more PCs from the data, the degrees of freedom of the data will decrease. Furthermore, it is not immediately clear which PCs will actually capture physiological, environmental, and systematic variation, which might lead to removal of genetically determined expression variation as well. Therefore a tradeoff has to be made on the number of PCs to subtract from the data. We assessed this systematically, by removing up to 100 PCs from the genetical genomics dataset (in steps of 5). Figure S7A shows that the number of significantly detected cis-eQTL probes increases two-fold when 50 PCs were removed from the expression data. There is a long plateau visible (around PC50), where the number of detected cis-eQTLs probes remains approximately constant, irrespective of removing for instance 10 fewer or 10 extra PCs (reported numbers in this figure also include false-positive eQTLs due to potential primer polymorphisms, as we here wanted to solely compare the performance of removing different numbers of PCs). Figure S7B shows that of the initial 5,950 significantly detected cis-eQTL probes (no PCs removed), 4,965 (83.5%) were still detected with 50 PCs subtracted. The 985 initially detected cis-eQTLs probes, yet no longer detected when 50 PCs had been removed from the expression data, all had a low significance (Figure S8). As we controlled the FDR at 0.05 in all analyses it is therefore likely that a considerable amount of these reflect false-positives. Figure S8C shows that for all the overlapping 4,965 detected cis-eQTLs probes between the different analyses, the allelic direction was identical, and effect size on expression correlate well (Pearson r = 0.95) although these were nearly always stronger after having subtracted 50 PCs. We assessed this for trans-eQTLs as well. An important aspect to consider is that trans-eQTL SNPs might affect multiple genes. If these effects are substantial (either in effect size or the number of affected genes), it is likely that a certain PC will capture this. Removal of such PCs from the expression data will therefore unintentionally result in the inability to detect these trans-eQTLs. In order to avoid such false-negatives we first performed a QTL analysis on the first 50 PCs (that had been removed from the expression data for the cis-eQTL analysis) to assess whether some of these PCs are under genetic control (genome-wide analysis, controlling FDR at 0.05). We did this for the large HT12 and the smaller H8v2 expression data separately, as PCA had been applied independently to these datasets. We observed that out of the first 25 PCs in the HT12 data three PCs and in the H8v2 two PCs were to some extent genetically determined (r2>5%). This was different for PCAs 26–50 in the HT12 data: 11 PCs were under substantial genetic control (Figure S9a). We therefore assumed that most trans-eQTLs could be detected when removing approximately 25 PCs. We quantified this systematically, by removing increasing amounts of PCs from the expression data and conducting a full genome-wide trans-eQTL mapping. Indeed, in these analyses at most 244 significant trans-eQTLs could be detected (at FDR 0.05, with potential false-positives due to cross-hybridizations removed), when removing 25 PCs (Figure S9b). The overlap with the expression with no PCs removed was substantial: 62 of the 82 trans-eQTLs (77%), detected in the original analysis were detected as well in the analysis with 25 PCs removed (Figure S9c), all with identical allelic directions (Figure S9d). One should be aware that sequence polymorphisms can cause many false cis-eQTLs [53]. Such false cis-eQTLs do not reflect actual expression differences caused by sequence polymorphisms in cis-acting factors that affect mRNA levels. Instead they indicate hybridization differences caused by sequence polymorphisms in the mRNA region that is targeted by the microarray expression probes. Therefore, SNP-probe combinations were excluded from the cis-eQTL analysis when the 50 bp long expression probe mapped to a genomic location that contained a known SNP that was showing at least some LD (r2>0.1) with the cis-SNP. We used SNP data from the 1000 Genomes Projects, as it contains LD information for 9,633,115 SNPs (April 2009 release, based on 57 CEU samples of European descent). Detected trans-eQTLs might also reflect false-positives, although we initially had attempted to map the expression probes as accurately as possible, by using the aforementioned three different mapping strategies: it is still well possible that some of the identified, putative trans-eQTLs in fact reflect very subtle cross-hybridization (e.g. pertaining to only a small subsequence of the probe). We therefore tried to falsify each of the putative trans-eQTLs by attempting to map each trans-probe into the vicinity of the SNP probe location, by using a highly relaxed mapping approach. All putative Illumina trans-expression probes were mapped using SHRiMP [54], which uses a global alignment approach, to the human reference genome (NCBI 36.3 build). The mapping settings were chosen very loosely to permit the identification of nearly all potential hybridization locations: match score was 10, the mismatch score was 0, the gap open penalty was −250, the gap extension penalty was −100, Smith and Waterman minimum identical alignment threshold was 30.0%, while other SHRiMP parameters were left at default. Using these settings all mappings with a minimum overlap of 15 bases, or with 20 matches with one mismatch, or 30 matches with 2 mismatches, or full-length (50 bp) probe hybridizations with no more than 15 mismatches were accepted. Any trans-eQTL was discarded, if the expression probe had a mapping that was within 2 Mb of the SNP that showed the trans-eQTL effect. Once these potential false-positive trans-eQTLs had been removed from the real, non-permuted data, we repeated the multiple testing correction (again controlling the FDR at 0.05). Using this strategy we observed several instances where only 20 out the 50 bases of a probe sequence mapped in the vicinity of the trans-SNP (data not shown). For these trans-eQTLs the Spearman's rank correlation p was often lower than 10−100, which would imply these SNPs explain over 25% of the total expression variation of the corresponding trans-genes. Given the small amount of trans-eQTLs we detected in total, such effect sizes are quite unlikely and therefore provide circumstantial evidence these indeed reflect cross-hybridization artifacts. We also assessed whether any of the Illumina SNPs that constitute trans-eQTLs might map to a different position than what is reported in dbSNP. As such we mapped the 50 bp Illumina SNP probe sequences to the genome assembly, permitting up to four mismatches per 50 bp SNP probe sequence. We did not observe any SNP that could map (with some mismatches) to the same chromosome of the trans-probe. It is still possible that some of the trans-eQTLs for which we did not find any evidence of cross-hybridization, still are false positives, e.g. by missing some cross-hybridizations due to imperfections in the NCBI v36 assembly we used. Although we have identified numerous occasions where a SNP affects two different probes within the same gene in trans, substantiating the likelihood these trans-eQTLs are real, providing unequivocal evidence that all our reported trans-eQTLs are real is not straightforward. To assess enrichment of trait-associated SNPs, we used a collection of 1,262 unique SNPs from 'A Catalog of Published Genome-Wide Association Studies' (accessed 09 February 2010, and each having at least one reported association p-value <5.0×10−7). We could successfully impute the genotypes for 1,167 of these SNPs and therefore confined all analyses to these SNPs. Of these SNPs 572 had been directly genotyped on the Illumina HumanHap300 platform, with a MAF>0.05, an HWE exact p-value >0.0001 and call-rate >95%. To ascertain whether these SNPs are more often constituting an eQTL than expected, we used a methodology that is not affected by the following potential confounders: non-even distribution of SNP markers and expression probe markers across the genome, differences in MAF between SNPs and LD structure within the genotype date and correlation between probes in the expression data. Additionally, this methodology is also not confounded by the fact that for certain traits different SNPs in strong LD can have been reported, due to differences in the platforms that were used to identify these loci. We first determined how many unique eQTL SNPs had been identified in the original eQTL mapping (with an FDR<0.05) and how many of these are trait-associated. Subsequently we permuted the expression phenotypes relative to the genotypes (thus keeping the correlation structure within the genotype data and the correlation structure within the expression data intact, yet assigning the genotypes of a sample to the expression data of a randomly chosen sample) and reran the eQTL mapping, sorting all tested eQTLs on highest significance. We then took an equal number of top associated, but permuted, eQTL SNPs and determined how many of these permuted eQTL SNPs are trait-associated. By performing 100 permutations we obtained an empiric distribution of the number of trait-associated SNPs expected by chance. We subsequently fitted a generalized extreme value distribution (EVD, using the EVD add-on package for R), permitting us to estimate realistic enrichment significance estimates (called EVD p throughout the manuscript). For the MHC enrichment analysis the followed procedure was identical, with the difference that we looked for enrichment for SNPs within the MHC, defined as SNPs physically mapping between 20 Mb and 40 Mb on chromosome 6 (NCBI 36 assembly). Replication of the detected eQTLs was performed in monocytes from 1,490 different samples [45] and in an independent population of 86 morbidly obese individuals that underwent elective bariatric surgery (Department of general surgery, Maastricht University Medical Centre, the Netherlands). Both these datasets also used the same Illumina HumanHT-12 expression platform. For the 1,490 monocyte samples eQTL P-Values summary statistics were available for all monocyte trans-eQTLs with a nominal p<1.0×10−5. We ascertained how many of the trans-eQTLs we had found in our peripheral blood data had a nominal eQTL p<1.0×10−5 in this monocyte dataset. We also assessed trans-eQTLs in four different tissues from the 86 morbidly obese individuals that underwent bariatric surgery. DNA was extracted from blood samples using the Chemagic Magnetic Separation Module 1 (Chemagen) integrated with a Multiprobe II Pipeting robot (PerkinElmer). All samples were genotyped using both Illumina HumanCytoSNP-12 BeadChips and Illumina HumanOmni1-Quad BeadChips (QC was identical as was applied to the peripheral blood samples). We imputed HapMap 2 genotypes using Impute version 2.0. In addition expression profiling was performed for four different tissues for each of these individuals using the Illumina HumanHT-12 arrays. Wedge biopsies of liver, visceral adipose tissue (VAT, omentum majus), subcutaneous adipose tissue (SAT, abdominal), and muscle (musculus rectus abdominis) were taken during surgery. RNA was isolated using the Qiagen Lipid Tissue Mini Kit (Qiagen, UK, 74804). Assessment of RNA quality and concentration was done with an Agilent Bioanalyzer (Agilent Technologies USA). Starting with 200 ng of RNA, the Ambion Illumina TotalPrep Amplification Kit was used for anti-sense RNA synthesis, amplification, and purification according to the protocol provided by the manufacturer (Ambion, USA). 750 ng of complementary RNA was hybridized to Illumina HumanHT12 BeadChips and scanned on the Illumina BeadArray Reader. Expression data preprocessing was as mentioned before. We first attempted to replicate the trait-associated trans-eQTLs per tissue, using an FDR of 0.05 and 100 permutations. Subsequently we conducted a meta-analysis, combining the four tissues. Per trans-eQTL we used a weighted Z-method to combine the four individual p-values. However, these four datasets are not independent, as they reflect the same individuals. We resolved this by conducting the permutations in such a way that in every permutation round the samples were permuted in exactly the same way for each of the four tissues. By doing this we retained the correlations that exist between the different tissues per sample, and were able to get a realistic empiric (null-)distribution of expected test-statistics. Per trait we assessed all the SNPs that have been reported to be associated with that particular trait. We analyzed per trait all possible SNP-pairs. If a pair of SNPs was not in LD (r2<0.001) we assessed whether they affected the same gene in cis or trans. When using the trait-associated cis- and trans-eQTLs that had been identified when controlling the FDR at 0.05, we identified 7 unique pairs of SNPs that caused both the same phenotype and also affected the same gene(s). When using a somewhat more relaxed set of trans-eQTLs, identified when controlling the FDR at 0.5, we identified 18 unique pairs of SNPs that affect the same downstream gene. We assessed whether these numbers were significantly higher than expected, by using the same strategy that we had used to assess the enrichment of trait-associated SNPs and the HLA; we ran 100 permutations. We kept per permutation the cis-eQTL list as it was, but generated a permuted set of trans-eQTLs, equal in size to the original set of non-permuted trans-eQTLs. This enabled us to determine per permutation round how many unique pairs of SNPs converge on the same gene(s). We subsequently fitted a generalized extreme value distribution, permitting us to estimate realistic enrichment significance estimates. If a particular SNP is cis- or trans-acting on multiple genes, it is plausible that those genes are biologically related. Co-expression between these genes provides circumstantial evidence this is the case, strengthening the likelihood such cis- and trans-eQTLs are real. We assessed this in the peripheral blood data, by using the expression data of the 1,240 samples, run on the comprehensive HT12 expression platform. As we had removed 25 PCs (to remove physiological, environmental variation, and systematic experimental variation) for the trans-eQTL analyses, we decided to confine co-expression analyses to this expression dataset. As there are 43,202 HT12 probes that we mapped to a known genomic location, 43,202×43,201/2 = 933,184,801 probe-pairs exist. Given 1,240 samples, a Pearson correlation coefficient r≥0.19 corresponds to a p<0.05 when applying stringent Bonferroni correction for these number of probe-pairs. Expression data for both the peripheral blood and the four non-blood datasets have been deposited in GEO with accession numbers GSE20142 (1,240 peripheral blood samples, hybridized to HT12 arrays), GSE20332 (229 peripheral blood samples, hybridized to H8v2 arrays) and GSE22070 (subcutaneous adipose, visceral adipose, muscle and liver samples).

10.1371/journal.pcbi.1003442

De Novo Prediction of PTBP1 Binding and Splicing Targets Reveals Unexpected Features of Its RNA Recognition and Function

The splicing regulator Polypyrimidine Tract Binding Protein (PTBP1) has four RNA binding domains that each binds a short pyrimidine element, allowing recognition of diverse pyrimidine-rich sequences. This variation makes it difficult to evaluate PTBP1 binding to particular sites based on sequence alone and thus to identify target RNAs. Conversely, transcriptome-wide binding assays such as CLIP identify many in vivo targets, but do not provide a quantitative assessment of binding and are informative only for the cells where the analysis is performed. A general method of predicting PTBP1 binding and possible targets in any cell type is needed. We developed computational models that predict the binding and splicing targets of PTBP1. A Hidden Markov Model (HMM), trained on CLIP-seq data, was used to score probable PTBP1 binding sites. Scores from this model are highly correlated (ρ = −0.9) with experimentally determined dissociation constants. Notably, we find that the protein is not strictly pyrimidine specific, as interspersed Guanosine residues are well tolerated within PTBP1 binding sites. This model identifies many previously unrecognized PTBP1 binding sites, and can score PTBP1 binding across the transcriptome in the absence of CLIP data. Using this model to examine the placement of PTBP1 binding sites in controlling splicing, we trained a multinomial logistic model on sets of PTBP1 regulated and unregulated exons. Applying this model to rank exons across the mouse transcriptome identifies known PTBP1 targets and many new exons that were confirmed as PTBP1-repressed by RT-PCR and RNA-seq after PTBP1 depletion. We find that PTBP1 dependent exons are diverse in structure and do not all fit previous descriptions of the placement of PTBP1 binding sites. Our study uncovers new features of RNA recognition and splicing regulation by PTBP1. This approach can be applied to other multi-RRM domain proteins to assess binding site degeneracy and multifactorial splicing regulation.

A key step in the regulation of mammalian genes is the splicing of the messenger RNA precursor to produce a mature mRNA that can be translated into a particular protein needed by the cell. Through the process of alternative splicing, mRNAs encoding different proteins can be derived from the same primary gene transcript. The regulation of this process plays essential roles in the development of differentiated tissues and is mediated by special pre-mRNA binding proteins. To understand how these proteins control gene expression, one must characterize what they recognize in RNA and identify these binding sites across the genome in order to predict their targets. Models that allow this prediction are essential to understanding developmental regulatory programs and their perturbation by disease causing mutations. In this study, we use statistical methods to build models of RNA recognition by the important splicing regulator PTBP1 and then apply these models to predict PTBP1 regulation of new gene transcripts. We show that PTBP1 has different specificity for RNA than was previously recognized and that its target exons are more diverse than was known before. There are many similar splicing regulators in mammalian cells, and these analyses provide a general framework for the computational analysis of their RNA binding and target identification.

Alternative splicing of pre-mRNA commonly determines the protein output of mammalian genes, with most genes generating multiple mRNA and protein products [1]. A typical alternative exon is affected by multiple pre-mRNA binding proteins that may either enhance or repress splicing [2]. The expression and activity of these splicing regulatory proteins can vary with development, cell type, or cellular stimulus [3]. This complex combinatorial regulation can be seen in the conserved sequences within and surrounding alternative exons, which generally contain the binding sites for many different regulators. These sequences make up what is sometimes called the splicing code as they determine where and when the exon is spliced into an mRNA [4], [5], [6], [7]. Such a code should allow the development of models that predict exon regulation based solely on the RNA binding affinity of the many regulatory proteins and their other interactions. However, this is not currently feasible, in part due to our incomplete understanding of RNA recognition by the splicing regulators and their mechanisms of action. Whole-transcriptome crosslinking methods for individual proteins in vivo are allowing the identification of large numbers of protein/RNA interaction sites [8], [9], [10], [11]. These data can be overlapped with functional data on splicing to identify possible direct target exons for particular proteins [12], [13], [14], [15]. However, there are limitations in the interpretation of these data. Crosslinking efficiency can vary between different proteins and between individual binding sites, making it difficult to relate the crosslinking signal to the actual binding affinity. These signals are also dependent on the expression of the bound RNA, and since these data are generated one tissue or cell type at a time it is not always feasible to extend the results from one setting to a new cell type or point in development. It would be extremely useful to be able to scan for binding affinity across the complete transcriptome and to predict exon targets in tissues that have not yet been subjected to experimental analysis. Splicing regulatory proteins commonly contain multiple RRM or other RNA binding domains, with each domain recognizing a short element of a few nucleotides [2], [16]. Subtle variation in the optimal binding element of each domain and flexible peptide linkers between them allow for significant degeneracy within high affinity binding sites. Although the short sequence motifs that are common to a set of binding sites are readily identified, these likely constitute only a portion of a full high affinity site. To rank binding sites and assess their finer structures, we need an approach to search for clusters of these short motifs and to score for binding affinity. The Polypyrimidine tract binding protein 1 (PTBP1) is a widely studied splicing regulatory protein [17], [18]. PTBP1 is known to repress the splicing of a large number of exons by binding in their adjacent introns or within the exons themselves. PTBP1 is down regulated in differentiating neurons and muscle cells to allow inclusion of PTBP1 repressed exons during development of these tissues [19], [20], [21]. In neurons the loss of PTBP1 is accompanied by the up-regulation of the homologous protein PTBP2 [17], [20], [22]. PTBP2 has similar binding properties to PTBP1 and represses some of the same exons [23]. Other exons are more sensitive to PTBP1 than PTBP2 and are induced to splice when PTBP2 replaces PTBP1 in early neurons [24]. PTBP1 contains four RRM domains that recognize short pyrimidine elements [25]. Flexible linkers separate RRM domains one and two, and domains two and three. RRM domains three and four interact through a hydrophobic interface that position their RNA binding surfaces on opposite faces of the two-domain structure. This orientation requires that the RNA elements interacting with the structure be separated by an RNA loop [26]. The structure of each of the PTBP1 RRM domains has been solved in complex with the hexanucleotide, CUCUCU [25]. These structures show each domain binding a nucleotide triplet with some additional contacts, and making similar base specific interactions with CU or UC dinucleotides. Other sequences can likely make different base specific contacts, and the optimal elements for each domain are not known. Moreover, the flexible linkers separating some of the RRM domains and the requirement for a gap between elements simultaneously bound to domains three and four allow for substantial degeneracy in PTBP1 binding sites. This degeneracy and the lack of understanding of the sequence features that contribute to binding affinity have made it difficult to identify PTBP1 binding sites based on sequence alone, and to assess which sequences surrounding an exon might contribute to PTBP1 regulation. Experiments with model substrates indicate that a single high affinity PTBP1 binding site placed upstream of an exon, or within it, can repress splicing [27]. However, strong repression of an efficiently spliced exon requires an additional binding site either within the exon or downstream from an exon with an upstream high affinity site [17], [27], [28]. PTBP1 is also known to enhance the splicing of certain exons [13], [19], [20]. The properties of these exons and how they differ from those that are repressed by PTBP1 are unclear, with different studies coming to different conclusions [13], [19]. An analysis of CLIP data in HeLa cells found that PTBP1 sites near the adjacent constitutive exons could enhance the inclusion of an alternative exon between them [13]. In contrast, examination of exons whose splicing was reduced by double knockdown of Ptbp1 and Ptbp2 found that they frequently had binding sites immediately downstream [19], whereas splicing repression often involved upstream binding sites: a pattern observed for other splicing regulators. These results are not mutually exclusive. It is possible that the two groups examined different subsets of the many exons regulated by PTBP1, and that the protein may show additional patterns of protein binding adjacent to its target exons. In this study we sought to understand the sequence features that determine RNA binding by PTBP1 and to examine how they are combined in exons that are targeted by the protein. We first developed a statistical model of PTBP1 binding sites that identifies new features of RNA recognition by the protein. This binding model was then applied to the assessment of exon regulation by PTBP1 across the transcriptome. To examine the interactions of PTBP1 across many binding sites, we used a set of PTBP1-bound sequences identified by crosslinking immunoprecipitation (CLIP) [13]. PTBP1 has four RRMs separated by linker peptides, with each RRM recognizing a pyrimidine triplet. In previous studies we found that a minimal high affinity binding site for the protein extended across 25 to 30 nucleotides, about the average size of the CLIP clusters (29 nt) [27]. Given the triplet recognition and the need for spacers between the direct RRM contacts, it is unlikely that every nucleotide within a CLIP cluster makes a direct base-specific contact with the protein or otherwise contributes to binding affinity. This information about direct binding is hidden in the examination of a CLIP tag, but should affect the triplet frequencies within the entire set of tags. We designed a two-state Hidden Markov Model (HMM) based on triplets to assess whether triplets would segregate into two states and whether these two states differed in their PTBP1 binding or non-binding potential. The 48,604 CLIP clusters from the human transcriptome were extracted and used to train the HMM (Figure 1A) [29], [30]. This training defined two states showing distinctly different triplet distributions (Figure 1B). Pleasingly, all of the pyrimidine triplets segregated into State 1. We called this state the PTBP1 binding state, as we confirm below. We found that 20 triplets have higher probabilities to be seen in the PTBP1 binding state. All triplets containing only pyrimidines were included in this 20-triplet set (Figure 1B), with the top-scoring triplet UCU showing the alternating C and U nucleotides seen in many characterized PTBP1 binding sites. Interestingly, multiple triplets containing G residues are also preferred in State1 (Figure 1B). These triplets often contain U residues as the other nucleotides. Some of these triplets, such as UGU, have output (emission) probabilities in State 1 that are similar to pyrimidine triplets, presumably also making them predictive of PTBP1 binding. In contrast, triplets containing A residues, even if the other two nucleotides are pyrimidines, were all preferred by the non-PTBP1 binding State 2. These results indicate that PTBP1 is not strictly pyrimidine specific. At least one of its RRM domains can presumably make specific contacts with G residues. On the other hand, all A containing triplets have modest positive emission probabilities for state 2 and are likely to be either neutral or to inhibit PTBP1 binding. We next tested the HMM scoring, which strongly weights the triplets from state 1 over state 2, for prediction of PTBP1 binding. We performed cross validation experiments on the Hela CLIP dataset. A background dataset was generated using ten randomly picked sequences from each gene identified as containing a CLIP cluster. Applying the model to this data set gave us a distribution of scores that was compared to scores generated by subsets of the CLIP clusters removed from the training set prior to training. As shown in Figure S1, sequences from subsets of the CLIP clusters scored significantly higher than background. We also tested our model on an independent iCLIP dataset from human embryonic stem cells (ESC) (Figure S2). Unlike standard CLIP, iCLIP tags define the probable crosslink site as being the 5′ terminus of the tag. We used a Viterbi algorithm to predict the most probable state path predicted by the PTBP1 HMM model for each iCLIP tag. Defining triplets from the State1 (PTBP1 binding) and triplets from State 2 (nonbinding), we found that the frequency of predicted binding triplets is highly enriched in the iCLIP cluster regions and peaks precisely at the crosslink site. This indicates that State 1 probability is highly associated with PTBP1 crosslinking in vivo. To more quantitatively assess the relationship between the HMM score and RNA binding, we applied the trained model to a set of 100,000 random 69 nucleotide sequences. This length allows for one hexanucleotide binding site for each of the four RRMs with 15 nucleotide gaps, the minimum gap required for simultaneous binding by RRMs 3 and 4 [25], [26]. The scores are calculated as a log-odds ratio of the probabilities of the sequence having been generated by the HMM over a background model that assigns equal probability to all triplets. The random sequences generated a distribution of scores that was used to normalize the binding scores, with the average score for random sequence set to zero, and the z-score defined as the deviation from the average as shown in Figure S3A [29]. Thus a sequence with a z-score of 2.74 is 2.74 standard deviations from the average (empirical p-value = 0.005), and is predicted to be a significantly stronger binder than the average sequence (500 of the 100,000 random sequences have scores equal or greater than this sequence). A negative z-score is predicted to bind less well than the average sequence. We isolated thirteen sequences from the mouse transcriptome that exhibited a range of scores from −2.62 to +4.40 (Figure 2A). These were transcribed in vitro and subjected to electrophoretic mobility shift assay to measure binding to recombinant PTBP1 (Figure 2B; Figure S3B). Sequences yielding negative scores all failed to bind PTBP1 within the protein concentration range tested, with the exception of probe 4, which bound weakly, below the level that would allow measurement of an affinity constant. Positive scoring sequences all yielded PTBP1 bound complexes that were assayable by gel shift to derive apparent binding affinities. The apparent Kds of these RNAs showed a very strong negative correlation with their binding score from the model (Pearson correlation coefficient = − 0.9), where a higher score predicts a lower Kd and hence a higher affinity (Figure 2A). Thus, the scoring system performed very well in predicting PTBP1 binding affinity. Two sequences (probes 9 and 11) showed variable binding that shifted their Kd's slightly off the fitted curve relating z-score to Kd. These may have secondary structures that reduce binding affinity thus increase their apparent Kd. To look at this, we examined the predicted structure of each probe using the RNA fold program [31]. Probes 9 and 11 did not show an overall free energy of folding substantially lower than other RNAs. However, it is difficult to rule out that they contain a local structure that sequesters some key feature for PTBP1 recognition. In addition to the background model using uniform triplet frequencies, we also tested control sequence sets using different nucleotide frequencies (Figure S4). Control sets that maintain the mono or dinucleotide frequencies of the PTBP1 CLIP tags while shuffling the triplet frequencies did not perform well. This is not surprising because these sequences are highly skewed in nucleotide content and the shuffling does not change the triplet frequencies dramatically. We also tested a background model based on random sequences selected from genes containing PTBP1 CLIP clusters (ten sequences from each gene). Like the random dataset, this background model generated scores that predicted affinity reasonably well. However, it did generate negative scores for a couple of probes that are shown to bind (data not shown). Thus, the uniform model gave the most accurate scoring of the background models we tested. The data demonstrate that HMM scoring based on triplet frequencies can accurately predict the observed binding affinities across a wide range of Kd values (from ∼250 nM to 1 nM). Probe 6 yields a z-score of 0.82 and binds with a Kd of 257 nM, whereas probe 10 scores 2.74 in the model and binds with a Kd of 73 nM (Figure 2B). These sequences include G containing triplets that contribute to the binding scores. This method allows any sequence to now be quantitatively assessed for possible PTBP1 binding, which was not previously possible by simply looking for clusters of a limited number of motifs. This HMM based approach should be applicable to the prediction of binding sites and affinity for other multi-domain RNA binding proteins. With our new method of defining PTBP1 binding sites, we next examined PTBP1 target exons for the location of predicted PTBP1 binding. In part, we wanted to reassess two previous studies that came to differing conclusions regarding the placement of PTBP1 sites adjacent to its target exons. One group mapped PTBP1 CLIP clusters adjacent to a limited number of PTBP1 repressed and enhanced exons [13]. This study described PTBP1 repressed exons as enriched for binding sites both upstream and downstream, as has been seen in studies of individual exons. They did not observe PTBP1 CLIP clusters within repressed exons, even though such exons have been described [17], [32], [33]. The PTBP1 enhanced exons they examined showed a trend in PTBP1 binding near the flanking constitutive exons. A second study examined exons showing altered splicing on splicing-sensitive microarrays after Ptbp1/Ptbp2 double knockdown [19]. CLIP clusters derived from the first study were mapped to these exons. The authors found CLIP cluster enrichment upstream and within PTBP1/PTBP2 repressed exons. In contrast to the previous study, they found that PTBP1/PTBP2 enhanced exons showed enrichment for CLIP tags in the downstream region. This pattern of binding site placement relative to repressed and enhanced exons has been observed for several other splicing regulatory proteins [14], [34]. In our study, we defined four groups of exons from a set of exons previously assessed for splicing after Ptbp1 knockdown [20], [35]. These included 68 PTBP1-repressed exons whose splicing increases after Ptbp1 knockdown, 37 PTBP1-enhanced exons whose splicing decreases after knockdown, 69 control exons that are not affected by Ptbp1 depletion but are known to be alternatively spliced (PTBP1-non regulated), and 1,000 constitutive exons. We determined the density of predicted PTBP1 binding states within a 24-nucleotide window sliding along the exon region. We also examined the sequence encompassing the adjacent constitutive exons (Figure 3A). As expected, the non-regulated control and constitutive exon sets did not exhibit high probabilities of PTBP1 binding except in the polypyrimidine tract of the 3′ splice site. On the other hand, the introns upstream of PTBP1 repressed exons show enrichment of potential PTBP1 binding sites starting from 250 nucleotides upstream of the exon. Relative to the control exons, exons repressed by PTBP1 also exhibited substantial enrichment of PTBP1 binding sites within the exon itself and within the first 100 nucleotides of the downstream intron. The repressed exons thus exhibit binding site placement that combines the findings of the two previous studies [13], [19]. The PTBP1-enhanced exon set also shows enrichment of PTBP1 binding sites within the downstream intron relative to control exons, although the distribution of binding sites across this region was different between the repressed and enhanced exon sets (Figure 3A). Similar to what was seen in the previous study by Llorian, we found little enrichment of PTBP1 sites within enhanced exons [19]. There is a limited enrichment adjacent to the exons flanking enhanced exons. Interestingly however, we find some PTBP1 enhanced exons that have PTBP1 binding sites upstream of the exon. These were not seen in either previous study. Our results are generally consistent with the known placement of PTBP1 binding sites in PTBP1 target exons and imply that rules correlating the position of PTBP1 binding to its effect on a target exon are not as strict as seen for some other splicing regulators. The mechanisms proposed from previous maps of PTBP1 binding do not appear to be generalizable to all PTBP1 targets [13], [19], [27]. Binding maps for PTBP1 and other splicing regulators show the averages of multiple exons. Since the data indicated a high level of variability in binding site placement between individual exons, we wanted to visualize target exons relative to each other. To display binding signals for individual exons we created heat maps of the binding scores upstream, within, and downstream of each exon in the PTBP1 target set (Figure 3B). This display makes clear that the location of PTBP1 binding sites within its known target exons is variable. We found that 60% of PTBP1 repressed exons are predicted to have strong binding sites within the upstream intron. Most of these exons also have strong binding sites within either the exon or the downstream intron, patterns that were observed previously [13], [19], [27]. However, other patterns of binding site placement are also seen, suggesting PTBP1 dependent exons are following multiple rules. Some repressed exons score highly for PTBP1 binding only within the exon or in both the exon and the downstream intron. About half of PTBP1 enhanced exons have strong PTBP1 binding sites downstream (Figure 3B). These can co-occur with upstream intron-binding sites, but rarely with exon binding sites. Interestingly, there are exons enhanced by PTBP1 with strong upstream binding in the absence of other sites. These data demonstrate the heterogeneity in the position of PTBP1 binding sites for its target exons. This heterogeneity needs to be considered for predicting PTBP1 dependent regulation. PTBP1 repressed exons exhibited significantly higher average binding scores in both the upstream intron and in the exon itself, than either the control group of alternative exons or the PTBP1 enhanced exons (Figure 3C). The average binding scores in the downstream introns were higher for both the PTBP1-repressed and PTBP1-enhanced exons than the control group (Figure 3C), although not at the same statistical significance. The variability of binding site placement within the smaller group of PTBP1-enhanced exons presumably contributes to the weaker statistical correlation of binding scores with positive regulation. We also compared the three exon sets for other features that might contribute to their ability to be regulated by PTBP1, including exon length, flanking intron length, and 5′ and 3′ splice site strength. Most of these features were not statistically different among the three-exon groups. However, both PTBP1 enhanced and PTBP1 repressed exons were found to carry significantly weaker 3′ splice sites than the control exon set, as measured by the Analyzer Splice Tool (Figure 3C) [36], [37]. These results indicate that PTBP1-repressed exons, and perhaps PTBP1-enhanced exons, exhibit an ensemble of sequence features that define them as PTBP1 regulated and that should allow their identification by sequence alone. Alternative exons are generally regulated by multiple factors that act both positively and negatively on their ability to be spliced. Thus, an exon controlled by a regulator in one context might not be affected by it under other conditions where counteracting factors are present, or required cofactors are absent. This means that the most accurate predictions of splicing regulation will need to consider many different factors. Nevertheless, models based on single factors will be useful for understanding the relative contributions of individual proteins to patterns of splicing regulation. Such models will be easier to interpret regarding the contributions of individual factors to individual exons than more complex models. Moreover in the longer term, models developed for different individual factors can be combined to make more accurate predictions. To assess how well one might model splicing regulation by a single factor, we examined whether the strength and placement of predicted PTBP1 binding sites could be used to predict new PTBP1 dependent exons. We plotted the scores for a variety of sequence features against the percent of exons exhibiting that score that also exhibit PTBP1 dependent exon repression (Figure S5). These plots produced distinct sigmoidal curves where most exons regulated by PTBP1 were found above or below a particular score. This strongly suggests that a logistic regression model incorporating each of these scores will be predictive of PTBP1 repression. We developed a multinomial logistic regression model and trained it on three classes of regulated exons (Figure 4A) [38]. The training set included PTBP1 repressed exons, PTBP1 enhanced exons, and non-regulated exons. Each exon in each class was scored for the four features found to correlate with PTBP1 regulation (x1 through x4), including the 3′ splice site strength, and the PTBP1 binding scores for each of three regions: the 250 nucleotides upstream of the exon, the exon itself, and the 100 nucleotides downstream of the exon. These intron lengths encompass the regions of binding site enrichment for PTBP1 dependent exons (Figure 3). The PTBP1-enhanced exons are fewer in number and show more limited enrichment of PTBP1 binding sites than PTBP1-repressed exons making the prediction for these exons less accurate. We first tested models that considered just PTBP1-repressed exons relative to control groups. However, we found that including the enhanced exons as a separate training group improved the prediction of repressed exons, even though enhanced exons themselves are not as easily identified (data not shown). The trained model yielded values for the β coefficients that weight the different features contributing to the regulation. As expected the upstream binding score was weighted most heavily in predicting PTBP1 repression (Table S1), although binding scores in all three regions contributed to the score for PTBP1 repression. In contrast, we found that only the downstream binding score was significantly associated with PTBP1 enhancement. The upstream score generated a β coefficient close to zero making it essentially neutral in the prediction of enhanced exons. The exon binding score was subject to a negative β coefficient, indicating that exon binding reduces the probability of PTBP1 enhancement. Using these β coefficients, the trained models for repression or enhancement each yield a value of the g-function (logit) for an exon (x) given by the log of the ratio of the probability of repression or enhancement over the probability that the exon is not regulated. From this, the probability that an exon is repressed by PTBP1 can be determined from the two g-values as shown in Figure 4A. We assessed the multinomial logistic regression model by recursively retraining on exon sets with one exon left out and then scoring the missing exon. This leave-one-out cross validation enabled assessment of the overall performance of the model [38] (Figure S6). The PTBP1 dependent exon repression logit showed good prediction, with an area under the curve (AUC) value of 0.72, substantially greater than random guessing (AUC = 0.5). As expected, the enhanced exon logit was not as accurate as the repression logit (AUC = 0.57), although it was better than random (Figure S6A). Using these data, we assessed the sensitivity and specificity across the range of scores to define a decision threshold for exon repression scores (Figure S6B). Increasing the threshold increases the specificity by eliminating many false positives, but decreases the sensitivity of the model in identifying maximum numbers of repressed exons. We sought to choose a threshold that gave a low false positive rate over one that yielded more regulated exons. We found that above a threshold score of 0.65 the false positive rate was 10% or lower (Figure S6B). Applying the model to 4494 alternative cassette exons from UCSC genome browser database, we found 243 exons (5.4%) that yielded a PTBP1 repression probability score greater than 0.65 and which were not in the training set. The 50 top-scoring cassette exons are listed in Table 1. These included two exons that were reported previously to be PTBP1 targets. An exon of Gabrg2 yields a probability score of 0.92. Although we could not confirm its repression in N2A cells because of low expression of the transcript, the orthologous exon in rat is a well-characterized PTBP1 repression target [39]. Exon 2 of Ptbp3 (Rod1), another known PTBP1 target [40], yielded a repression probability score of 0.89 and was confirmed by RT/PCR to show increased inclusion after Ptbp1 knockdown (Figure 4B). We performed additional RT-PCR validation in triplicate on a series of high and low scoring exons from transcripts expressed in N2A cells (Figure 5 & Figures S7, S8 and S9). Seven of ten exons scoring above 0.65 were de-repressed after Ptbp1 knockdown in N2A cells, yielding a validation rate of 70%. The actual false positive rate is difficult to estimate because exons with high repression scores that are not affected by Ptbp1 depletion in N2A cells might be regulated by PTBP1 in other cells. An indication that this might be occurring is that the average inclusion level (or percent spliced in value, PSI) of the putative false positives is significantly higher than the confirmed true positives in N2A cells, indicating that they will be less prone to change upon Ptbp1 depletion and be more difficult to validate (Figure S8B). Thus, the true positive rate may be greater than 70%. Importantly, the high validation rate for exons scoring above 0.65 indicates that the binding model and the regulation model based upon it can identify many new PTBP1 targets that were not previously known (Table1). High scoring exons might also fail to be validated because of regulation by other proteins. Knockdown of Ptbp1 induces expression of its close homolog Ptbp2, which targets some of the same exons [20] (Figure S7). To test whether PTBP2 was also targeting the predicted PTBP1 repressed exons, we knocked down Ptbp2 or both Ptbp1 and Ptbp2 expression in N2A cells and re-assayed the exons in triplicate (Figures S10, S11 & S8A). Although some exons showed greater inclusion in the double knockdown compared to depletion of Ptbp1 alone, this did not validate any additional predicted PTBP1 repressed exons. We did identify some high and low scoring exons showing more complex regulation by the two PTB proteins (Figure S10 & S11). We also examined a set of low scoring exons (probability score≤0.2) by RT-PCR after Ptbp1 and/or Ptbp2 depletion (Figure 5B and Figure S11). All of these exons (8 of 8) failed to respond to the loss of PTBP1 and are likely true negatives. Thus, PTBP1 repression scores above 0.65 and below 0.2 were highly predictive for regulation and its absence, respectively. As expected, intermediate scores were less consistent in their predictive value (Figure S9). Some exons in the intermediate scoring group were affected by PTB proteins and will be interesting to assess further. The prediction of PTBP1-repressed exons was improved by treating PTBP1-enhanced exons as a separate class, but the probability scores for PTBP1 enhancement did not consistently identify new PTBP1 target exons (data not shown). This is likely in part due to the smaller number of exons in the training set and their heterogeneity, with some possibly being indirect targets. These predictions will likely improve with training on larger numbers of PTBP1 enhanced exons as they are identified. However, it is possible that simply the presence of the PTBP1 binding site is not sufficient for predicting PTBP1 enhancement and that binding sites for other factors will need to be considered. We next tested the model on a genomewide scale, by applying it to a set of 168,111 mouse internal exons and ranking them by their probability of PTBP1 repression. This analysis yielded 3824 exons (2.3%) with probability scores above 0.65 for being repressed by PTBP1. Among other activities, these exons were enriched in genes that function in calcium ion transport, cytoskeletal organization, intracellular transport, and synaptic transmission, all functions affected by previously known PTB targets (Table S2). To assess splicing of this large set of predicted PTB targets, we used RNA-seq to generate a large dataset of exons that change after Ptbp1 knockdown. RNA from control and PTBP1-depleted N2A cells was subjected to high density short read sequencing on the Illumina HiSeq platform using a strand specific, paired end protocol [41]. Exons whose inclusion changed between the two samples were identified by alignment to an exon database and quantification of exon inclusion using the SpliceTrap program [42]. After filtering for read coverage and removing the training set, we identified 573 alternative exons whose splicing was assayable in N2A cells. These exons exhibit changes in percent exon inclusion (delta PSI) ranging from −29% to 62% upon PTBP1 depletion. The exons were binned by their PTBP1 repression probability scores and plotted for their change in PSI (Figure 6). The average changes in splicing were significantly correlated with the repression probability. Exons scoring below 0.5 distributed around zero change in PSI, but above this score the average exon inclusion is altered by PTBP1 depletion. Most notably, exons with a repression probability score above 0.65 exhibited significantly larger changes in splicing than exons with lower scores. Exons with intermediate scores and hence weaker binding sites show smaller changes in splicing than high scoring exons. Setting a threshold of a 5% change in PSI as validation, 22 of 33 exons (67%) that scored above 0.65 for PTBP1 regulation were confirmed as PTBP1 repression targets in N2A cells. At least some of the other 11 exons are presumably PTBP1 targets in other cells. To test the model in another cell type, we examined exons reported to change after Ptbp1 knockdown in mouse C2C12 myoblasts, as measured on splicing sensitive microarrays [43]. Very similar to what was observed in N2A cells, we found that exons with high repression probabilities showed significant de-repression upon the Ptbp1 knockdown compared to exons with low repression probabilities (Figure S12). Of 29 exons assayed on the arrays with a repression probability above 0.65, 19 exons were confirmed as PTBP1 repressed on the array (q-value<0.05), yielding a validation rate of 66%. Thus the model performed very similarly in C2C12 and N2A cells. Among the 11 high scoring exons identified as unchanged after PTBP1 knockdown in N2A cells only 3 were assayed on the array and expressed in C2C12 cells. These again showed high inclusion in C2C12 prior to knockdown and so were difficult to assay for derepression. Thus, it is difficult to use the C2C12 data to draw conclusions about the false positive rate. The logistical model gives us a new tool for studying the regulation of alternative splicing. Using it, we can now scan genomic sequence to score exons for PTBP1 regulation. Applying the model genomewide, the PTBP1 repression probability scores were integrated into the UCSC genome browser. These data, displayed with the RNAseq data from N2A cells are available at our website (http://www.mimg.ucla.edu/faculty/black/ptbatweb/). A novel PTBP1 repressed exon in the Kcnq2 gene is shown in Figure 6B. The logistic model thus allows the assessment of any exon across the transcriptome for likely PTBP1 regulation. We have developed two computational models, one that allows accurate prediction of PTBP1 binding sites and another that predicts likelihood of PTBP1 repression of exons across the transcriptome. These models uncovered several new features of RNA recognition by PTBP1 and the properties of its target exons. The PTBP1 binding model was based on triplets following the structures of the PTBP1 RRM domains, whose sequence specific contacts are each primarily to three nucleotides. We find that the set of triplets that increase the probability of binding includes the expected pyrimidine motifs, particularly those with alternating cytosines and uridines. However, many triplets with guanosine residues also increase binding probability. In contrast, adenosine residues have a negative effect on binding. Thus, RNA recognition by PTBP1 is not solely dependent on pyrimidine nucleotides. The recognition of G residues by PTB was unexpected, although some previously characterized PTB binding sites did contain G residues [13], [44]. With this model, we can now predict PTBP1 binding affinity to any site in the transcriptome. The base-specific contacts that PTBP1 makes with Guanosine are not yet clear. Recent studies of RNA recognition by SRSF2 (SC35) protein have shown that the element GGAG can be recognized by the same RRM as CCAG by flipping the initial two G nucleotides to the syn conformation [45]. It will be very interesting to investigate whether a similar anti to syn switch occurs in RNA bound by PTBP1, when C residues are replaced with G. Previous characterizations of PTBP1 binding sites have focused on finding enriched short motifs within populations of bound RNAs or regulated exon sequences [13], [44], [46], [47], [48]. These methods generally identify elements whose short length will allow interaction with only one RRM domain. Searching for new binding sites comprised of clusters of these short elements can identify higher affinity sites but does not consider all elements or rank them. Crosslinking-immunoprecipitation experiments allow large numbers of binding regions to be identified. However, not all the sequence within a CLIP tag will be contacting the protein and it is difficult to relate CLIP signals to binding affinity. The HMM allowed the individual assessment of different short elements within the CLIP clusters, showing that they segregated into two states. The ranking of the triplets for their contributions to one of these states yielded a model where complex clusters of short elements could be assessed for binding and yielded accurate predictions of binding affinity. Many RNA binding proteins are similar to PTBP1 in having multiple domains that may each make different base specific contacts with RNA. The widespread generation of CLIP-seq datasets will allow the modeling of RNA recognition by almost any protein based on a large number of known binding sites. Using the same modeling approach, we also developed a binding model for PTBP2 (neuronal PTB) using a published PTBP2 CLIP dataset [49]. PTBP2 is about 70% identical to PTBP1 in sequence, and has only two amino acid changes among the residues making direct contact with RNA [17]. We found that the binding models for two PTB proteins were also nearly identical indicating that the two proteins are likely to differ more in their protein/protein interactions than in their RNA binding sites (Data not shown). Several PTBP1 target exons have been analyzed in detail [17], [50]. These exons vary in the placement and action of their PTBP1 binding sites. It is common for PTBP1-repressed exons to have a binding site upstream, often encompassing the branch point of the 3′ splice site [39]. Exons can also be repressed by PTBP1 binding within the exon [19], [32], [33]. Other exons contain downstream binding sites that are needed in conjunction with an upstream site to achieve splicing repression [51], [52], [53]. Although acting as a repressor for most of its targets, PTBP1 also activates the splicing of a group of exons. There have been divergent reports about placement of PTBP1 binding sites needed to mediate PTBP1 enhancement of splicing. The PTBP1 binding model allowed us to examine PTBP1 binding site placement across a large set of known PTBP1 target exons. Nearly all exons had predicted high affinity PTBP1 binding sites nearby. We found that more than half of PTBP1 repressed exons have high affinity binding sites upstream, and a fraction of PTBP1 enhanced exons have high affinity sites downstream. These exons fit with recent results on several other splicing regulators where the placement of the binding site determines the direction of the regulatory effect [12], [14], [34]. However, for PTBP1 these rules are not so clear. Some PTBP1 repressed exons have their strongest predicted binding site downstream or within the exon. These results indicate that there are fundamental differences between the mechanisms of PTBP1 mediated splicing regulation, and those governing regulation by certain other splicing factors. To quantify the predictive value of the PTBP1 binding scores for PTBP1 repression, we built a logistic model for PTBP1 regulation. For exons repressed by PTBP1, binding scores for the upstream, downstream and exon sequences all contribute to the probability of repression. Exons enhanced by PTBP1 were too few to achieve accurate predictions from the model. However, treating these as a separate exon class improves the prediction of PTBP1 repression. We find that for probability scores above 0.65 the model is strongly predictive of PTBP1 repression. Applying this criterion across the transcriptome, we identified hundreds of new PTBP1 target exons. Alternative exons are generally regulated by multiple proteins acting in combination, and a particular exon will often be subject to both positive and negative regulation by antagonistic factors. For a model based on one factor, these other proteins will confound predictions. Exons with high PTBP1 binding scores may be counteracted by antagonistic factors in some cell types. Alternatively, synergistic factors may allow an exon with a relatively weak binding site to still recruit PTBP1. Thus, a model based on one factor will be limited in its predictive power. In this study, our intent was to measure the effect of PTBP1 binding alone before considering the contributions of other factors. The logistic modeling allowed the contributions of different binding site placements to PTBP1 regulation to be measured. Several studies have used Bayesian models to dissect the regulatory properties of exons [7], [54]. These models can generate accurate predictions by incorporating a wide variety of sequence, expression and conservation data. However, because so many disparate variables are incorporated, it can be difficult to draw mechanistic conclusions from these models regarding any one protein. For example, the presence of high pyrimidine density upstream from the branch point can be predictive of exons showing neuronal specific inclusion [7], [55]. This is presumably in part due to many neuronal exons being regulated by PTBP1 and PTBP2. However, a subset of these exons may be regulated by other factors with pyrimidine rich binding sites. In the long term, it will be most accurate to develop predictive binding models for each protein, similar to the PTBP1 model here, and then to incorporate each of these binding models into a larger network model. Such an approach will allow the analysis of the many overlapping regulatory programs controlled by RNA binding proteins. A Hidden Markov Model (HMM) was designed and trained by an expectation–maximization (EM) method (Baum-Welch algorithm) using published PTBP1 CLIP data [13], [29], [30]. In total, 48,604 PTBP1-CLIP cluster sequences were used to train model parameters. During the training step, multiple initial values were tested to avoid a local maximum problem. Trained parameters included emission probabilities for nucleotide triplets, initial probabilities and transition probabilities between states [29], [30]. The trained model was used to score RNA sequences. The raw PTBP1 binding score is defined as a log-odds ratio that compares the score of a sequence from the HMM over the score from a background model. Since CLIP experiments do not have an inherent corresponding negative dataset, we generated computational negative datasets and tested different background models (Figure S4). We found that a background model that values all triplets equally yielded the most accurate binding scores [29]. Raw scores were further normalized and converted to z-scores. For the 69 mer RNA sequences used in binding assays, scores were normalized by 100,000 random sequences with same length (Figure S3). This yielded very accurate predictions of binding affinity (Figure 2). When considering binding scores in genomic sequence, exons and upstream or downstream intron regions have different base compositions and will yield different average binding scores. Thus, to score binding sites adjacent to possible regulated exons, it is more informative to score sites relative to equivalent sequence regions. From the annotated mouse genome, we retrieved 168,111 internal exons and their flanking introns as separate sequence sets using a python library, Pygr. We scored log odds of these sequences with the trained model. Since the lengths and base compositions of intronic and exonic sequences are different, and binding scores automatically increase with length (Figure S13) [29], we grouped sequences by their location and sequences in each group were sorted according to length into bins of 1000 sequences each. The average score and standard deviation were determined for each bin. These values were used to transform the raw scores into z-scores for each upstream intron, downstream intron, and exon sequence. We localized the PTBP1 binding sites along each RNA sequence using the Viterbi algorithm [29], [30]. To test predicted PTBP1 binding scores, we selected thirteen mouse exon/intron RNA sequences (69 nucleotides) exhibiting a range scores. In the selection, other sequence features such as secondary structure were not considered. Target RNAs were transcribed in vitro from dsDNA using T7 RNA polymerase and subjected to an electrophoretic mobility shift assay (EMSA). During the transcription, radioactive α-32P UTP was incorporated into RNA to visualize the probes. The RNA probes were then denatured for 2 min at 85°C and cooled down on ice immediately to reduce secondary structure formation. Binding assays were carried out as previously described with some modifications [27]. Specifically, each gel mobility shift reaction (10 µL) contained the indicated amounts of recombinant human PTBP1 in 6 µL DG buffer (20 mM Hepes-KOH ph 7.9, 20% glycerol, 80 mM potassium glutamate, 0.2 mM EDTA, 0.2 mM PMSF), 1 µL 22 mM MgCl2, 1 µL 0.5 mg/ml tRNA, 0.5 µL RNase inhibitor (20 unit, RNaseOut from invitrogen), 0.5 µL DEPC treated H2O, and 1 µL 100 nM RNA probe. At first, all reaction components excluding RNase inhibitor, tRNA, and RNA probes were mixed and incubated for 8 min at 30°C. Then RNase inhibitor and tRNA were added and mixed. RNA probe was then added and the reaction was incubated for an additional 15 min. The reactions were put on ice for 5 min and mixed with 1.2 µL glycerol loading dye (30% glycerol). They were separated on 8% native polyacrylamide gels with 25 mM Tris-Gly running buffer in a cold room. Gels were dried and exposed to a phosphor screen. Then images were scanned using Typhoon 9410 and quantified using ImageQuant TL program (GE Lifesciences). The apparent Kd values were estimated by fitting the data to non-linear curves using Prism software. An exon training set was compiled from previous microarray and RT-PCR experiments [20], [35]. The training set was composed with 68 PTBP1 repressed, 37 PTBP1 enhanced, and 69 non-PTBP1 regulated simple cassette exons. We only considered exons with canonical splice sites (GU-AG). An exon was classified as PTBP1 repressed or enhanced when 1) the inclusion level (PSI) of its minor isoform was greater than 5% in both the control and knock-down samples and 2) the inclusion level of its minor isoform was changed by 30% or more in the Ptbp1 knock down condition compared to the control sample. Next, we collected sequence features for each exon and its flanking exons. The features included PTBP1 binding scores, 5′ and 3′ splice site strengths, exon/intron lengths, and word frequencies. The PTBP1 binding scores were calculated from the PTBP1 binding model described above. The strength of splice sites was calculated by the splice-site analyzer tool [37]. Using a mouse whole internal exon set, we normalized features and fed them into the model. The PTBP1 splicing model is based on a multinomial logistic regression framework using the following steps: 1) selection of initial variables with a moderate level of association (p-value from t-test<0.25), 2) removal of outlier exons, 3) stepwise variable selection [38]. We scored mouse internal exons with the trained PTBP1 splicing model and validated candidate exons with RT-PCR and RNA-seq experiments. Exons from the training set were excluded from the validation. To test alternative splicing events for candidate exons, we assayed exon inclusion levels in cells following Ptbp1, Ptbp2, and double Ptbp1 & Ptbp2 knock down. The knockdown experiment was performed as described previously with minor modification [20]. Mouse neuroblastoma (N2A) cells were cultured in DMEM with 10% FBS and 2 mM L-glutamine. At 70 to 80% confluency, cells were trypsinized and suspended in the growth medium. DNA–Lipofectamine 2k (Invitrogen) complexes were prepared and mixed with cells in a tube according to manufacturer's instructions. Tubes were incubated for 5 h with mixing every half hour. Then cells were centrifuged and cultured in plates for 3 d. Proteins and RNA was extracted from collected cells. Protein samples were subjected to fluorescence immunoblotting to monitor knockdown efficiency of Ptbp1 and Ptbp2. Total RNA was collected using Trizol (Invitrogen) according to the manufacturer's instructions. The RNA was further treated with DNase I to avoid DNA contamination. For RT-PCR (Reverse Transcription-PCR) assays, the RNA was reverse transcribed to cDNA with random hexamers using SuperScript enzyme (Invitrogen) following the manufacturer's instructions. PCR reactions were performed to assay alternative splicing of particular target exons. First, forward and reverse PCR primers were designed for the flanking exons using PRIMER3 program [56]. To label PCR products, a 5′ fluorescent-labeled universal primer (5′-FAM-CGTCGCCGTCCAGCTCGACCAG-3′) was added to the PCR reaction and a universal priming site was introduced to the 5′ end of the forward primer (5′-CGTCGCCGTCCAGCTCGACCAG-Forward Primer-3′). Each PCR reaction (15 µL) was carried out with 1.5 picomole of the forward primer and 6.75 picomole of the reverse and universal primers [57]. PCR amplification proceeded with an initial denaturation at 94°C for 4 m followed by 24 cycles of 94°C for 30 s, at a melting temperature of the reverse primer for 45 s, and 72°C for 45 s, with a final extension step at 72°C for 10 m. The samples were mixed with 2× formamide buffer (Formamide with 1 mM EDTA pH 8.0) and denatured at 95°C for 5 min. Then samples were chilled on ice and run on 8% denaturing polyacrylamide gels. Gels were directly scanned by Typhoon and quantified by ImageQuant program. RNA-seq libraries were constructed following standard protocols (Illumina TruSeq RNA Sample Prep Kit). To make strand-specific libraries, we added two extra steps to the protocol [41]. After first strand cDNA synthesis, remaining dNTPs were removed by a size selection on beads (AMPure XP). Second-strand cDNA was synthesized with a dNTP mix containing dUTP instead of dTTP. The reaction contained samples eluted in 50 µl resuspension buffer, 2 µl 5× FS buffer, 1 µl 50 mM MgCl2, 1 µl 100 mM DTT, 2 µl 10 mM dUTP nucleotides mix, 15 µl Second Strand Buffer (Invitrogen), 0.5 µl E.coli DNA Ligase (10 U/µl;NEB), 0.5 µl RNase H (2 U/µl;Invitrogen), 2 µl DNA E.coli Polymerase I (10 U/µl;NEB). The reaction was incubated for 2 h at 16°C. After sequencing adaptors were ligated, 1 µl USER (Uracil-Specific Excision Reagent enzyme; NEB) was added to reactions to degrade the second strand cDNA. The samples were incubated for 15 min at 37°C and the reaction were inactivated at 94°C for 5 min. The samples were put in ice and then subjected to PCR amplification. Average size of inserts was about 225 bp and the libraries were subjected to 100 bp paired-end sequencing (Illumina HiSeq2000 platform). Using SpliceTrap [42], 60–65% of reads were mapped to exon duos or trios. In total, 180M (179,511,116) and 145M (145,334,711) paired end reads were used to infer exon inclusion ratios in the control and Ptbp1 knockdown conditions, respectively. The data have been deposited in NCBI's Gene Expression Omnibus [58] and are accessible through GEO Series accession number GSE45119.

10.1371/journal.pcbi.1002657

Artificial Neural Networks Trained to Detect Viral and Phage Structural Proteins

Phages play critical roles in the survival and pathogenicity of their hosts, via lysogenic conversion factors, and in nutrient redistribution, via cell lysis. Analyses of phage- and viral-encoded genes in environmental samples provide insights into the physiological impact of viruses on microbial communities and human health. However, phage ORFs are extremely diverse of which over 70% of them are dissimilar to any genes with annotated functions in GenBank. Better identification of viruses would also aid in better detection and diagnosis of disease, in vaccine development, and generally in better understanding the physiological potential of any environment. In contrast to enzymes, viral structural protein function can be much more challenging to detect from sequence data because of low sequence conservation, few known conserved catalytic sites or sequence domains, and relatively limited experimental data. We have designed a method of predicting phage structural protein sequences that uses Artificial Neural Networks (ANNs). First, we trained ANNs to classify viral structural proteins using amino acid frequency; these correctly classify a large fraction of test cases with a high degree of specificity and sensitivity. Subsequently, we added estimates of protein isoelectric points as a feature to ANNs that classify specialized families of proteins, namely major capsid and tail proteins. As expected, these more specialized ANNs are more accurate than the structural ANNs. To experimentally validate the ANN predictions, several ORFs with no significant similarities to known sequences that are ANN-predicted structural proteins were examined by transmission electron microscopy. Some of these self-assembled into structures strongly resembling virion structures. Thus, our ANNs are new tools for identifying phage and potential prophage structural proteins that are difficult or impossible to detect by other bioinformatic analysis. The networks will be valuable when sequence is available but in vitro propagation of the phage may not be practical or possible.

Bacteriophages are extremely abundant and diverse biological entities. All phage particles are comprised of nucleic acids and structural proteins, with few other packaged proteins. Despite their simplicity and abundance, more than 70% of phage sequences in the viral Reference Sequence database encode proteins with unknown function based on FASTA annotations. As a result, the use of sequence similarity is often insufficient for detecting virus structural proteins among unknown viral sequences. Viral structural protein function is challenging to detect from sequence data because structural proteins possess few known conserved catalytic motifs and sequence domains. To address these issues we investigated the use of Artificial Neural Networks as an alternative means of predicting function. Here, we trained thousands of networks using the amino acid frequency of structural protein sequences and identified the optimal architectures with the highest accuracies. Some hypothetical protein sequences detected by our networks were expressed and visualized by TEM, and produced images that strongly resemble virion structures. Our results support the utility of our neural networks in predicting the functions of unknown viral sequences.

As modern sequencing technologies exponentially increase the amount of DNA sequence data available, the discovery of sequences that encode proteins with unknown functions continue to accumulate. For example, a large majority of microbial and viral metagenome sequences sampled from different environments have unknown function based on similarity to known sequences [1]–[4]. The remarkable biodiversity of viruses and the fact that sampling and in-depth genetic and biochemical studies of protein functions have been biased until relatively recently toward biomedically important or model organisms limits the utility of similarity-based annotation methods. Viruses, largely prokaryotic viruses (bacteriophages or phages) are the most abundant carrier of genetic material in marine environments [5], most of which are phages [6] that directly influence their host populations by lysing their hosts or by providing genes that confer selective advantages, such as antibiotic resistance, detoxifying enzymes, etc. Viral diversity is partly driven by viral structural protein genes, such as those encoding tails and tail fibers, which participate directly in the evolutionary contest between viruses and their hosts. Moreover, phage genes that encode proteins used in recombination mechanisms accelerate bacterial evolution through horizontal gene transfer and the development of new varieties of pathogenic strains [5]. Discovering the functions of unknown viral sequences is important for understanding the lifestyle and effects of viruses in the environment, the genetic relationship between viruses and their hosts, and the influence of viruses on the development of new pathogens. Roughly 85% of phages have a double stranded (ds) DNA genome [7], which is protected by a protein shell. The genomes of most characterized phages are introduced into a host cell through a tail structure [8]. Both head and tail structures are much more complex than previously thought [9]. The protein shell of a ds DNA bacteriophage is composed of subunits called capsomeres that polymerize into structures called procapsids or proheads. Further assembly and restructuring of procapsids generate the head structure that houses and protects the phage genome. Attached to the phage head via portal or connector proteins is a tail structure that has been used to classify tailed phages into families (http://www.ictvdb.org). Myophages have contractile tails, Siphophages have long non-contractile tails, and Podophages have short tails. Other proteins that are involved in the assembly of the phage particle may be degraded or left behind after phage assembly is completed and do not become part of the phage particle. Examples of these types of proteins are proteases, some scaffold proteins, and chaperone proteins. Evolutionary information from secondary structure alignments of the λ tail structure [10] and T4-like capsids is known [11], and the number of crystal and cryo-EM structures of numerous capsid and tail proteins from tailed phages is increasing. However, this information is restricted to a limited number of viruses, and the degree to which all phage structural protein sequences are similar to one another is not fully understood. The lack of sequence similarity is problematic because nearly all machine learning algorithms applied to biological data rely on conserved sequence motifs or functional domains. Dynamic Bayesian networks have been used to classify signal peptides [12] and to study secondary structure [13]. Support Vector Machines have been applied to classification tasks, such as the recognition of cysteine and histidine metal binding sites [14], and predicting sequence motifs in tertiary structures [15]. Hidden Markov Models have been successfully used in the prediction of HTH domains [16] and transcription factor binding sites [17]. In addition, neural networks have previously been trained by protein sequences with at least one conserved motif, such as 3 conserved catalytic residues in the phage integrase enzyme [18], conserved signal sequences in signal peptides [19]–[20], metal binding sites [14], transmembrane proteins [21], and protein functional domains from primary sequence alignments [22]. As an alternate to sequence similarity, protein fold recognition servers such as PHYRE [23], CSBLAST [24], and pGenTHREADER [25] may be used to compare an unknown sequence to known 3D structures by “threading”, a process that compares the fold profile of a query sequence to the fold profiles from known structures. Structure prediction servers, however, are poor at predicting the orientation of protein domains [23] and may match a query to several different types of proteins with similar domains, which may lead to false predictions. These are only a few examples from a long list of machine learning applications that predict protein function from primary DNA or protein sequence data. Invariably, these approaches rely on known sequence motifs or multiple sequence alignments to generate models. Sequence alignments between known and unknown structural sequences generally form the basis for homology assignments, in which an unknown primary sequence is annotated with the function of a known sequence that best aligns to the unknown sequence. Identifying conserved phage structural protein sequences by pair-wise sequence alignments, however, is very difficult if possible, because of inadequate data showing sequence similarity between known phage genomes [26]. Almost none of the structural proteins encoded by tailed-phages, with the exception of portal proteins, are identifiable by sequence similarity, which is too weak to be useful for classification tasks. Tail fiber proteins are also difficult to identify by computational methods because of extensive swapping of gene fragments between loci [27]. This highlights the challenge of predicting phage structural proteins: unlike enzymes that may share metal or nucleotide binding motifs, signal peptides, or transmembrane regions, structural protein sub-domains are not well conserved or well characterized. In addition, some structural proteins possess multiple functional domains. For instance, the procapsid of bacteriophage phi6 [28]–[29] and the major coat protein of filamentous phage M13 [30] contribute both to the virion structure and bind directly to nucleic acids. The morphogenesis protein of phage φ29 is a structural component of a tail fiber that also lyses the host cell wall during infection [31], and hence has both structural and enzymatic functions. Like many proteins of RNA viruses, nearly all hepatitis C virus proteins are multifunctional and contribute to viral assembly as well as replication [32]. As mentioned above, sequence similarities among viral structural proteins are known but extremely limited, although evidence of structural similarities for viral structural proteins have been accumulating. For example, a fusion of two-barrel folds, commonly called the double “jelly roll” fold, is found in the capsid proteins of the mammalian adenovirus [33], Sulfolobus turreted icosahedral virus (STIV), bacteriophage PRD1, and Paramecium bursaria Chlorella virus [34]. The capsid proteins of bacteriophage SPO1 and Herpes viruses share some structural similarity based on asymmetrical capsid surface molecules and triangulation number [35]. In addition, orthogonal sheets and loops are common in the proteins of the non-contractile tail of phage λ and the contractile tail of the induced prophage PBSx [10]. The use of structural information by X-ray crystallography may be ideal for predicting the function of an unknown protein sequence; however, crystallography is a lengthy and expensive process with a relatively low rate of success. Structural protein sequences are ideal targets for the detection of viruses because they are absolutely required and present in essentially all viruses, and in principle may serve as the analog of rRNA genes in the classification of cells. Here we describe the design of Artificial Neural Networks (ANNs) that detect virus and phage structural protein sequences based on the frequencies of amino acids predicted from the translated gene sequence. The sections below describe our methods of training, testing, and evaluating ANNs to identify virus and phage structural protein sequences without the direct use of sequence similarity. We chose to use ANNs because they have been successfully used in a multitude of problems involving pattern recognition and classification. Neural networks have been trained using different methods of encoding sequence data, such as sliding windows [36] or focusing on residues surrounding a catalytic site [18]. Due to a lack of sequence homology among all phage structural proteins, we represented our protein sequences in the most general way possible, by the percent composition of the 20 naturally occurring amino acids. To determine an optimal ANN architecture, we trained thousands of ANNs with varying training parameters, then assessed the optimized networks for their ability to correctly classify test cases using K-fold cross validation. The estimate of a network's accuracy in classifying data that was not used in training is known as generalization. We assessed an ensemble's ability to generalize phage structural proteins by specificity and sensitivity measures, which are based on the classification of a curated test set of phage sequences by an optimized voting scheme among ANNs. Sensitivity and specificity are commonly used, for example, to assess the performance of trained ANNs that recognize the biochemical markers associated with various forms of cancers [37]–[40]. Statistical measures have also been used to assess ANNs that were trained from sequence data to predict the presence of protein features, such as functional groups [41], secretory proteins [42], and protein functional domains [43]. In addition to testing our ANNs against phage structural protein sequences, we assessed network classifications of capsid and coat protein sequences from the genomes of viruses that infect archaea and eukarya. Lastly, we describe our method of experimentally validating ANN predictions of hypothetical proteins using transmission electron microscopy. Most of our validation results corroborate the predictions of our neural networks. Our aim was to recognize phage structural proteins by ANNs having minimum possible error rates, and to use this computational tool to predict the functions of unknown viral sequences. Training ANNs for high accuracy and to generalize patterns well is dependent upon many factors, such as data complexity, network architecture, and validation set size. The results that address these issues are described below. Our training strategy is summarized in Figure 1. For training the ANNs, we selected positive examples consisting of over 6000 phage structural protein sequences from GenBank's non-redundant database. An equal number of negative examples consisted of randomly chosen non-structural protein sequences from phage and prokaryotic genomes. The complexity of our data was highlighted by the diversity of annotated phage structural protein sequences in GenBank. For instance, we were unable to determine a consensus sequence from a single structural protein family, such as major capsid proteins. We chose to represent protein sequences by amino acid percent composition because ANNs trained by other encodings, such as the hydropathy index of individual amino acids, were not as successful [44]. We chose our parameters by comparing network performances trained with a range of parameter values, which follows a prescription for evolutionary programming [45]. For such tuning purposes, neural networks were trained and evaluated using 10-fold cross validation (explained further below). The ANNs with the highest mean accuracy were used to define our network architecture and to determine the best division of our data into training and validation sets. To increase performance even further, the resulting ANNs “voted” on the classification of ORFs, and we assessed various possible levels of voting. The total number of ANNs used for voting is 160 (see Cross Validation Partitioning in Materials and Methods) from which the optimum values of the training error, specificity, and sensitivity were then assessed using the best voting scheme against a curated test set of phage sequences that we manually labeled as structural or non-structural proteins. Our structural protein sequences came from genome sequences of organisms and viruses, which are summarized in Figure 2. The pie chart in the center of Figure 2 is divided into four slices that represent the sources of our protein sequences. We collected 6,303 protein sequences based on keyword searches against the non-redundant database (see Materials and Methods). Although we intended to focus on phage proteins, our positive training set contained 1001 proteins from over 300 phage genomes and 2,216 proteins from over 1,200 virus genomes. Among 2,603 proteins from 2,214 microbial, archaeal, or eukaryotic genomes (“Other” slice), 245 non-structural protein sequences came from eukaryotic genomes that inadvertently passed our filtering process because a keyword used to search for structural proteins was part of the name of a gene or organism. Although nearly all of the phage major capsid and tail protein sequences in our positive training set came from the genomes of tailed phages, our training set used to train structural protein neural networks contained the structural protein sequences from a variety of viral genomes (11 phage families and 81 virus families). Furthermore, our training set contains viral protein sequences from 7 archaeal, 277 eukaryotic, and 1,929 prokaryotic genomes. Seven archaeal structural protein sequences contained in our dataset are capsid portal (gi148552749) and minor tail (gi148552761) proteins from Methanobrevibacter smithii ATCC 35061, a tape measure protein (gi159885966) Methanococcus maripaludis C6, a head-tail adapter (gi170935066) Thermoproteus neutrophilus V24Sta, and a minor tail protein (gi118194002) from Cenarchaeum symbiosum A. To identify an optimum architecture, we examined the performance of a large number of neural networks that have between 1 and 100 neurons in one hidden layer, and between 1 and 30 neurons in a second hidden layer. To determine an optimum validation set size, we tested networks that were trained with sequences that were distributed differently between training and validation sets, as follows: 50∶50, 60∶40, 70∶30, 80∶20, and 95∶5, where the second number denotes the fraction of ORFs in the validation set. While no single architecture or ratio was found to be statistically different from others, the final set of voting ANNs were trained using the parameters that gave the best classifications of test cases. Initially, we trained ANNs using all structural protein types (including, for example, capsid proteins, tail proteins, tail fiber proteins, portal and connector proteins, etc.). Of all single and double hidden layer ANNs we tested, the ANNs with 20×90×1 topology (20 input neurons, 90 hidden layer neurons, one output neuron) correctly classified the greatest number of test cases, or 85.6% (left panel of Figure 3A). The validation set was used to decide when training should stop, i.e. when the classification error on the validation set remained the same or increased within 6 consecutive training iterations. We observed the highest mean accuracy, or 86.2%, from the fully trained ANNs when the distribution of sequence data into training and validation sets was 80/20, respectively. Figure 3A summarizes the performance of the resulting ANNs, which were assessed by 10-fold cross validation. K-fold cross validation refers to an evaluation process that splits a dataset into K disjoint subsets, each of which is used to train and evaluate an ANN's performance. The accuracy of a network is evaluated by test sequences that were not used for training. Networks with the optimal topology (20×90×1) and validation set size (20% of total sequences) had an average accuracy of 86.2% based on 160-fold cross validation. The resulting 160 networks were used to determine the optimal number of voting ANNs that gave the highest instance of accurate classifications of the curated test sequences, which are sequences that we manually labeled as structural or non-structural proteins. To test whether the performance of the structural ANNs could be improved by focusing the training on sub-classes of structural proteins, we trained ANNs to classify either major capsid proteins (MCP) or tail proteins (this training set included tail proteins as well as tail fibers, etc.). We also tested the effect of different ratios of positive and negative examples on ANN performance. All capsid and tail network architectures described in this section, however, were tested using data sets that contained equal numbers of positive and negative examples (1∶1 ratio), and one hidden layer of neurons. The performance of the trained ANNs was based on the average output of 10 voting ANNs. Figure 3B shows that MCP neural networks with 40 hidden layer neurons correctly classified the most test cases (91.3%). Tail neural networks with 10 hidden layer neurons correctly classified the most test cases (79.9%) (Figure 3C). The MCP and Tail networks with the highest accuracies were those trained with a training:validation set ratio of 70∶30. In addition to accuracy, we used sensitivity and specificity to measure neural network performance. Specificity and sensitivity may be used in different contexts, for example, to describe biochemical interactions between molecules or the performance of binary classifiers. The latter sense was used here, in which we classified phage protein sequences into two categories, positive and negative examples. The correct classification frequency, specificity, and sensitivity measures were calculated from trained ANNs that voted on 3,012 curated phage test sequences from 51 phage genomes. These genomes, listed in Table S2, were sequenced in 2010 and 2011, after our original data set was collected in 2009. Each of our voting ANNs correctly classified between 72% and 96% of all test cases, and each ANN voted independently. Based on the results described above, we used 20×90×1 ANNs that were trained with a validation set size containing approximately 20% of total sequences. Validation set sequences were not part of either the training or the test sets. An odd number of ANNs with the top mean correct classification frequencies from 160-fold cross validation were used to vote on curated phage test sequences. We also tested the correct classification frequency of all 160 ANNs. Voting results (Figure 4A) indicate that using the 5 most accurate ANNs increased the number of test cases that were correctly classified by nearly 2% (76.4%) versus our single most accurate ANN (74.5%). Similarly, the specificity of ensemble predictions increased by 2% over the specificity of the ANN with the highest accuracy. Moreover, the sensitivity of the top 5 ANNs was nearly 3% higher than the top single ANN. As expected, averaging the outputs of the voting ANNs produced very similar results to voting by a majority rule (data not shown). To visualize the performance of our networks we mapped the predictions of the Structural Protein ANNs against two phage genomes, T4 and T7, using CGView [46] (results are shown in Figure S1). These phage genomes were chosen because their genomes have been extensively studied and gene constructs that were used to validate network predictions come from two marine phages, φP-SSM2 and φMa-LMM01, that have T4-like genomes (below). The detection of structural proteins by our networks appears to be accurate; however, several proteins were missed by the networks: the “internal virion protein A” from T7, and two “prohead core” proteins, two “internal head” proteins, and a Soc small outer capsid proteins from T4 (Figure S1). Networks trained to detect MCPs correctly classified ∼90% of test cases. Tail ANNs accurately categorized ∼80% of positive (tail proteins) and negative sequences. Figure 5A shows that the Major Capsid Protein ANNs using a 1∶1 ratio of positive (major capsid protein) to negative examples correctly distinguished more capsid protein sequences from non-capsid protein sequences than did the Structural Protein ANNs (Figure 4B, MCP Test Data histograms) by as much as 15%. While the Tail ANNs showed only slight improvements in accuracy and specificity (Figure 5B) over the Structural Protein ANNs, the sensitivity of the Structural Protein networks to detect tail proteins was slightly higher (Figure 4B, Tail Test Data histograms). A marked improvement in the performance of Capsid and Tail ANNs was observed when isoelectric point information was added to the training data. The accuracy of the Capsid ANNs increased by as much as 5% in the ANNs that were trained with a 1∶1 ratio of positive to negative sequences. For the Capsid networks that were trained at ratios other than 1∶1, we observed 5–20% increases in ANN sensitivities. Isoelectric point information appears to have little effect on the specificity of the Capsid ANNs. The accuracy and specificity of the Tail networks are roughly the same, though slightly lower (1–2%) in the Tail PC + pI data, between the networks trained with and without isoelectric point values. The sensitivity of the Tail ANNs, however, was improved by as much as 15% in comparison to the Tail networks that were trained by amino acid frequency alone. We observed a trend in the sensitivity and specificity of the Capsid and Tail ANNs with respect to changes in the ratio of positive to negative examples in the training sets. As the ratio of negative to positive examples increased (X axes in Figure 5), the sensitivities of the networks decreased while the specificities increased. These trends can be explained by the frequency with which the networks classified true negative examples. Decreases in the positive to negative ratios in the training set produced networks that predict negative sequences at a higher rate. An increase in the number of negative ANN calls, however, increases the false negative frequencies of the networks, and hence decreases the overall network sensitivity (see Equation 1 in Materials and Methods). To test the accuracy of network predictions against capsid proteins from non-phage genomes, we collected additional capsid and coat protein sequences from the Reference sequence database. Table 1 summarizes the performance of our Major Capsid and Structural Protein ANNs. Protein sequences used for this round of testing did not overlap with any of our training sequences or test sequences from previous testings. We grouped protein sequences by the genome type from which each sequence came: double-stranded DNA, double-stranded RNA, single-stranded DNA, and single-stranded RNA. We also show capsid proteins sequences from archaeal viruses, which includes phage sequences to form a test set with as many sequences as possible. Also shown in Table 1 are the performances of our ANNs based on capsid sequences from phage genomes that were added to the Reference Sequence Database between February and May of 2011. The accuracy of predictions made by our Structural Protein neural networks were quite high (77–95% accuracy) for eukaryotic viruses and 73% for archaeal viruses in comparison to the performance of the Major Capsid 1∶1 ANNs (4–15%). Although there are very few capsid proteins from archaeal virus genomes in GenBank, our Structural Protein ANNs correctly classified 19 of 26, or roughly 3/4, of capsid protein sequences from archaeal viruses. The reliability of our neural network predictions at a 99% confidence level is shown in Figure S2 for different levels of network output, or threshold values. We observed an increase in the accuracies of our ensembles' true positive predictions as the threshold value increases. At the 99% confidence level, Structural protein ANN outputs of ≥0.6 correspond to 71–76% accurate predictions of true positive examples, while outputs of ≥0.9 correspond to 85–88% accurate predictions. The more specific MCP protein (1∶1) neural network scores of ≥0.6 correspond to 81–91% accurate predictions of true positive examples, as expected for networks that classify a more homogeneous sample of proteins. Due to the need to use a test set of proteins not “seen” before by the ANNs and the relatively few newly available tail proteins relative to the diversity of protein classes included in the Tail ANNs training set, we have not yet been able to calculate accurate confidence intervals in the case of these ANNs; this will be done as new phage genomes are entered in GenBank, and the information will be added to the iVIREONS web site that we are building as a web-based interface for the ANNs. In order to validate experimentally the predictions of the structural ANNs, we explored whether proteins predicted as structural could self-assemble into structures that resemble structural features of phages, using TEM (gray boxes, right side in Figure 6). We chose 16 genes whose functions were unknown from the genomes of two marine phages, φP-SSM2 and φMa-LMM01, one Bacillus phage, φIEBH, and Burkholderia φBcepC6B. We are aware that phage assembly is a highly ordered and complicated process, and that normally neither head nor tail structures are assembled from single, isolated proteins. We thus expected that relatively few of our test cases would in fact be able to self-assemble, presumably driven by a relatively high in vitro concentration of the single proteins. Moreover, we did not expect that a single protein would be able to self-assemble into the correct structure, as many proteins are necessary to regulate the correct assembly of the final structure of either phages or viruses. Nevertheless, for validation purposes, we considered that, even with the aforesaid limitations, self-assembly into a structure resembling the head or tail of a phage would be sufficiently indicative of the success of the ANN predictions. In addition to predicting the functions of unknown viral sequences, we were also interested in using ANNs to help detect prophages or prophage fragments present in bacterial chromosomes. We pre-processed and presented a few bacterial chromosomes to our ensemble of voting ANNs, which were each trained with a 1∶1 ratio of positive to negative sequences. Our training distribution, however, does not accurately represent the number of structural prophage genes in a bacterial chromosome. For example, in the case of E. coli MG1655 our data set would have to contain approximately 829 bacterial and prophage non-structural proteins for every prophage structural protein. (This approximation was obtained by dividing the total number of genes in MG1655 (4146) by the number of prophage genes that are annotated as structural proteins (5) in the MG1655 genome.) Although the ratio of prophage structural protein sequences to MG1655 sequences was not appropriately represented in our data sets, we investigated the ability of our voting ANNs to correctly classify sequences from the chromosomes of S. enterica LT2, S. aureus COL, and E. coli MG1655. Table S2 reports the number of prophage genes that were identified as structural proteins from the total number of prophage genes. The ANN predictions identified a number of genes that is within the expected range of structural genes relative to the size of a prophage genome. The percentage of prophage structural genes predicted by our ANNs from each bacterial genome, however, is 13% (347/2615), 19% (868/4423), and 18% (779/4145) of the S. aureus COL, S. enterica LT2, and E. coli MG1655 chromosomes, respectively. A list of ANN positive bacterial genes and average neural network outputs are listed in Supporting Dataset S1. Although all of the ANN predictions of all bacterial genes are given, it is important to keep in mind the thresholds at the 99% confidence interval (Supporting Figure S2) when interpreting the data. We investigated the classification of structural proteins based on conserved sequence motifs using the sequence analysis tool MEME [52]. We analyzed 120 sequences that were randomly chosen from our dataset of 757 major capsid proteins. The MEME analysis tool identified a pattern of motifs shared by only 16 major capsid sequences including a capsid sequence (gi 258545859) that was not detected by our Structural or Capsid ANNs. However, common patterns of motifs were discovered in <13% of capsid sequences we tested (data not shown). Thus MEME is not as sensitive as the ANNs at identifying members of structural, major capsid, or tail protein families, but may be useful in grouping the ANN-identified proteins into subfamilies that are more closely related. We will investigate the use of this tool further in the future. The analysis of genetic capacity of a microbiome is frequently performed using rRNA sequences as classifiers of different bacterial genera and species in the community. rRNA sequences are useful because they contain both highly conserved and highly variable regions. Thus metabolic capacity, determined by the actual genes encoded within a microbiome, can be related to a phylogenetic and taxonomic analysis of the cells present as indicated by rRNA sequences. In the case of viriomes, however, similar analyses are severely limited by the lack of any single gene that is shared among all viruses. In principle, structural components of viruses should fulfill a similar identification function to cellular rRNA sequences, but their sequences are simply too diverse. We aimed to design a computational tool that did not rely solely on sequence similarity in order to identify structural components of viruses, in this study of bacteriophages in particular. Herein we have described the training of feed-forward, back-propagation neural networks that classified phage protein sequences by amino acid percent compositions as well as, in the case of MCP and Tail ANNs, protein isoelectric point (pI). Each amino acid's functional group has its own characteristic pKa, and the overall pI of each protein can be estimated as a function of amino acid composition. It is possible that the accuracies of the ANNs benefitted from the addition of pI as an individual feature because this emphasized the charge aspect of the protein. The accuracies of our MCP ANNs increased with the inclusion of pI data in our training set, presumably because the pI of the capsid proteins lie in a relatively narrow range (pI 5–6), and because the pI distribution for MCP proteins differs significantly from the pI distribution of the proteins in the negative training set (Figure S3). In contrast to major capsid proteins used to train the MCP ANNs, the proteins used to train the Tail ANNs are more heterogeneous and the similarity between the bimodal distributions of the IEP values for the Tail protein positive and negative training sequences caused only a slight increase in the accuracies of the Tail, not comparable to that of the MCP ANNs. The function of a protein does not relate to a unique pI value, which was shown for proteins having the same function but did not have conserved cross-species pI values [53]. Thus, pI values were useful for improving the performances of our Capsid and Tail ANNs, but amino acid percent composition provided a much better signature for the function of a protein than pI values alone (data not shown). Although percent composition of amino acids and pI estimates are weakly tied to sequence similarity, the classification of viral sequences by these characteristics is much less dependent on sequence similarity than sequence alignments between nucleic acid or protein sequences. Our goals were two-fold. First, we investigated the potential of ANNs to recognize classes of virion structural proteins by training neural networks with sequences from prophages, proviruses, and the genomes of viruses that infect prokarya, eukarya, and archaea. Second, we examined the accuracy of predictions of networks that were trained exclusively on phage major capsid or tail proteins. We achieved our first goal with an ensemble of five voting networks that identify a broad spectrum of phage and viral structural proteins with ∼80% accuracy. We achieved our second goal with two ensembles of 10 ANNs each, whose specificity is as great or greater than the specificity of the structural neural networks at identifying phage capsid or tail proteins. In summary, we trained and evaluated thousands of networks from which our network ensembles have collectively classified a high percentage (∼80–95%) of test cases. Despite the lack of similarities in the sequence or predicted secondary structure of phage proteins across all structural protein families, the predictive accuracies of our trained neural networks were quite good at the 99% confidence level. Our trained Structural Protein neural networks have higher sensitivity than specificity, a condition that will detect more true positives but also more false positives than networks with high specificity and low sensitivity. Highly sensitive networks are ideal for identifying candidate sequences that are very diverged from known capsid or tail proteins sequences. For example, ORFs 29 and 30 from the Vibriophage VP16T phage genome (Figure S4) were identified as structural proteins by our neural networks and experimentally verified as structural proteins present in the phage lysate [54], but 8 years after their identification, there are still no similar sequences with known function. Our Structural Protein networks also correctly predicted a majority of capsid and coat protein sequences from the genomes of both phages and eukaryotic viruses, as well as a few from archaeal viruses, because our training set included protein sequences from a broad spectrum of virus genomes. In contrast, our Capsid and Tail ANNs have higher specificity than sensitivity. For example, we observed that 85% of all positive predictions made by our Capsid ANNs are indeed true positive capsid sequences, which is ideal when the goal of experimentally validating capsid protein sequences must be balanced by maintaining low experimental costs. Our capsid ANNs predicted capsid protein sequences from archaeal and eukaryotic viruses very poorly, which was expected because these ANNs were trained to recognize capsid proteins of bacteriophages. The specificity of the capsid ANNs for phage capsid proteins indicates that the frequencies of amino acids in phage capsid/coat proteins are inherently different from the frequencies in archaeal or eukaryotic virus capsids. Overall, our three network ensembles (Structural, MCP, and Tail) provided a means of detecting putative protein sequences that serve as virion structural components. Independent network predictions were used together to strengthen the predictions. Positive predictions of the Structural ANNs, for example, should produce similar predictions from the Capsid or Tail ANNs, and vice versa. A sequence that produces a positive output from the Structural ANNs and negative outputs from Capsid and Tail ANNs may be a structural protein that is neither a capsid nor a tail protein, but is either a structural protein such as a baseplate, tape measure, or portal protein, or a eukaryotic or archaeal virus capsid or tail protein. The former is probably the case for the sequences 5520-22 and 5524 from phage P-SSM2 (Figure 8B). An ORF that has positive outputs from Capsid or Tail ANNs but negative outputs from Structural ANNs may belong to a class of capsid or tail proteins whose sequences were represented poorly or not at all in the training set of the Structural ANNs, and thus not recognized as a structural protein, i.e. a false negative Structural ANN prediction. Another possible interpretation is that the Capsid or Tail ANNs made a false positive prediction. If experimentally validated, sequences that are positively identified by the Capsid or Tail ANNs but not by the Structural ANNs would be useful to re-train and improve the performance of the Structural ANNs. Structural protein sequences that have been predicted by our ensembles and experimentally verified will be used to re-train future versions of our networks. Network responses to all RefSeq test cases are listed in Supporting Dataset S2. By interpreting our network predictions in the manner just described, our ensembles were able to find nearly all the prophage structural proteins in the three bacterial chromosomes examined (Table S2). In cases where putative capsid or tail proteins are found in the context of a bacterial chromosome, i.e. as part of a prophage, the networks with greater specificity may be necessary since the positive ORFs will be found in the context of a preponderance of negative ORFs. Interestingly, 15–16% of all ORFs that were scored as ANN positive in the bacterial chromosomes of E. coli MG1655 and S. enterica LT2 were fimbrial, pilin, flagellar, and membrane proteins, which may share similar features present in some phage structural proteins. For example, the immunoglobulin (Ig) domain or fold that is found in phage tail-associated proteins, bacterial pili and fimbriae, and the bacterial type VI secretion system [55]–[60][10]. In addition to measuring the accuracy of our ANNs from test cases, we used our ANNs to predict the functions of proteins that have no known function, and used TEM to investigate the structures of hypothetical proteins that were expressed in vivo. Experimentally validating ANN predictions was challenging primarily because we attempted to assemble phage structures using a single protein, without the benefit of all the phage or bacterial accessory proteins that normally contribute to assembly. Despite the complexity of phage structures, at least some phage procapsids can assemble with just one or two proteins [61]. The assembly of the major capsid protein gp23 of T4 into polyheads occurs in simple buffer conditions [62]. Procapsids or polyheads form with just the Pb8p capsid protein of T5 phage, in certain buffers [63]. Here we experimentally validated, by TEM, a putative tail fiber gene (VCID5525) that was identified by both the Structural ANNs and Tail ANNs. The structures we observed were very similar to the structure of the T4 tail fiber, differing only in length, which may be due to the absence of phage accessory proteins. Despite the fact that T4 and P-SSM2 tail fiber genes are not similar in sequence, our TEM images suggest that the genes have the same function based on the similarity of their structures. Likewise, we have EM images showing procapsid-like structures in four protein samples. Images from 2 of our samples are nearly identical in morphology to the virion head structures of the known phages IEBH and BcepC6B (Figures 9 and 10). The procapsid and tail structures that were assembled are certainly not “finished” structures, although they are highly suggestive. It is very unlikely that images of self-assembled protein structures are those of inclusion bodies, because all of our purified proteins were soluble. These results strongly suggest that our ANNs are able to detect structural proteins that are otherwise not detectable by sequence similarity. Neural networks have been criticized as black boxes because the weights learned from the attributes of a data set are not easily deciphered; this also applies to other machine learning methods, such as support vector machine and Bayesian networks. Although some groups disfavor black box approaches for the reason just mentioned, we chose to use artificial neural networks for their ability to correctly classify our data and for the lack of good alternative options. We certainly expect that adding criteria such as the position of an ORF relative to genes encoding other structural components would improve the reliability of the predictions made by the ANNs. However, structural genes are frequently present in several distinct clusters within phage genomes, and the heterogeneity among phage genomes makes synteny a difficult parameter to encode for ANN training. Moreover, we wished to train ANNs that could be applied more generally in situations where ORFs are not genetically linked, such as is the case for metagenomic data where contigs frequently average no more than ∼1000 bp. Synteny information, when available, can be used in conjunction with the ANNs. In the future, we will also investigate whether the accuracy of our networks may be improved by the addition of sequence motif information, such as that explored by the MEME suite [52]. Perhaps results from MEME may also be useful for grouping ANN-predicted sequences into subclasses of structural proteins rather than during training, because MEME detects less general, more specific features than the ANNs. We have presented instances showing that our network predictions have been correct. The ANNs we have presented should serve as a building block to train other ANNs, or other machine learning methods, to accurately classify sequences from a broad range of protein families without direct dependence on sequence similarities to known sequences. This will be useful in identifying evolutionarily distant structural proteins that, if experimentally validated by (ideally) X-ray crystallography, will in turn increase the sensitivity of homology-based algorithms as well. The methods we used to train, test, and evaluate neural networks are summarized in Figure 1 and described in detail below. All neural networks were trained and tested using the Neural Network Toolbox 7.0 in Matlab version 7.6.0.324 (R2008a, The MathWorks, Natick, MA). All other computations and data manipulations were done with Java, UNIX shell utilities, and Perl and Bash scripts. Box plots were generated by Matlab and the R statistical package [64], and circular genome maps were created by CGView [46]. Isoelectric point estimates were calculated by BioPerl's pICalculator (http://doc.bioperl.org/releases/bioperl-1.4/Bio/Tools/pICalculator.html). Our neural networks are available to analyze translated coding sequences through the iVIREONS (identification of VIRions by Ensembles Of Neural networkS) web interface that is hosted at the SDSU Viral Dark Matter website (http://vdm.sdsu.edu/ivireons). We used the MEME suite of motif-based sequence analysis tools [52] to examine presence of motifs in our major capsid sequences. We uploaded 120 randomly selected major capsid sequences to the MEME web site (http: http://meme.nbcr.net/meme/cgi-bin/meme.cgi). We set the search parameters to look for a maximum of 20 motifs that are between 6 and 100 amino acids in length. We used default values for required MEME parameters and did not use optional MEME features. All sequences were obtained from NCBI by keyword search followed by several rounds of removing unwanted sequence by keyword searches through NCBI annotations. Positive sequences are proteins that have a target function, which the neural networks learns to distinguish from proteins that have other functions. A neural network that is trained to detect tail proteins, for example, was able to distinguish between tail protein sequences and protein sequences that do not function as tail proteins. The following sections describe the methods we used to gather positive sequences and negative sequences. All sequences are available in Datasets S3–S5. Training, test, and validation sets served different roles in the training and evaluation of neural networks. Sequences in the training set were used to calculate network errors, which were back-propagated throughout the network to update neuronal weights by Matlab's implementation of the Levenberg-Marquard learning algorithm, or trainlm. A validation set was used to stop the training process if the network performance fails to improve or remains the same for max fail consecutive epochs. Figure S5 shows an example of a training session that was stopped after the max fail stopping criteria were met. The max fail parameter was set to 6 by default. All sequences not used in the training and validation sets, were used as test cases, which were used to determine correct classification rates. To generate training and test sets from structural protein sequences, we used Matlab's cvpartition function (described in Cross Validation Partitioning). In other words, 90% of the original data set was randomly selected for the training set and the remaining 10% of the sequences was used for testing. A portion of the training data was reserved for the validation set, however, the optimum size of the validation set was uncertain. The percent distributions of training and validation set sequences we tested were 95/5, 80/20, 70/30, 60/40, and 50/50. The ANNs that were observed to have the best accuracy based on 10-fold cross validation determined the optimum ratio of sequences that were allocated into the training and validation sets. Due to the paucity of capsid and tail protein sequences in NCBI's RefSeq database we were unable to determine an optimum distribution of sequences by 10-fold cross validation. Instead, we used 10 ANNs to determine the accuracy of MCP and Tail ANNs in correctly classified test cases. Each of the 10 networks was trained using a different ratio of training and validation sequences that were randomly selected. The Structural Protein neural networks were trained using an equal number of positive and negative examples; however, we examined the accuracies of trained ANN using different ratios of negative to positive examples (1∶1, 2∶1, 3∶1, and 4∶1). In addition, we used the ratios of capsid to non-capsid and tail to non-tail proteins that are found in the genome of λ phage; these ratios are 22∶1 and 6.6∶1, respectively. The connection weights to neurons were adjusted by the Levenberg-Marquardt supervised learning algorithm [66], which has been implemented by the trainlm Matlab function. The LM algorithm has been used extensively for training neural networks [67] to iteratively update connection weights such that network error is minimized. The default parameter values set by the Matlab function newff were used. All neurons used the hyperbolic tangent, or tansig, squashing function. Matlab by default creates neural networks with 20 (amino acid frequency) or 21 (amino acid frequency with isoelectric point estimates) in the input layer and 1 neuron in the output layer, which is based on the structure of our data. To determine an appropriate Structural Protein ANN architecture we evaluated, by 10-fold cross validation, neural networks with 1 hidden layer that contained between 1 and 100 neurons. We also used 10-fold cross validation to evaluate networks with 2 hidden layers; we used between 1 and 100 neurons in the first hidden layer, and between 1 to 30 neurons in the second layer. Our goal was to find the smallest neural network architecture with the best accuracy of all the ANNs with these configurations. To determine the architectures of the Capsid and Tail ANNs we used the same method to train ANNs, as previously described with only a single layer of hidden neurons. Training and test sets were chosen by logical indices generated by Matlab's cvpartition function using the kfold option. The cvpartition function defines a random partition of K disjoint subsamples from N observations. For 160-fold cross validation, K = 160 was chosen because 160 is a number that divides the number of sequences in our data set (12,604 sequences) such that the size of each test set is equal to the average number of genes in a phage genome, or approximately 78 genes. The average number of phage genes was calculated by dividing the number of known phage coding sequences (46,754) by the number of known phage genomes (596), according to NCBI's Reference Sequence database in 2010. Hence, each test set consists of 78 disjoint subsamples from N>12,500 observations. For 10-fold cross validation, the default value of K = 10 is used. In comparing the performances of ANNs, all networks were trained using the same set of training, test, and validation sets. We chose to use voting ANNs because Hansen and Salamon [68] showed that the likelihood of error decreases with a majority decision if each network vote is independent and if each network produces a correct response more than half of the time. Trained ANNs from 160-fold cross validations were used to vote on curated phage test sequences, which were classified by averaging the outputs of all voting ANNs. The ANNs were trained by supervised learning to output a value of 1 for all phage structural protein sequences, and a −1 for all other sequences. Fully trained ANNs, however, will have a range of output values between −1 and 1. To better interpret intermediate values between −1 and 1, we impose the criteria that a positive (>0) mean output from the voting ANNs suggests that networks recognized the amino acid composition of a phage structural protein. Otherwise a negative (≤0) mean vote suggests that a protein sequence was not a phage structural protein sequence. To determine the optimal number of voting ANNs to use, we looked at the correct classification rates of voting ANNs to find an ensemble with the highest accuracy, specificity, and sensitivity. Only the ANNs with the highest correct classifications were used for voting and the sizes of the ensembles that we tested were 1, 5, 11, 21, 41, 61, 81, 101, 121, and 141 ANNs. We also assessed the performance of all 160 ANN responses to determine whether all ANNs are useful for classifying protein sequences. The small number of capsid and tail sequences prevented us from performing 160-fold cross validation, however, 10 voting ANNs were used to evaluate the accuracies of the Capsid and Tail networks. To evaluate the performances of the Structural Protein neural networks, we manually classified coding sequences from 51 phage genomes, which we used as a final test set to determine the performance of our neural network ensembles. The genomes were sequenced after our data set was collected in 2009 and hence, were not present in the training, test, or validation sets. A total of 64 phage genomes were sequenced in 2010 and 2011, however, 13 of the genomes were not used because they were poorly annotated. The genome names and accession numbers are listed in Table S1. The protein product description of each protein sequence was used to classify each ORF as a structural protein (positive example) or a nucleic acid modification enzyme (negative example). We also classified a sequence as a negative example if the function of the protein is annotated as an accessory protein for an enzyme that modified nucleic acids. All other protein sequences that were not annotated as an enzyme/accessory protein or structural protein were not considered. We classified a total of 3,012 sequences, 1,093 of which are structural proteins and 1,919 are non-structural protein sequences. To evaluate the performance of the Capsid and Tail neural networks, we used 59 capsid, 73 tail, and 507 negative sequences that were added to the Reference Sequence Database between February and May of 2011. Test sequences were added to the database after our training set was downloaded and therefore, our test set was not used in training. To make a fair comparison against the performance of the Structural Protein ANNs, the same sequences that were used to test the Capsid and Tail ANNs were also used to evaluate the performance of the Structural Protein ANNs. Similarly, the test sequences were not used to train the Structural Protein ANNs. Additional capsid and coat protein sequences from archaeal and eukaryotic genomes were collected from the RefSeq database (Dataset S6). These capsid and coat protein sequences were not used in the previous rounds of testing just described and used to test the Structural and Major Capsid Protein neural networks. Specificity and sensitivity may be used in different contexts, two of which describe interactions between molecules or the performance of a binary classifier. In a biochemical context, sensitivity describes the affinity of a molecule for its target, such as the recognition and binding of a sequence of DNA by a repressor protein [69], or the binding of an antibody to an antigen [70]. Specificity, on the other hand, describes the selectivity of a molecule for its target from among other potential or similar targets. Specificity and selectivity in the context of binary classifiers expresses the accuracy of a classifier's predictions, such as the accuracy of diagnostics in predicting disease states [71]–[75]. The task of predicting disease states is essentially a binary classification problem, which was the task of identifying structural from non-structural proteins. We used equations (1) and (2) to calculate sensitivity and specificity from all of our ANNs. Sensitivity measures the rate at which a neural network correctly classifies structural protein sequences, or the true positive rate. A structural protein sequence was identified by keywords in the sequence annotation, which strongly suggests that a protein is an integral part of the phage particle. Such keywords are tail, tail fiber, major and minor capsids, baseplate, tail sheath, tape measure, and collar. Specificity, or the neural network's true negative rate, is the percentage of non-structural protein sequences that are correctly identified. In relation to a structural protein, a true negative sequence encodes a protein that does not associate with, or is not known to be physically attached, to a phage particle.(1)(2) Confidence intervals were determined for all of the ANNs at the 99% confidence level using a bootstrap method. The pool of true positive sequences used was subsampled 1000 times, each time using a random 80% of the entire pool. Confidence intervals for Structural ANN output values were calculated at 0.1 intervals, ranging from 0 to 0.9. For each subsample, we calculated the accuracy of network predictions as described below for each class of networks. We determined confidence intervals from the minimum and maximum averaged network accuracies after excluding 0.5% of the highest and lowest accuracies. For the Structural ANNs, we used 1,093 true positive structural protein sequences from the curated test sequences described in the Curated Test Phage Sequences section. For the Capsid ANNs, we used 59 true positive sequences (this was limited by the number of new capsid sequences available that were not previously seen by the neural networks).

10.1371/journal.pbio.1001591

Transcription-Factor-Mediated DNA Looping Probed by High-Resolution, Single-Molecule Imaging in Live E. coli Cells

DNA looping mediated by transcription factors plays critical roles in prokaryotic gene regulation. The “genetic switch” of bacteriophage λ determines whether a prophage stays incorporated in the E. coli chromosome or enters the lytic cycle of phage propagation and cell lysis. Past studies have shown that long-range DNA interactions between the operator sequences OR and OL (separated by 2.3 kb), mediated by the λ repressor CI (accession number P03034), play key roles in regulating the λ switch. In vitro, it was demonstrated that DNA segments harboring the operator sequences formed loops in the presence of CI, but CI-mediated DNA looping has not been directly visualized in vivo, hindering a deep understanding of the corresponding dynamics in realistic cellular environments. We report a high-resolution, single-molecule imaging method to probe CI-mediated DNA looping in live E. coli cells. We labeled two DNA loci with differently colored fluorescent fusion proteins and tracked their separations in real time with ∼40 nm accuracy, enabling the first direct analysis of transcription-factor-mediated DNA looping in live cells. Combining looping measurements with measurements of CI expression levels in different operator mutants, we show quantitatively that DNA looping activates transcription and enhances repression. Further, we estimated the upper bound of the rate of conformational change from the unlooped to the looped state, and discuss how chromosome compaction may impact looping kinetics. Our results provide insights into transcription-factor-mediated DNA looping in a variety of operator and CI mutant backgrounds in vivo, and our methodology can be applied to a broad range of questions regarding chromosome conformations in prokaryotes and higher organisms.

One mechanism cells use to regulate gene expression is DNA looping, whereby two distant DNA sites are brought together by regulatory proteins. The looping then either enhances interactions between other regulatory proteins bound at the separate sites or brings those regulatory proteins close to RNA polymerase at the promoter. Recent work in bacteriophage λ has suggested that DNA looping mediated by a transcription factor called λ repressor CI plays a critical role in regulating the expression of λ genes and consequently in determining the fate of the host E. coli bacterial cells. CI-mediated DNA looping has been directly demonstrated in vitro, but it has only been indirectly inferred in vivo. For the current study we developed a method to visualize CI-mediated DNA looping in individual live E. coli cells. We labeled two DNA sites—one each side of the proposed loop—with differently colored fluorescent fusion proteins, allowing us to measure their separation with an accuracy of a few tens of nanometers. Using this method, we directly analyzed CI-mediated DNA looping, providing insight into how transcription factor-mediated DNA looping influences gene regulation in live E. coli cells. Our methodology can be applied to a broad range of questions regarding chromosome conformation in prokaryotes and higher organisms.

Looping between two DNA sites, mediated by transcription factors, is a ubiquitous mechanism in prokaryotic transcription regulation [1]. DNA looping brings two distal DNA sites into close proximity, enhancing interactions between transcription factors bound at separate sites or bringing transcription factors close to RNA polymerase at the promoter. Knowing when and how DNA loops in vivo is important to understand the role of DNA looping in gene regulation and cell decision-making; some studies found molecular details of gene regulation have little influence on gene expression [2]–[4], while others suggested that DNA looping could trigger cell phenotype switching [5] and influence fluctuations in transcription activity [6]. DNA looping was first suggested for the transcription factor AraC (accession number P0A9E0) in the E. coli arabinose operon. Disruption of an AraC binding site ∼280 bp upstream of the promoter reduced AraC-mediated repression nearly 10-fold, indicating a long-range interaction between the promoter and upstream DNA [7]. Subsequently, DNA looping mediated by transcription factors LacI [8] (accession number P03023), DeoR [9] (accession number P0ACK5), NtrC [10] (accession number P0AFB8), GalR [11] (accession number P03024), and bacteriophage λ repressor CI [12],[13] was reported. The length of the intervening DNA in these loops can be as short as 58 bp (lac operon [8]) or as long as ∼5 kilobases (deo operon [9]). Biochemical, biophysical, and genetic studies have established important roles of DNA looping in transcription regulation. However, transcription-factor-mediated DNA looping on the length scale of a few kilobases in prokaryotic cells has not been directly visualized in vivo, and the in vivo dynamics of DNA looping are difficult to investigate. Chromosome conformation capture (3C) has been used to detect juxtaposition of DNA sites separated by hundreds of kilobases in both eukaryotic and prokaryotic cells [14],[15], but high background of interactions at the kilobase scale limits the utility of these methods in studying typical prokaryotic DNA loops [16]. An in vivo imaging method using fluorescent proteins fused to DNA-binding proteins bound to tandem arrays of hundreds of binding sites has been employed to visualize homologous chromosome pairing in yeast induced by double-strand breaks [17]; however, an array of several kilobases of binding sites makes this method unsuitable for studying DNA loops of only a few kilobases. In addition, the long array of tightly bound protein molecules may be detrimental to cells [18]. We developed a two-color, high-resolution imaging method to directly measure the end-to-end separation of two DNA sites 2.3 kb apart in live E. coli cells (Figure 1a). This method is based on the ability to precisely determine the location of a specific DNA site in vivo [19]. By expressing a fluorescent protein in fusion with a DNA-binding protein in a cell with only three tandem binding sites (spanning less than 100 bp), the resulting fluorescent spot is diffraction-limited, and the location of the binding site can be determined with sub-diffraction-limited precision by fitting its fluorescence profile to a two-dimensional Gaussian function [20]. By labeling two ends of a DNA segment with two unique sets of binding sequences and co-expressing corresponding fluorescent DNA-binding fusion proteins of different colors, the distance between the two DNA sites can be determined with a precision of a few tens of nanometers. An in vitro experiment employing the same principle measured intramolecular distances using organic dyes [21], but this approach has not been demonstrated in vivo with comparable resolution using fluorescent proteins. We used our method to probe the mechanisms and dynamics of DNA looping mediated by the bacteriophage λ repressor CI [22] in live E. coli cells and investigate its regulation of transcription from the CI promoter PRM. The λ repressor CI is an essential transcription factor in determining the fate of an E. coli cell infected by the bacteriophage λ. When CI is expressed, it represses lytic promoters to commit to an extraordinarily stable lysogenic state that persists for millions of generations [23]–[25]. However, upon induction by UV irradiation or other specific events, CI degradation can trigger an irreversible switch from lysogenic to lytic gene expression within one cell generation time [26]. The robustness of the λ regulatory circuit has been extensively studied. Among many important features of the system such as promoter-operator arrangement [27],[28], CI autoregulation [3],[29],[30], and cooperative binding [31]–[34], DNA looping between the homologous rightward and leftward operators OR and OL, separated by 2.3 kb, was shown to play significant, fate-determining roles in the λ lifecycle [13],[35]. Cooperative binding of CI dimers at the subsites OR1 and OR2 of OR represses the lytic promoter PR (reviewed in [36]) and simultaneously activates CI's own promoter, PRM, by accelerating transcription initiation [37]–[39]. At higher CI concentrations, an additional CI dimer binds to OR3 and represses PRM [40]. As illustrated in Figure 1a, an octameric CI complex (with or without an additional CI tetramer) can mediate DNA looping by bridging OR and OL. These higher-order complexes result from interactions between CI dimers bound to subsites at OR123 and OL123, and were first identified in vitro by ultracentrifugation [41] and later visualized by EM [12] and AFM [42]. Looping dynamics were investigated in vitro using tethered particle motion (TPM) [43]–[46]. To gain quantitative insight into the relationship between CI-mediated DNA looping and transcription regulation, thermodynamic models and numerical simulations were developed [33],[35],[44],[47]–[52]. Key parameters in these studies were the free energies of octameric and tetrameric CI interactions that mediate DNA looping [35]. These free energies specify the DNA looping probability at a given condition (temperature, CI concentration, etc.) and hence the extent to which distal DNA sites affect each other. To date, DNA-looping probabilities and free energies were either estimated indirectly in in vivo studies by measuring PRM and PR activities in various operator mutants with a priori assumptions of DNA looping states [35],[49],[51] or measured using purified components in vitro, where conditions differ from those in a cellular environment [42]–[46]. Consequently, these studies yielded varying estimates for the free energies of DNA looping and the degree to which DNA looping influences PRM activity. Hence, the roles of CI-mediated DNA looping in transcription regulation are still in debate [13],[35],[49],[51],[53]. In this study, we tracked the apparent separation between the OR and OL sites on a λ DNA segment (termed OR–OL DNA below) in real time in live E. coli cells, from which we obtained the first direct estimates of in vivo looping frequencies and kinetics for both wild-type DNA and for DNA carrying mutations in OR3 and OL3. We also measured corresponding CI expression levels in these strains by counting the number of CI transcripts in individual cells. Applying these independent, in vivo measurements to a thermodynamic model, we were able to obtain looping free energies and quantify the influence of DNA looping on PRM expression. Furthermore, we discuss how the compaction of the E. coli chromosome may impact DNA looping kinetics. The methodology established in this work can be extended to a broad range of questions regarding chromosomal DNA conformation and/or gene activities in prokaryotes and higher organisms. We inserted the construct shown in Figure 1a into the E. coli chromosome. It contains three tandem tetO sites (tetO3) [54] and three tandem lacOsym sites (lacO3) [55] flanking the wild-type λ lysogen sequence from OR to OL (including the PR, PRM and PL promoters and the cI, rexA (accession number P68924) and rexB (accession number P03759) genes). In this construct, called λWT, CI is expressed from PRM and regulates its own expression. The lacO-binding and tetO-binding proteins LacI and TetR (accession number P04483) were fused with red and yellow fluorescent proteins to generate LacI-mCherry and TetR-EYFP, and were expressed from an inducible plasmid (Figure 1b). With the combination of strong induction, weak ribosome binding sites, and carefully controlled growth, we achieved sufficiently low LacI-mCherry and TetR-EYFP expression levels to detect distinct, diffraction-limited mCherry and EYFP spots in single cells. We then fit the fluorescence intensity profile of each individual spot with a two-dimensional Gaussian function to estimate its centroid position. The average localization precisions for individual spots of LacI-mCherry and TetR-EYFP were 17 and 14 nm, respectively (Figure S1a). Subsequently, we transformed EYFP coordinates into mCherry coordinates using fiducial data to calculate the vector between the mCherry and EYFP spots arising from LacI-mCherry and TetR-EYFP protein molecules bound to the same OR–OL DNA segment. We called this vector (Figure 1c). The magnitude of the vector, , is the two-dimensional projection of the distance between lacO3 and tetO3 onto the image plane; on average, it is proportional to the end-to-end distance between lacO3 and tetO3 in three dimensions. The total error for an measurement, including fitting errors in determining centroid of individual spots (Figure S1a), registration errors in aligning EYFP and mCherry two-color images (∼10 nm based upon experiments using fluorescent beads), and contributions from local fluorescent background, was on average ∼40 nm (see below). With very low TetR-EYFP and LacI-mCherry expression, it was inevitable that not all lacO3 and tetO3 sites were bound by fusion protein molecules. Furthermore, not all fusion protein molecules were fluorescent due to stochastic chromophore maturation. Figure 2a contains typical data showing that a subset of cells was successfully labeled at both sites. We analyzed all cells having distinct fluorescent spots in both emission channels to calculate . We expected to decrease when DNA between lacO3 and tetO3 looped. To determine whether our two-color imaging method was sufficient to distinguish between looped and unlooped DNA in the crowded intracellular environment, we constructed two control strains (Table 1). In the positive control λnull, the centers of lacO3 and tetO3 sites are separated by 66 bp (Figure 1d). The outmost lacOsym and tetO sites are separated by less than 40 nm (Figure S2a). The close proximity of lacO3 and tetO3 mimicked permanently looped DNA. In the negative control λΔOL, we inserted the λ sequence from OR up to but not including OL between lacO3 and tetO3 (Figure 1e). The resulting λΔOL DNA has comparable length as the wild-type λ DNA, but CI-mediated DNA looping between OR and OL is abolished. We first examined λnull and λΔOL in two-color fluorescence images to determine whether we could discriminate between looped and unlooped DNA by eye. We obtained at least sixty 20-frame movies (100 ms exposures; 2 s total) for each strain in each of three independent experiments. Typical fluorescence images are shown in Figure 2a and b. Crosstalk between the two emission channels was negligible, as bright mCherry and EYFP spots only appeared in the corresponding channel but not the other. Figure 2c and d show 1 s of typical data for individual λnull and λΔOL spots. Representative movies for the two strains and others discussed below are included as Movies S1, S2, S3, S4, S5, S6. As expected for a permanently looped configuration, the positive control λnull exhibited overlapping EYFP and mCherry spots (Figure 2c). Generally, λΔOL molecules did not exhibit spot separation that was easily identifiable by eye (Figure 2d). However, some λΔOL molecules displayed large displacements between the LacI-mCherry and TetR-EYFP spots that were distinguishable by eye (Figure 2e); such images were not observed for λnull. Visual inspection of the apparent separation between the LacI-mCherry and TetR-EYFP spots suggested that comparing the end-to-end separation in OR–OL DNAs required a more quantitative approach. We calculated for all OR–OL DNA molecules in the λΔOL and λnull strains that exhibited fluorescent spots in both EYFP and mCherry images. Figure 2f–h shows calculations for movies in Figure 2c–e, respectively, and Figure S3 shows vectors for all movies lasting 0.8 s or longer. We then compiled the corresponding probability density distributions (PDF, , Figure 3a) and cumulative density distributions (CDF, , Figure 3b) of the vector magnitude, . The long-tailed PDF observed for λnull (Figure 3a) is consistent with the expected end-to-end distance distribution measured from two spots with a fixed separation when the localization of each spot is subject to Gaussian fitting error [56]. A simple numerical simulation of the end-to-end distance PDF for two sites separated by 22 nm and each subject to 22-nm localization error largely recapitulates the long-tailed distribution (Figure S2c). We found that the distribution for λΔOL was distinctly different from that of λnull (p<10−3); the difference was reproduced in three independent experiments (Figure S1b). The mean separations, , were 47 (N = 1,153) and 71 nm (N = 979) for λnull and λΔOL respectively (results and measurement errors summarized in Table 2). Peaks in plots centered at ∼40 nm, reflecting our experimental precision in determining ; that is, OR–OL molecules with below 40 nm could not be distinguished from each other. Hence, it was more meaningful to compare distributions of at large values where distributions differed most prominently. The cumulative probability of being 75 nm or more was ∼40% for λΔOL and only ∼15% for λnull (Figure 3b). Furthermore, two-dimensional distributions of vectors (Figure S4) were clearly wider for λΔOL than for λnull. Thus, by examining distributions, we could distinguish between the looped and unlooped control strains, suggesting that this approach could be used to probe CI-mediated DNA looping. We measured the mean end-to-end separation for λΔOL at 71-nm, much shorter than the ∼200-nm distance expected for B-form DNA with a typical 50-nm in vitro persistence length [57]. While such a result is expected given the many factors known to compact prokaryotic chromosomes [58], it is possible that nonspecifically bound CI on the λΔOL DNA and/or PRM transcription activity could influence the distribution, as indicated by a series of recent studies in vitro and in higher eukaryotic systems [46],[59],[60]. To examine these possibilities, we first compared the distribution of the λΔOL strain to that of a control strain λΔOLPRM−cI−/cItrans (Table 1, Figure S5a and b). In this control strain, promoter PRM was mutated to abolish transcription and the cI start codon was eliminated, but CI binding to OR was unaffected (Figure S5c, d, and e). In addition, we expressed CI from a plasmid at ∼9 times its level in λWT (Table S8). We found that the distributions of the λΔOL and λΔOLPRM−cI−/cItrans strains were indistinguishable (Figure S5a and b), demonstrating that the compact λΔOL distribution does not depend on PRM transcription. Furthermore, distributions for the same λΔOLPRM−cI− strain with or without the CI-expressing plasmid were indistinguishable (Figure S5a and b), suggesting that nonspecifically bound CI did not interact with specifically bound CI at OR operator sites to condense DNA in vivo [46]. We next investigated DNA looping in the context of wild-type and mutant OR–OL DNAs. In λWT, the wild-type λ sequence from OR through OL was inserted between lacO3 and tetO3. CI could bind all OR and OL sites to mediate looping with both octameric and tetrameric CI complexes (Figure 1a). In λOR3− and λOL3−, mutations in OR3 and OL3 essentially eliminated CI binding to these operators at lysogenic CI concentrations (Table 1) [35],[61]. We measured for these three strains and found that distributions differed significantly from those of the positive and negative controls λnull and λΔOL (p<10−3, except p = 0.004 for λWT and λnull), with and being intermediate to those of the controls (Figure 3c and d). Mean values for the three strains also fell in between those of λnull and λΔOL (Table 2). The wild-type strain had lower than λOR3− and λOL3−, and its distribution differed from those of the mutant strains with moderate to high significance (p = 0.001 and 0.048 for λOR3− and λOL3−, respectively); distributions for λOR3− and λOL3− were indistinguishable from each other (p = 0.493). The trend of λnull<λWT<λOR3−≈λOL3−<λΔOL for was reproduced in three independent experiments (Figure S1b). Assuming that a DNA molecule in the λWT, λOR3−, and λOL3− strains is in either a looped or unlooped state, the intermediate values of the three strains suggested that the fraction of looped DNA molecules (herein termed looping frequency) could be estimated by comparing distributions of these strains to those of the looped and unlooped controls λnull and λΔOL. To further investigate whether the observed DNA looping in the λWT, λOR3−, and λOL3− strains could be abolished by eliminating CI cooperative binding rather than by deleting OL, we constructed a control strain λCIG147D (Table 1). This strain differs from λWT by a CI mutation G147D known to be defective in pairwise cooperative interaction [62],[63]. Structural evidence suggests that cooperative binding interfaces are shared for pairwise binding to adjacent operator sites and the formation of CI tetramers or octamers via DNA loops [64]. We found that the distribution of the λCIG147D strain was indistinguishable from that of λΔOL (Figure S5f and g, Table S7). We note that this G147D mutant also diminishes PRM transcription because of its weakened ability to form a CI tetramer at the OR1 and OR2 sites; hence its expression level is lower than that with wild-type CI (Table S8). Therefore, we constructed another control strain (λCIG147D/cIG147D,trans), in which the CIG147D mutant protein was expressed constitutively at ∼11 times the CI expression level in λWT from a plasmid transformed into the λCIG147D strain (Table S8). We found that distribution of this strain was indistinguishable from that of the λΔOL and the λCIG147D strains, demonstrating that DNA looping could be abolished by eliminating CI cooperative binding. To quantitatively examine how operator mutations influence DNA looping, we estimated looping frequencies for λWT, λOR3−, and λOL3− by assuming a simple model. In this model, DNA molecule can only exist in one of two states, looped or unlooped, with distributions for each state resembling those of the looped and unlooped controls, λnull and λΔOL, respectively. Therefore, the distribution or for one of the three strains is the linear combination of that of λnull and λΔOL, with their distributions weighted by the looping frequency, :Using this model, we found that the looping frequency was 79% for λWT, and reduced to 53% for λOR3− and 60% for λOL3− (results with errors summarized in Table 2). The results were indistinguishable within error regardless of whether cumulative or probability density distributions were used, or whether data points from all frames or only the first frame of each molecule's movie were used (Table S1). The looping frequencies for λOR3− and λOL3− were indistinguishable from each other within error, suggesting a similar role of OR3 and OL3 in loop formation. Reduced looping frequencies of λOR3− and λOL3− compared to λWT suggest that while a CI octamer at OR12 and OL12 is sufficient to loop DNA, the resulting loop can be further stabilized by an additional CI tetramer only if both OR3 and OL3 are intact. To our knowledge, these measurements provide the first quantitative in vivo estimates of DNA looping frequencies that are independent of gene regulation models. In the above analyses, we only utilized , the magnitude of the vector, and discarded information about the direction of and its evolution in time. Looping frequencies estimated from distributions are analogous to equilibrium constants and lack kinetic information. While many DNA molecules only exhibited fluorescent spots in both EYFP and mCherry channels for one or two consecutive frames due to photobleaching, some molecules had fluorescent spots lasting for several consecutive frames in both channels (Figure 2c–h; also see plots from molecules with many frames in Figure S3). By analyzing how evolves in time, we can obtain additional information about DNA looping kinetics. We calculated the autocorrelation of (the average dot product of two vectors separated by a time lag) up to 0.5 s for each strain using all movies in which fluorescent spots in both channels lasted two or more frames (Figure 4a). The autocorrelation curves of all strains showed an initial drop of ∼2,500 nm2 at the first time lag, corresponding to uncorrelated errors in determining . After the initial drops, all autocorrelation curves showed positive correlation values that were approximately constant at time lags up to 0.5 s. The observation of near constant autocorrelation values after the first time lag for all the strains indicated that the conformation of each DNA molecule, characterized by both the magnitude and orientation of , persisted for at least 0.5 s. This provides a lower limit for the amount of time it takes for two DNA sites in the relaxed, unlooped state to move relative to each other and potentially form a DNA loop, and thus an upper limit of ∼2 s−1 for the rate of DNA looping. The plateau values are related to the averaged mean end-to-end separations—λΔOL has the highest autocorrelation plateau and λWT, λOR3−, and λOL3− have intermediate values because they contain a mixture of looped and unlooped DNA molecules. Next, we measured average CI expression levels, , in all strains in order to understand to what different extent DNA looping influences PRM regulation. We used single-molecule fluorescence in situ hybridization (smFISH, [2],[65],[66]), in which multiple fluorescently labeled oligonucleotides probe targeted nonoverlapping regions of cI mRNA, to count the number of PRM transcripts in individual cells (Figure 4b and c). Given the assumption that the average number of CI molecules translated per PRM transcript is the same in all strains and the observation of indistinguishable cell growth rates (Figure S6a and b), we expected average mRNA expression levels proportional to . The λnull strain does not contain the cI gene and was used as a negative control. All other strains were transcriptionally active. Under our experimental conditions, the false positive rate using the λnull strain was ∼1 transcript per 50 cells, two orders of magnitude below the levels of all other strains; false positives arise when nonspecifically bound probes occasionally co-localize to create a fluorescent spot above the detection threshold. Typical smFISH images of the five strains are shown in Figure 4b. We quantified the number of transcripts in each individual cell by dividing the total intensity of fluorescent spots in each cell by the average intensity of a single-transcript spot (Figure 4c). We then determined in wild-type λ units (WLU) by dividing the average number of transcripts in cells of a given strain by the average number of transcripts in λWT cells. We found that deleting OL increased to ∼1.4 WLU (Table 2), indicating that the DNA loop formed between OL and OR in λWT enhances PRM repression. Mutating either OR3 or OL3 further increased to ∼2.5 WLU. These observations are consistent with previous observations that although OL3 is 2.3 kb away from the PRM promoter, it has as important a role as OR3 in repressing PRM at lysogenic CI concentrations [13]. This suggests that PRM was not strongly repressed by CI binding to OR3 in the absence of a tetrameric interaction with an additional dimer at OL3. Finally, elevated in λOL3− relative to λΔOL indicated that DNA looping could also activate PRM, which was likely mediated by the binding of a CI octamer at OL12 and OR12, and was consistent with recent in vivo [49],[51] and in vitro [53] experiments. We have shown that reduced looping frequencies in λOL3− and λOR3− compared to that in λWT corresponded to increased expression levels of CI in the two strains, and that unlooped λΔOL has a higher expression level than the λWT strain. To establish a quantitative framework that explains all observed relationships between looping and CI expression levels, we refined a thermodynamic model, with which we estimated looping free energies and the degree to which DNA looping changes the activity of PRM. These parameters are important because free energies describe the likelihood of interaction between two distal DNA sites, and changes in promoter activity directly reflect the influence of DNA looping on gene regulation. The thermodynamic approach was first applied to model repression and activation of PRM by CI bound to OR [52] and recently modified to address looping [35],[44],[49],[51]. Our modeling approach is unique in that we used two independent, in vivo measurements, looping frequencies, and corresponding CI expression levels, to refine parameters for DNA-looping free energies and transcription activities. In previous modeling work, DNA-looping free energies were either inferred from PRM and PR expression-level measurements [35],[49],[51] or estimated using in vitro data [44]. The thermodynamic model and fixed physical parameters from previous reports we used to estimate PRM expression levels and DNA looping frequencies are essentially identical to the one used to analyze in vivo gene expression experiments [35]. Briefly, we assume that DNA states can be enumerated, that steady-state, in vitro DNA-binding measurements are applicable in vivo, and that mean expression rate, , equals the sum of all products , where is the transcription rate in a particular state and is the probability of the state at a given concentration of free CI dimers :Each state is defined by its free energy, , the number of bound CI dimers, , and the degeneracy, , which is the number of states with the same , , and . The model is described in greater detail in the Materials and Methods section; all states considered are listed in Table S2. is normalized by the partition function, , so that the sum of all state probabilities is 1. Following earlier work [49] and considering that the CI-mediated loop is relatively long, we assumed looping free energies to be independent of parallel or antiparallel orientation. Note that loop orientation is important in shorter DNA loops such as those mediated by Gal repressor [67]. We approximated the average CI concentration, , as the concentration at which the degradation rate equaled the production rate. We refined our model to fit seven experimental observables: CI expression levels for λΔOL, λWT, λOR3−, and λOL3−, and the looping frequencies for λWT, λOR3−, and λOL3−. We varied four free parameters: the free energies of forming a CI octamer and tetramer in the DNA loop as defined by Dodd et al. [35], , and , and the PRM expression rates when OR12 is bound by CI and DNA is either looped () or unlooped (). is the free energy of bringing together OR and OL when both are bound by two adjacent CI dimers to form a CI octamer, resulting in a looped conformation. is the free energy of adding a CI tetramer to a loop already secured by a CI octamer. All other free energies and parameters such as specific and nonspecific DNA binding of CI were fixed at the values used by Dodd et al. [35]. The wild-type CI concentration was fixed to 220 nM (∼150 molecules/cell) based upon our previous experiment in which CI molecules were counted at the single-molecule level in a similar strain at similar growth conditions [3]. The CI degradation rate was fixed to give a half-life equal to the observed 2-h doubling time in our experiments. The four free parameters were adjusted to best fit our experimental measurements of looping frequencies and CI expression levels. Modeled looping frequencies and CI expression rates at different CI concentrations are shown in Figure 5a and b. The best fit estimated and at 0.3 and −3.2 kcal/mol, respectively, and the CI expression rates at 1.9 nM/s and 4.5 nM/s for unlooped () and looped () DNA when CI binds OR12. These results suggest that the DNA looping mediated by only a CI octamer is not strongly favored, while looping mediated by both an octamer and tetramer is the dominant configuration if all six binding sites are bound by CI dimers. Note that a small, positive is consistent with measured looping frequencies greater than 50% for ΔλOL3− and ΔλOR3−, as one unlooped configuration could lead to multiple looped configurations (Table S2). The higher CI expression rate from the looped configuration suggests that, in the absence of OR3 binding, bringing the distal OL and OR sites together to form a DNA loop activates PRM to 2.4 times the unlooped level. To test how sensitive the fitting results were to two fixed parameters that are poorly defined in previous work, we varied CI expression levels and nonspecific DNA binding affinity. We found that across the examined ranges, octameric looping energies, , were consistently near 0 and tetrameric looping energies, , were strongly favorable between −2.8 to −4.6 kcal/mol (Table S3). Similarly, CI expression rates and remained close to the original fit values, giving activation ratios between 1.7 and 2.5 (Table S3). We also verified that our fit parameters were unique—as shown in Figure 5c and d, the values of fit parameters corresponded to a well-defined minimum in the sum of squared residuals in the four-dimensional (two free energies and two expression rates) parameter space (Figure 5c and d). Hence we conclude that the four fit parameters resulted from the model were robust and well defined. In this work, we directly measure the end-to-end separation between two DNA sites separated by only 2.3 kb on the E. coli chromosome with high spatial resolution, and report the first estimates of CI-mediated DNA looping frequencies in live E. coli cells. We improved a thermodynamic model to estimate the free energies of DNA looping as well as the degree to which DNA looping enhances transcription regulation. Combining independent, single-molecule measurements of looping frequencies and CI expression levels increased confidence in this model. Our results provide insight into transcription-factor-mediated DNA looping in vivo, and the new method reported here also has the potential to address questions beyond DNA looping, including understanding of chromosome structure and dynamics in vivo. In the following, we compare our results with previous work, and discuss unique information provided by our new method. Our estimated looping frequencies of 79% for λWT and greater than 50% for λOR3− and λOL3− are larger than those observed in vitro by TPM and AFM, where looping frequencies at lysogenic CI concentrations were approximately 60% with wild-type operators and 10%–40% in the absence of OR3 and OL3 [42],[44],[46]. As looping frequency is directly linked to looping free energy, comparison of values showed the same trend: values estimated in these in vitro experiments were similar to our estimate of −3.2 kcal/mol, while in vitro values were 1–2 kcal/mol higher than ours [44],[46]. Significantly different values likely resulted from differences between naked DNA in an in vitro environment and the compact, protein-decorated E. coli chromosome in the crowded cellular environment. Factors such as supercoiling and nonspecific, “histone-like” DNA-binding proteins could compact DNA and lead to more frequent encounters between OR and OL. Our observation that the unlooped λΔOL DNA was extremely compact (discussed in more detail below) was consistent with this view; this level of compaction (comparable to a polymer with a 3-nm rather than a 50-nm persistence length) could lead to a 50-fold increase in the rate at which OR and OL encounter each other [68]. The relatively unchanged values could reflect the fact that the entropic and energetic costs of bringing OR and OL together are included in . Our looping frequency estimates confirm what were predicted by in vivo gene expression experiments—DNA was estimated to loop ∼72% of the time for wild-type OR–OL DNA and ∼69% for DNAs similar to our λOR3− and λOL3− constructs [35]. Correspondingly, the and estimated in the in vivo work (−0.5 and −3.0 kcal/mol) [35] compared well to ours (0.3 and −3.2 kcal/mol). One important assumption we employed in calculating looping frequencies is that that looped and unlooped λWT, λOR3−, and λOL3− DNA molecules had similar distributions to those of the looped control λnull and unlooped control λΔOL, respectively. It is possible that the unlooped states in the λWT, λOR3−, and λOL3− strains were more compact than that in λΔOL if after a DNA loop breaks OR–OL DNA does not always completely relax before it reforms again. In such a case, looping frequencies estimated using the linear-combination model would be upper limits on the true looping frequencies. Nevertheless, as we show above, our looping frequency estimates broadly agree with expectations from previous studies. Since this simple model only requires one free parameter and gives reasonable results, it is unnecessary to invoke more complicated models. By comparing looping frequencies and corresponding CI expression levels in λWT, λΔOL, λOR3−, and λOL3−, we showed that loop stabilization by the CI tetramer between OR3 and OL3 is important for efficient PRM repression, and that looping mediated by a CI octamer at OR1 and OR2 is important for PRM activation. We note that while it is possible that the presence of tetO3 and lacO3 binding sites flanking OR–OL DNA may influence CI binding and/or transcription, this influence is negligible. This is because CI expression levels in these strains measured using smFISH are comparable to that of a wild-type λ lysogen (Table S8), and our results are consistent with previous observations [13],[49],[51],[53]. Furthermore, results are directly comparable as all strains used in this study are identical with respect to the presence and positioning of these binding sites. Combining these results in our thermodynamic model, we estimated that CI-mediated DNA looping activates PRM to 2.4 times its level when the DNA does not loop. This compares well to earlier estimates of 2–4 fold [49], and 1.6-fold for a high-expression PRM mutant [53]. Another study did not find looping activates transcription, modeling CI-concentration-dependent PR and PRM activities without invoking activation via looping (by assuming ) [35]. A later study indicated that this discrepancy may have resulted from different constructs used in the earlier study [49]. The molecular basis for DNA loop-enhanced PRM activation is unclear. One possibility is that a CI dimer bound to OR2 interacts with RNA polymerase to a greater extent if it is part of a higher-order CI octamer [53]. Alternatively, a recent work showed that a DNA UP element proximal to OL [49],[69] enhances CI expression from PRM in looped DNA by contacting the α-C-terminal domain of RNA polymerase [51]. The activation mechanism could be clarified in future experiments measuring both looping frequency and PRM activity while varying operator and UP element sequences and introducing CI mutations affecting operator binding, oligomerization, and RNA polymerase interaction. We estimated the time scale a DNA molecule stays in a particular state by calculating the autocorrelation function of the vector (Figure 4a). The vector was strongly correlated for at least 0.5 s, suggesting that a particular DNA conformational state, either compact or extended, persisted for at least 0.5 s. This implies an upper limit of 2 s−1 for the rate of loop formation from the extended state. This upper bound of transition rate is in the range of what was observed in a previous TPM experiment, in which looped and unlooped states lasted for tens of seconds [44], and argues against a significantly faster rate used in a recent computer simulation (∼60 s−1) [50]. We note that although it is possible that transient CI unbinding does not necessarily lead to immediate and complete DNA conformational relaxation at our measurement time scale, the autocorrelation analysis puts an upper limit for the true transition rate between the looped and unlooped states. The same concern also applies to in vivo 3C and in vitro TPM experiments. Slow transitions between looped and unlooped states imply that low or high expression states resulting from a particular DNA conformation could be long-lived, potentially committing a cell to a particular fate. Supporting this is a recent study that suggested that a single unlooping event could trigger induction of the lac operon [5]. We were unable to obtain time trajectories long enough to clearly identify looped/unlooped transitions for single DNA molecules. Development of brighter, faster maturing, and more photostable fluorescent proteins or in vivo labeling with synthetic fluorophores [70],[71] will help in increasing the number of measurements made on one DNA molecule, possibly enabling accurate measurement of DNA looping kinetics in vivo. We observed very small end-to-end separation for the unlooped control ( = 71 nm). This distance was shorter than expected from modeling the unlooped DNA as a noninteracting worm-chain with an in vitro persistence length of 50 nm [72], but consistent with the recently observed extreme bendability of short DNA molecules [73]. A noninteracting chain with an equivalent to that of λΔOL would have a persistence length of only 3 nm, which is physically infeasible. Our measurements of indistinguishable conformational distributions in the absence of PRM transcription and the presence of CI overexpression suggest that neither transcription nor nonspecifically bound CI played a major DNA-compacting role in our experiments. Furthermore, C. crescentus chromosomal DNA segments of ∼5 kb were found to be similarly compact and consistent with Brownian dynamics simulations of supercoiled DNA [74]. We attribute the small end-to-end separation observed for λΔOL to the high compaction of the E. coli chromosome in the crowded cellular environment. While the exact molecular mechanisms responsible for compaction remain unclear, previous studies found that in vitro binding of the histone-like HU proteins [75] (accession numbers P0ACF0, P0ACF4) and in vivo mammalian chromatin packing [76] reduced the apparent persistence length of DNA. Hence, it is possible that nucleoid-associated proteins such as HU may bring distal DNA sites together by protein–protein interactions and/or affect local DNA conformations by introducing bends and relieving torsional strain [77]. Another important factor could be negative supercoiling, which has been shown to compact the chromosomal DNA globally [78]. However, the exact effect of negative supercoiling on a 2.3-kb DNA segment is difficult to predict, because negative supercoiling could also introduce extended, plectonemic structures that promote large separations between DNA sites on relatively short length scales [78]. Our two-color, high-resolution method can be applied to examine how chromosomal location, DNA length, genetic background, and growth conditions affect the distance between any two DNA sites on the E. coli chromosome. Furthermore, the spatial organization of the E. coli chromosome can be determined by systematically measuring distributions between DNA sites throughout the chromosome. This method is similar to how chromosome conformation capture was used to generate a 3D model of the C. crescentus chromosome [79], but with significantly improved spatial resolution and without potential artifacts from fixation. A plasmid, pS2391, containing lacO3 and tetO3 (the tetO2 sequence [54] was used for each repeat in tetO3) sites was synthesized by Genewiz, Inc. Segments of λ DNA (OR through OL for λWT, OR up to but not including OL for λΔOL) from the wild-type lysogen JL5392 (a gift from John Little, University of Arizona) were amplified by PCR. This DNA was sequenced and inserted between lacO3 and tetO3 using the In-Fusion PCR cloning system (Clontech). A kanamycin-resistance cassette flanked by BamHI sites was amplified by PCR and inserted after lacO3. For strains with mutated operators, mutations r1 [80], OL3–4 [13], and cIG147D [62] were introduced to the λWT template via QuikChange (Agilent). A plasmid carrying the PRM−cI− mutations (Figure S5c) (λΔOLPRM−cI−) was constructed by overlapping PCR mutagenesis using complementary primers carrying the desired mutations, flanked by a forward primer that sits at the EcoRI site on the upstream end of the operon and a reverse primer at the ClaI site in the rexA gene downstream of cI. The 1.13 kb PCR product was introduced to the λΔOL plasmid by restriction ligation. This procedure resulted in seven plasmids that were used as templates in subsequent chromosome insertion: pZH105 (λnull), pZH016 (λΔOL), pZH107 (λWT), pZH107r1 (λOR3−), pZH107OL3–4 (λOL3−), pACL006 (λWTG147D), and pACL007 (λΔOLPRM−cI−). Note that we use shorthand names such as λnull here for clarity; corresponding names used in our laboratory are listed in Table S4. The DNA sequence including lacO3, the λ DNA segment, tetO3, and the kanamycin resistance cassette was inserted into the chromosome of E. coli strain MG1655 by λ Red recombination [81], excising the lac operon, lacI, and all lacO sites. To express the CI protein in trans from a plasmid, we constructed the plasmid pACL18 in which the wild-type cI ORF is driven by a constitutive promoter, PRMc, which has the wild-type −35 (TAGATA) and −10 (TAGATT) sequences, lacks OR2, and has a mutated OR1 sequence (CGCCTCGTGAGACCA) that eliminates binding by CI. The pRMc–cI fragment was then cloned to the ClaI site of the low-copy vector pACYC184. The plasmid pACL17 was generated similarly using a template containing the CIG147D mutation. The two-color reporter plasmid pLau53, which expresses LacI-ECFP and TetR-EYFP polycistronically under the control of the PBAD promoter [82], was obtained from the Yale Coli Genetic Stock Center. Because the autofluorescence spectrum of live cells is generally strongest at wavelengths around 500 nm [83], single-molecule imaging of blue-shifted fluorescent proteins such as ECFP is difficult. The red fluorescent protein mCherry, which further benefits from a large Stokes shift, fast chromophore maturation rate, and high brightness relative to other monomeric RFPs [84], was inserted in place of ECFP. We also created a tandem LacI-mCherry-EYFP reporter, which was used as a fiducial marker, by inserting the linker sequence from the tandem-dimer fluorescent protein tdTomato [84] in between mCherry and EYFP. To accurately localize a fluorescent spot arising from only a few fluorescent protein molecules above the background of unbound molecules within a cell, we reduced the reporter expression level by weakening the ribosome binding sites (RBSs). Weakened RBS sequences were designed using an online RBS calculator [85]. For example, the RBS for TetR-EYFP translation was the consensus AGGAGG Shine-Delgarno sequence in the parent plasmid pLau53. Our reporter plasmid had an ACCAGG Shine-Delgarno sequence, with a predicted ∼300-fold decrease in the TetR-EYFP translation rate. All sequences including chromosome insertions were verified by sequencing (Genewiz Inc). Reporter plasmids are described in Table 1. For all experiments reported in this study, cells were grown and imaged at room temperature (∼25°C) in M9 minimal media supplemented with MEM amino acids (Sigma). Cells were grown overnight with 0.4% glucose and 50 µg/ml carbenicillin to an optical density (OD600) of 0.4. After centrifugation at room temperature, cells were resuspended at OD600≈0.2 with 0.4% glycerol plus 0.2% L-arabinose and grown for 2 h (∼1 cell cycle) to induce LacI-mCherry and TetR-EYFP expression. Cells were again resuspended at OD600≈0.2 with 0.4% glucose and grown for another 2 h before immediate observation to allow time for fluorescent protein chromophores to mature. We compared growth rates for the parent strain MG1655 to the experimental strain λnull to determine whether inserting the lacO3 and tetO3 construct into the chromosome and/or inducing expression from the reporter plasmid introduced a significant growth defect. Under induction growth conditions (∼27°C, M9 media with 0.4% glycerol and 1× MEM amino acids) starting at OD600≈0.1 and observing 8 h of growth, we measured doubling times of 2.7 h for MG1655 and 3.4 and 3.3 h for λnull harboring the reporter plasmid (in the absence and presence of 0.3% L-arabinose, respectively), indicating that there is no large growth defect associated with the insertion of the tandem operator sites into the chromosome and/or the expression of TetR-EYFP and LacI-mCherry fluorescent fusion proteins (Figure S6c). In each experiment, samples of all strains were placed on separate gel pads in the same growth chamber. Two sets of at least 30 movies were acquired for each strain, with the second set acquired in the reverse order to minimize any bias possibly introduced by observing some strains in a particular order. All images were acquired within less than one cell doubling time. Cells were put on a gel pad made of 3% low-melting-temperature SeaPlaque agarose (Lonza) in M9 with glucose and imaged on an Olympus IX-81 inverted microscope with a 100× oil immersion objective (Olympus, PlanApo 100× NA 1.45) and additional 1.6× amplification. Images were split into red and yellow channels using an Optosplit II adaptor (Andor) and captured with an Ixon DU-895 (Andor) EM-CCD with a 13-µm pixel width using MetaMorph software (Molecular Devices). Laser illumination was provided at 514 nm by an argon ion laser (Coherent I-308), which also pumped a rhodamine dye laser (Coherent 599) tuned to ∼570 nm. A quarter-wave plate (Thorlabs) was used to circularly polarize excitation light. Emitted light was split by a long-pass filter, and the red and yellow images were filtered using HQ630/60 and ET540/30 bandpass filters (Chroma). Images were inspected manually using a custom MATLAB script to identify spots that appeared in both EYFP and mCherry images. Images from all strains were displayed in random order without knowing the strain identify to avoid bias in spot selection. Pixel intensities within 3 pixels of the initial spot location were fitted with a symmetric, two-dimensional Gaussian distribution to estimate spot coordinates. The variance of the fit distribution was constrained to be less than 2 pixels. Spot-fitting error was estimated by scrambling residuals from a fit to the fluorescence data in 10 random permutations, adding them to the data, and fitting the resulting images; the reported error for a spot is the standard deviation of the distances between these fits and the initial fit to the raw data. Fitting error distributions are shown in Figure S1a. The LacI-mCherry-EYFP tandem dimer (Figure S2b) in which the two fluorescent proteins were directly fused together was used to acquire fiducial control points to transform between the mCherry and EYFP coordinate systems. A projective transform was calculated from the control points using the cp2tform function in MATLAB. We found that relatively simple, global transformations were sufficient to transform coordinates of fluorescent beads (Tetraspeck, Invitrogen) with ∼10-nm registration error in our microscope setup, and did not see any further improvement with a locally weighted transformation used in in vitro two-color experiments [21]. This transformation was also used to generate the overlay images in Figure 2, Figure 3, and all supplemental movies. Fluorescent beads were not used as fiducial markers because the beads' emission spectra were different from those of the fluorescent proteins. Analysis was restricted to molecules in which mCherry and transformed EYFP coordinates were separated by less than 200 nm. Separations beyond this threshold were rare (∼1% of data, see two-dimensional distributions in Figure S4) and did not correlate with strain identity in any reasonable way. They possibly arose from data in which cells contained two labeled copies of OR–OL DNA. After transformation into a uniform coordinate system, was calculated from the mCherry and EYFP coordinates and multiplied by an 81-nm pixel size (resulting from 160× magnification on a CCD with a 13-µm pixel width). Probability and cumulative distributions and were calculated for 10-nm bins using the kernel smoothing probability density estimation (ksdensity) function in MATLAB, restricting the density to positive values and employing a uniform kernel width small enough to follow empirical cumulative density distributions without any systematic errors. Significant differences between distributions were determined using a two-sample Kolmogorov–Smirnov test; two-tailed Student's t tests of sample means returned smaller, more significant p values. Errors in and were determined by calculating the means of 1,000 bootstrapped samples; the reported error is the standard deviation of the calculated means. Looping frequencies were estimated by least squares fitting of 1,000 bootstrapped distributions (control distributions were also randomized on each iteration) and their error was calculated similarly. Concentration measurements by smFISH followed a previously described protocol [66]. Transcripts from PRM were labeled with a mixture of 42 oligonucleotides labeled with CAL Fluor Red 610 (Biosearch Technologies), 31 of which hybridized to cI (11 targeted sequences not found in E. coli and did not cause a problematic level of false positives). Table S5 lists all 42 oligonucleotides. Labeled cells were imaged with 561-nm excitation at six imaging planes separated by 200 nm z-depth with negligible photobleaching. For each frame, fluorescent spots were automatically detected and fit to a Gaussian using a custom MATLAB routine. Nearly all molecules appeared in multiple image slices; the slice with the largest fit amplitude was kept. The integrated fluorescence of spots was observed to be quantized with one or a few molecules localized within one diffraction-limited spot. The intensity of one transcript was estimated from the distribution of spot intensities, and the number of molecules contributing to each spot was estimated from this quantization. The number of transcripts in each cell was estimated from the sum of the number of molecules in each spot within that cell. Alternatively, the number of molecules in one cell is proportional to its integrated fluorescence; this measurement provided the same average expression levels within error. The experiment was repeated to ensure that differences in labeling efficiency between samples were not responsible for differences in the number of detected molecules; combined data from both experiments were used for analysis. To generate simulated distributions, we first generated 10,000 random radial distances for a chain with a contour length and persistence length from a worm-like, noninteracting chain model using a Gaussian distribution with Daniels' approximation, which is accurate in the regime [72]:Each simulated was projected onto the plane at a random angle to give a distance . Simulated spots were placed at coordinates and . The MATLAB function mvnrnd was then used to simulate normally distributed measurement error with a standard deviation of 22 nm to the coordinates of each simulated spot. This procedure was sufficient to simulate the λnull distribution (Figure S2c) using a fixed end-to-end distance of 22-nm (approximate distance between the centers of the lacO3 and lacO3 sites; Figure S2a). Note here that the simulation is simplified in that it assumes that each spot has the same 22-nm localization error. In reality, localization error varies between different spots (Figure S1a) and there are other sources of measurement error. These differences may explain the slight deviation of the simulated distribution from the experimental distribution. The same procedure was used to estimate the expected for 2.3-kb, B-form DNA with a 50-nm persistence (∼200 nm) as well as the apparently persistence length (3 nm) implied by the 71-nm observed for λΔOL. Additional descriptions of thermodynamic states are listed in Table S3. Parameter values were determined by first scoring a wide range of parameter values and iteratively searching narrower and more finely grained parameter ranges to manually minimize the sum of the squares of the differences between experimental and modeled values for looping frequency and CI expression level. We then refined this fit by least-squares minimization using MATLAB. This was done using a minimized model that only accounted for states likely to be populated near or above lysogenic CI concentrations (e.g., disregarding states in which OR1 and OR2 are unbound by CI). Using the same parameters and accounting for all 176 possible states (122 unique states accounting for degeneracy) did not significantly change the fit results. Fitting with this much more complex model gave octameric and tetrameric looping free energies of 0.6 and −3.3 kcal/mol and unlooped and looped expression rates of 2.1 and 5.3 nM/min. When determining parameters, rates were expressed in terms of changes in concentration per unit time; we followed earlier work in assuming that in a typical E. coli cell, a single molecule is at a concentration of ∼1.47 nM [35]. We do not report any estimate of fitting error; instead, we present only the parameters most consistent with our data and assumptions. Figure 5c and d shows that fit parameters were well-determined at a given combination of wild-type CI concentration and nonspecific binding parameters. As noted in the main text, varying these two parameters changed the absolute best-fit parameters, but did not dramatically change our conclusions. Furthermore, fixed parameters of previous studies were determined in a number of separate experiments employing different methods at temperatures other than 25°C; a rigorous estimate of modeling error would require knowing the error in the measurements of fixed parameters in our experimental conditions. The basal CI expression rate, , was arbitrarily fixed at ; this did not have any significant impact on determining other parameters, as our measurements were all at or above lysogenic , where OR2 is almost always bound by a CI dimer. Additionally, the fraction of free CI dimers was fixed at its value for 150 CI molecules per cell at a given concentration of nonspecific binding sites and nonspecific binding affinity. Fixing the concentration of free CI dimers is a reasonable approximation if (1) nearly all CI molecules are in dimers and (2) the number of free nonspecific binding sites is not significantly changed by nonspecifically bound CI dimers. Figure 2a–e, Figure 4b, and Movies S1, S2, S3, S4, S5, S6 were prepared using NIH ImageJ [86]. Raw fluorescence image intensities were scaled linearly from the lowest to highest values in region shown. For EYFP/mCherry overlay images, brightfield images were inverted and converted to 8-bit RGB. Fluorescence images were bandpass filtered and background subtracted before being used to generate magenta (mCherry) and green (EYFP) 8-bit RGB images that were added to the brightfield image. The EYFP images were first transformed in MATLAB using the imtransform function and the same fiducial data that were used to transform EYFP spot locations into mCherry coordinates. For smFISH images (Figure 4b), the value of each pixel is the maximum value of that pixel in six images collected at different z-axis positions. Intensities for all images were scaled linearly from the minimum to the maximum of all pictures (117–4,840 counts in 16-bit images).

10.1371/journal.pntd.0006682

An exploratory study on rabies exposure through contact tracing in a rural area near Bengaluru, Karnataka, India

Rabies is a neglected zoonotic disease. Given the low incidence, apart from the existing reporting syst, there is a need to look for other means of case detection strategies for rabies. Contact tracing is one such method to efficiently capture information. To find out the rabid status of biting animal through contact tracing and to determine health seeking behavior of the bite victims. An exploratory study using contact tracing was conducted during the first quarter of 2017 in villages coming under three Public Health Centers. The households of the bite victims were visited and details of rabies exposure obtained from the bite victim/ adult responsible respondent using a standardized questionnaire. A total of 69 dog/cat bite cases were identified. 69.5% of bites were by stray dogs. 97.1% bite victims had Category III bites. Only 4.5% bite victims had taken PEP. 70.1% of animal bite cases were administered ARV. Only 7.2% bite victims had exposure to probable rabid animals. All dog bite victims were alive after 3 months of follow up. Contact tracing was successful in case detection of probable rabid animal exposures and suitable for a period of one year.

In India, Rabies is a neglected zoonotic disease and the burden of rabies is usually not captured by the health system due to varied reasons. Hence, an exploratory study was attempted to find out the rabid status of the biting animal through contact tracing during the first quarter of 2017 in villages coming under three Public Health Centers. A total of 69 dog/cat bite cases were identified. Majority of bites were by stray dogs, Category-III, exposure to non cases and few received PEP. Percentage of bite victims who actually required PEP was calculated. These findings support the fact that contact tracing can be used as an additional tool by the health care provider for measurement of rabid animal bites, where resources are scarce, reporting systems are weak and priorities for disease vary. A rabies check list/score card to be made available at Public Health Centers, which shall ensure better utilization of limited but lifesaving Rabies Immunobiologicals.

Rabies is a 100% fatal viral zoonotic disease.[1] However, Rabies can be prevented if, health care providers are able to identify the type of exposure, categorize the wounds and provide post exposure prophylaxis(PEP) as early as possible.[2,3] South East Asia region accounts approximately for 60% of human rabies deaths in the world. An estimated 20,000 rabies deaths (approximately 2/100,000 population) and 17.4 million exposures to animal bite occur every year in India.[4] The data on incidence of human and canine rabies is currently not known and needs to be updated. In reality, the burden of rabies is usually not captured by the health system due to varied reasons. Moreover, Rabies is not a notifiable disease in India and acts as a surveillance barrier for measuring the burden of disease. Also, the laboratory-based animal rabies surveillance program to measure the burden is non-existent in many regions of the country. Human and animal rabies cases are under reported in most of the developing countries due to ineffective rabies surveillance methods.[5] To understand the national burden of rabies, estimation methods must be periodically conducted.[6]In India, the populations most at risk from rabies exposures are living in rural areas, in poor, marginalized communities where surveillance and reporting systems are weak. The need of the hour is measurement of bite victims bitten by a rabid dog and having access to PEP. Contact tracing provides an efficient method of documenting such information. Contact tracing consists of sequential steps of data collection and investigation, revealing different aspects of health seeking behavior, treatment cost and health outcomes. Contact tracing is finding everyone who comes in direct contact with rabies exposure usually occurring through dogs and helps in targeting at risk population [7] In India, A number of changes have taken place over the years including introduction of intra dermal rabies vaccine (IDRV), abolition of nerve tissue vaccine, availability and accessibility to rabies immunobiologicals and better awareness about rabies in the population. [8]Hence it is assumed that the rabies burden would have come down. In this background, an attempt has been made to find out the rabid status of the biting animal through contact tracing in a rural area, to determine the health seeking behavior of the bite victims and to recommend measures for rabies prevention. An exploratory study was conducted during the first quarter of 2017 in villages (rural area) coming under three Public Health Centers (K.Golahally, Kengeri and Sullikere). The villages were located about 25 kilometers away from Bengaluru, Karnataka, India. The average population was about 1100 and population of the villages ranged from 81 to 3165. The study area is located in a dry belt with little agriculture activity and the mean temperature was about 25°C. [9] Map of study area is described in Fig 1. Typical contact tracing includes diseases which involves human to human transmission. As far as rabies is concerned, one of the participants in transmission could not be interviewed i.e. the dog. Hence, the methodology for contact tracing was modified, where the index case was interviewed to identify other exposures of the biting animal, which would herewith be considered as contacts. The local government primary care provider (Anganwadi Centre) in each village was the starting point of contact tracing. The female staff was the first informant, as she is the front line primary care worker and would have had basic information on the morbidities reported. Through her cooperation, the initial process of identifying the index bite victims was started, which would have otherwise been difficult for the investigator in terms of acceptability and feasibility. The bite victims were people who had been exposed to animal bites in the last one year from the date of survey. Ideally, the bite victims details should have been got from the health center. However, the address and village would have been difficult to get as they are usually not recorded in the health center. Hence, the index case was the starting point of contact tracing in the village. Majority of the bite victims were interviewed within 6 months of the exposure, the minimum and maximum duration of exposure was one week and 11 months from the date of survey. The bite victims were followed up for a period of three months to know the outcome i.e alive/dead. The household of each bite victim was visited by the investigators and detailed information regarding biting animal, availability of animal, rabies exposure, PEP, rabies-related deaths following animal bites and clinical signs of rabies in the biting animal, etc was obtained using a standardized semi-structured questionnaire. The bite victims were the respondents in majority of the interviews and when they were not available, the head of family /adult responsible respondent were interviewed. For contact tracing of animal bite, the respondents were asked about the index biting animal having bitten other animals or humans and if they had come across people who had been bitten by a dog/cat.[10]Subsequently, the households of other bite victims were visited and detailed information on exposure was obtained. This procedure of contact tracing (snowball sampling technique) was repeated until all probable exposures were identified. To maximize the efforts to trace all bite victims, information was also obtained from villagers, Accredited Social Health Activist (ASHA) workers, formal and informal leaders. Additionally, the investigators searched for the availability of the biting dogs/cat during the survey. The bite victim must have been a resident of the village for a minimum of six months. Subjects bitten by animals not known to cause rabies in humans were excluded from the study. Finally, a total of 69 dog/cat bite victims were selected through contact tracing during first quarter of 2017 covering 17 villages. A total of 69 bite victims were followed up. The age range of bite victims was 3 to 89 years respectively. The median age of bite victim with Inter Quartile Range (IQR) was 17 (9,18) years. 52% of the bite victims were males. 33 animals (32 dogs & 1 cat) were responsible for the exposure among 69 bite victims. 69.5% of the bites were by stray dogs. There was one (1.4%) event of a dog having bitten a cat. There were no wild animal bites. Information regarding outcome (alive, dead, sick, unknown) of all the 33 dogs/cat was available and given by the bite victims. Out of the 69, 5(7.2%) bite victims had exposure to 3 probable rabid dogs and 1 probable rabid cat. The remaining bite victims were exposed to non cases. All 4(100.0%) probable rabid dogs/cat had died within 10 days. The clinical signs observed in the probable rabid dogs/cat were hypersalivation in 3(75.0%) and aggressive behavior in 3(75.0%). Laboratory confirmation of rabies diagnosis was not done in any of the dogs/cat that had died. All the 69(100%) bite victims were alive at the end of three months of follow up (Fig 2). There was no relationship between any two dog bites within the village and between adjacent villages. Each of the bites had occurred in different time periods as sporadic event. In majority of the villages, one dog was responsible for each exposure. A maximum of 7 persons were bitten by a single dog. Only in one village, 4 dogs were responsible for 4 different exposures. An example of contact tracing in village X is given in Fig 3. 67(97.1%) bite victims had category-III exposures and 2(2.9%) had category-II exposures. 3 (4.4%) category-III exposures had received PEP (Rabies Immunogobulin & Anti Rabies Vaccination), 46(68.7%) had received Anti rabies Vaccination (ARV) alone and 18(26.9%) had not sought any treatment. 1(50.0%) category-II bite victims had received ARV and 1 (50.0%) had not sought any treatment. Among those who had received ARV (PEP included), 31(62%) bite victims had completed full course of ARV (5 doses) and 19 (38%) had incomplete Anti Rabies Vaccination.14(73.6%) bite victims informed that doctor /health worker did not advice and 5(26.4%) said busy as reason for not completing ARV. Among the 5 bite victims exposed to probable rabid dogs/cat, the age of the victims were 14,16,17,32,and 60 years. 3 (60.0%) were males and 2(40.0%) were females, all the exposures were unprovoked in nature, 2 (40.0%)bites were over the hand, 1(20.0%) on the back, 1(20.0%) forearm and 1(20.0%) was on the leg. 2 (40%) bite victims had washed the wound with soap and water,1 (20%) washed only with water,1(20%) had applied antiseptic and 1(20%) did not do anything. Two (40%) probable rabid bite victims had completed ARV (5 doses), 1 (20%) had incomplete Anti Rabies Vaccination, while the other 2(20%) had not sought PEP and reason being ignorant of ARV. An average of one dose of ARV was taken by non cases and average of three doses of ARV by probable rabid animal bite victims. The average number of people bitten by probable rabid dogs was one and average number of people bitten by non rabid dog was two. None of the bite victims had received first aid by the village primary care provider and all of them were referred to higher centers. 26(52%) bite victims had visited the private health care provider for availing treatment and 10(20%) had visited both government and private health care centers. The average cost of PEP incurred per person was Rs.1049.70(16$) and average cost of transport Rs.165.24(3$). However, assessment of costs is additional information gathered and was not looked as barrier for PEP. Table 1 describe the ratio of rabies exposure and PEP. Contact tracing conducted immediately with little or no delay between the bite and the interview, will give detailed and accurate information. Contact tracing can also be carried out retrospectively to maximum of one year for more reliable data. Data collected beyond one year can result in recall error. [10] From contact tracing it was observed that, less than 10% of the exposures were by probable rabid animals and no brain sample was examined for confirmation of rabies. The interview of index case, family members and stake holders to trace bite victims and animal cases in the villages was similar to the studies done in Bali and Bohol.[12,13] Different studies on contact tracing among community members and healthcare workers, were either to trace bite events, or to look for clinical signs of rabies in the biting dogs, or to see if they could find out strategies for rabies control or to identify contacts for post-exposure prophylaxis to prevent the disease.[14,15,16] Majority of the bites were category-III exposures. Thirteen stray dogs were responsible for more than half of the bite victims in concordance to the observations from other studies.[17,18] Majority of bite victims had first visited a clinic/hospital in the town for treatment as the primary care provider in the village was not aware about rabies PEP. Two (40.0%) probable rabies exposure victims did not seek treatment in the present study compared to contact tracing survey in Tanzania, which had showed that 15% and 24% of suspect rabies exposure did not seek medical attention.[6] In a hospital based study in Bhutan, 32% of the subjects mentioned that the biting dog looked normal and 9% mentioned that the biting dogs looked like suspect rabies contrary to the observation of the present study.[19] Studies have shown that hypersalivation and aggression are the common clinical signs observed for diagnosis of rabies in dogs similar to the finding in the present study. [12,14] The World Health Organization reported that, almost all rabies death victims had not sought rabies PEP and there were no facilities or health personnel available to provide PEP in many areas where the disease is prevalent and suggested strengthening availability of Rabies Immunobiologicals in these places.[20,7] It is concerning that 60% victims of probable rabid animal exposures were either not aware or, did not seek care or did not complete PEP. The study is too small to make any determinations about the rate or risk of human rabies in the study area. The standard surveillance practices applied to many human and animal diseases consist of case identification, contact tracing, epidemiologic investigation, and laboratory confirmation.[21] The outcome of animal (alive, dead, sick, unknown) was mainly based on information provided by the bite victim/people in the village. The surveillance system for dog rabies diagnosis was non existent in the villages surveyed. The methods of rabies surveillance practiced in many countries suffer from fundamental problems including a lack of trained professionals and lack of diagnostic laboratory capacity. This results in a lack of awareness of case burden, reduced funding for control, and poor community engagement around prevention.[8,11] Having a functioning surveillance system in villages will go a long way in achieving the WHO, OIE and FAO goal to educate, vaccinate and eliminate dog-mediated human rabies deaths in the world by 2030.[22]Laboratory diagnosis is critical to confirm the status of a suspect case, in part, to justify prophylaxis in exposed persons or animals.[23] Human rabies is underreported and the disease is not a priority in endemic countries.[24] One Health emphasizes that the rabies control and elimination should be a joint effort of veterinary and medical field.[25] The Indian rabies survey had estimated that, for every 870 bites, there will be one rabies case. However, it was not possible to elicit, which bite was responsible for the rabies death.[4] The present study through contact tracing was able to identify bite victims who were exposed to probable rabid animal exposures. India is a vast country with limited resources. Data on availability, accessibility and affordability of PEP is needed to plan for better intervention strategies. Nearly 28% of the subjects did not receive any rabies prophylaxis in the present study. In such a situation, a rabies risk score card based on the knowledge of local rabies transmission, category of bite and dog rabid status, etc can be developed. The rabies score card/check list would be able to identify the bite victims who need rabies PEP. This can be made available to the primary care providers in the village along with campaigns for strengthening of rabies IEC in the community. Yes, there are chances of missing out genuine cases, however these can be overcome if the scale/check list has very high sensitivity, treating physicians are asked to administer PEP in case of doubt and individuals with high risk are targeted. Snowball methods are typically performed to find rare events, such as human rabies deaths. They are less accurate at describing rates of more common events, like bites or healthcare seeking behaviors. Some of the bite victims may have been missed because of the contact tracing methodology followed (selection bias). Information on details of animal bites, PEP seeking behavior and follow up vaccination of the bite victims was based on the history revealed by the bite victim (Information bias). Categorization of bites was according to bite victims observation (observer bias), money spent and treatment taken are as revealed by bite victims (recall bias). The classification of biting animal as non cases and probable rabid is based on facts given by the bite victims/people (information bias). Clinical and laboratory confirmation of rabies in the biting animal was not possible. A study covering a wider geographical area was not possible due to feasibility issues. Contact tracing was successful in case detection of probable rabid animal exposures and suitable for a period of one year. In addition, contact tracing identified that only one tenth of the bite victims required PEP. Implementation of contact tracing technique for identification of rabid status of the biting animal. To conduct a prospective study in a larger geographic area. Development of a rabies risk score card for the health care provider.

10.1371/journal.pgen.1007799

Genetic regulation of the placental transcriptome underlies birth weight and risk of childhood obesity

GWAS identified variants associated with birth weight (BW), childhood obesity (CO) and childhood BMI (CBMI), and placenta is a critical organ for fetal development and postnatal health. We examined the role of placental transcriptome and eQTLs in mediating the genetic causes for BW, CO and CBMI, and applied integrative analysis (Colocalization and MetaXcan). GWAS loci associated with BW, CO, and CBMI were substantially enriched for placenta eQTLs (6.76, 4.83 and 2.26 folds, respectively). Importantly, compared to eQTLs of adult tissues, only placental eQTLs contribute significantly to both anthropometry outcomes at birth (BW) and childhood phenotypes (CO/CBMI). Eight, six and one transcripts colocalized with BW, CO and CBMI risk loci, respectively. Our study reveals that placental transcription in utero likely plays a key role in determining postnatal body size, and as such may hold new possibilities for therapeutic interventions to prevent childhood obesity.

Genetic studies (e.g GWAS) revealed substantial heritability on birth weight (BW), childhood obesity (CO) and childhood body mass index (CBMI), however, the etiological mechanisms and relevant tissue(s) underlying these traits/conditions are not clear. We incorporated the data from largest GWASes to date and placenta expressional quantitative trait loci (eQTL) that have been newly published, and showed the variants associated with BW, CO and CBMI were substantially enriched for placenta eQTLs (6.76, 4.83 and 2.26 folds, respectively). Importantly, compared to eQTLs in 7 adult tissues such as adipose and liver, only eQTLs in the placenta were found to contribute significantly not only to anthropometry outcomes at birth (BW) but also to childhood phenotypes (CO/CBMI). Further, we employed COLOC and MetaXcan analyses and identified placenta transcripts potential mediate the genetic effect of BW/CO/CBMI GWAS loci. In summary, our study strongly supports a key role for the placenta in determining BW, CO and CMBI at the molecular level, and pinpointed genes whose expression levels in placenta potentially influences BW and CO risk.

Birth weight (BW) is influenced by both fetal and maternal factors, including race, infant sex, plurality, altitude, education, and smoking [1], and is consistently associated with future risk of adult metabolic diseases including type 2 diabetes (T2D) and cardiovascular disease [2]. Childhood obesity (CO) has also emerged as an important health problem in the United States and most other countries in the world [3,4]. In the USA, 1 in 3 children is afflicted with overweight or obesity [3]. The increasing prevalence of CO is also associated with diseases later in life including T2D, hypertension, nonalcoholic fatty liver disease, obstructive sleep apnea, and dyslipidemia [3]. The placenta, situated at the maternal-fetal interface, is a key organ for fetal growth and development; it performs a variety of functions including controlling fetal access to nutrients, hormone production, and mitigation of adverse effects from the environment [2,5,6]. The placenta and associated extraembryonic membranes are of fetal organ (i.e, have the same genetic composition as the fetus), formed from the zygote at the start of each pregnancy [7]. Several studies, including our own, have reported associations linking placenta markers to variation in birth weight [8,9]. Recently, we leveraged the Rhode Island Child Health Study (RICHS) [6,10] resource and reported the first placenta expression quantitative trait loci (eQTL) profile based on a large sample size [11]. We discovered more than 3,000 cis- and trans-eQTLs at ≤10% false discovery rate (FDR), demonstrating that the placenta is a transcriptionally active tissue and that this activity is largely controlled by fetal genetics [11]. Large meta-analysis of BW [12], CO [13] and childhood BMI (CBMI) [14] genome-wide association studies (GWAS) have identified multiple genome-wide significant loci suggesting a strong genetic determinant for these conditions. At genome-wide significance, 60 and 15 fetal loci have been associated with BW (explaining 15% of its variance)[12] and CBMI[14], respectively. As many of these genome-wide significant variants are in non-coding regions, they likely affect these phenotypes through regulation of gene expression. Furthermore, while genetic variants are static within an individual, the regulatory effect that a genetic polymorphism exerts on outcomes will be likely to vary spatially, temporally and in response to differing environmental exposures and influences. As a noted limitation of the GWAS design, while informative for identifying genetic loci implicated in disease, this approach does not provide information on down-stream genes and their pathophysiological functions. More specifically, the GWAS design is unable to identify; (1) the tissue types and developmental stages where the causal variants and the underlying genes have their effects and (2) the mechanisms of effect of a causal variant. An eQTL analysis provides a profile of significant associations between transcription levels of a gene and genetic polymorphisms. A related concept, eSNP, denotes a particular SNP whose genotype is significantly associated with the transcription level of a gene. A distinction is made as to the location of the eQTL relative to the physical location of the gene whose expression level it is associated with; if the eQTL is close (within 500kb) to the gene (which encodes the mRNA), it is defined as a cis-eQTL. In contrast, if the eQTL is distant (typically more than 500kb away or on another chromosome altogether) from the gene whose expression level it is associated with, it is defined as a trans-eQTL. In this study, the cis- and trans-eQTLs in placenta and adult tissues were defined using these criteria, as previously reported [11]. An important approach is to apply knowledge of eQTLs in disease-relevant tissues to mine GWAS risk loci datasets to discover underlying mechanisms linking genetic variants to diseases [15,16]. Recently, using integrative genomics applied to eQTLs has become an even more powerful tool to mine GWAS results and shed light on disease pathways, as it can identify true disease causal variants (as opposed to merely undirected associations) with down-stream genes in a tissue-specific manner. Recently, methods like COLOC [17], SMR [18], MetaXcan [19], and TWAS [20] have shown superior ability to identify causal genes of genetic risk loci compared to earlier integrative analysis simply looking at overlap between eQTLs and GWAS lead SNPs. Herein, to study the in utero genes/pathways influencing BW, CO and CBMI, and to define the specific role of the placenta for neonatal and long-term child health outcomes, we compared placenta eQTLs with 7 STARNET adult tissue eQTLs (blood, atherosclerotic-lesion-free internal mammary artery, atherosclerotic aortic root, subcutaneous fat, visceral abdominal fat, skeletal muscle, and liver). Further, we applied COLOC and MetaXcan analyses on placenta eQTLs integrated with GWAS data. Summary level data of three GWAS meta-analyses were retrieved (Materials and methods): (1) a multi-ancestry GWAS meta-analysis of BW in 153,781 individuals [12]; (2) a CBMI GWAS meta-analysis of 20 studies with a sample size of 35,668 [14], where the BMI was measured between 2 and 10 years of age; (3) a CO meta-analysis of 14 studies consisting of 5,530 cases and 8,318 controls [13], where cases were defined as BMI>95th percentile and controls defined as BMI<50th percentile. We overlapped the GWAS results (summary level data, Materials and methods) of BW, CO and CBMI with placenta eSNPs (10% FDR), and found that placenta eSNPs were substantially enriched for GWAS hits with highly significant p-values (Fig 1). Among GWAS hits with p-value ≤ 10−5, placenta eSNPs were 6.76 fold enriched for BW, 4.83 fold for CO; and 2.26 for CBMI (S1 Table). We explored similar GWAS-eQTL overlapping in tissue types collected in adult subjects using eSNPs (FDR≤10%) from the STARNET study[21] (Materials and methods). All tissues’ eSNPs were enriched for BW GWAS signals, and the enrichment were significant (pvalue<1e-6) in Kolmogorov—Smirnov test (K-S test) even after LD pruning (Materials and methods). The eSNPs in subcutaneous fat and visceral abdominal fat tissues demonstrated most substantial enrichment among adult tissues eQTLs, comparable to that of placenta eSNPs (Fig 1A). Interestingly, unlike the placenta, none of the adult tissue eQTLs in STARNET demonstrated magnitude of enrichment in CO or CBMI GWAS hits comparable to placenta eSNPs (Fig 1B and 1C), indicating a more prominent role of the placenta transcriptome in postnatal anthropometry outcomes. We integrated placenta eQTLs with GWAS summary data using the COLOC approach [17] to identify transcripts influenced by the same SNPs driving disease phenotype. COLOC was conducted on 1,863 non-overlapping intervals around BW GWAS hits (Materials and methods), and eight transcripts (ITPR2, PLEKHA1, TBX20, METTL21A, SPATA20, TMEM87B, RPS9, and HAUS4) were identified as PP.H4 >75% (Table 1). The finding suggested, at these loci, a single genetic variant controlled both BW and transcript levels in placenta. Six genes were genetically co-localized with CO [13] with PP.H4 >75% and one gene was co-localized with CBMI (Table 1). All transcripts identified by COLOC analysis were protein coding genes. We also conducted COLOC integrating placenta eQTLs and adult BMI [22], and no transcript showed large PP.H4 (e.g. ≥0.75). We implemented MetaXcan [19] to integrate GWAS summary data (BW, CO and CBMI) and sample level placenta eQTL data to identify genes whose imputed expression level was associated with GWAS traits. Table 2 lists the findings that passed Bonferroni correction (Materials and methods). Nineteen genes were associated with BW. Importantly, seven out eight genes genetically colocalized with BW were also identified by MetaXcan (Table 2). Further, MetaXcan inferred the directionality of the transcription-trait association, for example, higher PLEKHA1 mRNA level in placenta was associated with low birth weight. The expression level of three genes (ADCY3, DNAJC27, TYW3) were associated with childhood obesity risk. ADCY3 and DNAJC27 belong to the genomic locus, and both have positive correlations with CO and CBMI (Table 2). We corroborated the MetaXcan-based inferred directionality of the imputed transcription vs. BW associations using available data: observed BW and placenta transcription (whole transcriptome RNAseq, Materials and methods) in the RICHS cohort. Out of the 19 genes identified by MetaXcan (Table 2), placenta mRNA level of seven genes was associated with BW with FDR<0.1, including XRN1, HSPA4, PLEKHA1, AGTR1, RSPO3, MRPL10, and CLDN7 (S2 Table). Importantly, among these seven genes, the direction of six genes (except CLDN7) association with BW were consistent with that inferred by MetaXcan. The association between observed BW and measured expression level of these seven genes were robust even we adjusted the regression model for gestational age, baby sex, race, peak eSNP and GWAS peak SNP. Based on the GWAS summary level data on BW, CO and CBMI, we applied LD score regression [23,24] and found relatively high heritability (h2) of BW, CO and CBMI, being 0.1023, 0.4067, and 0.2421, respectively (Table 3). The heritability of CO is particularly large compared to other heritable traits. For example, h2 of T2D is 0.0872 and h2 of BMI is 0.1855[24]. In fact, the heritability of CO is comparable to that of height (h2 = 0.4623) [24,25], indicating it is strongly controlled by genetic background. Birth weight has significant genetic correlation (rg) with CO and CBMI at 0.1847 and 0.2038, respectively (Table 3), indicating BW and CO/CMBI share genetic determinants. Importantly, the positive rg demonstrates that higher BW is genetically associated with higher CBMI value and increased CO risk. We identified the “shared SNPs” that were associated with both BW and CO at GWAS p-value cutoffs of 1e-3, 1e-2 and 1e-1 (Table 4). Consistent with the positive rg, most shared SNPs have consistent allele direction in BW and CO (Table 4). It should be noted that few “shared SNPs” showed highly significant (e.g. p value <1e-4) association with both traits, and in fact, most of the shared SNPs had moderate significance (e.g. 1e-2), indicating that the common genetic predisposition of BW and CO is not attributable to a small number of loci of large effects, but rather to many genetic variants of small-to-moderate effect size. There is increasing recognition that the period of intrauterine development constitutes one of the most critical periods for defining disease risk later in life [26]. We report the first study to use comprehensive transcriptome data from the placenta (being one of the most relevant tissues in controlling in utero development) to interpret the causal effects of GWAS genetic risk loci for BW, CO and CBMI. We found extensive enrichment of placenta eQTLs for GWAS hits (Fig 1), and importantly, discovered hundreds of transcripts controlled by the genetic variants that influence these anthropometric traits. Our findings demonstrate that placental transcription and related functions defined by identified trait genes play important roles in determining BW, CO and CBMI BW, CO and CBMI are of major clinical and public health importance. BW has been convincingly shown to be inversely associated with risk for T2D, CAD and hypertension in later life [12]. CO or excess BMI are risk factors for T2D and hypertension [3]. These traits are also heritable, especially CO (h2 = 0.4067). Furthermore, these traits are genetically correlated (Table 3), where higher BW is associated with increased risk of CO. Importantly, we demonstrate that the placenta is a highly relevant tissue for determining not only weight at birth (BW) but also body weight in childhood (CO and CBMI). Identifying the genes mediating the genetic predisposition for BW, CO and CBMI is of great importance for understanding the mechanisms controlling these traits and developing clinical prevention strategies. Integrative genomics can be a powerful tool to mine GWAS results. Many methods have been proposed to integrate GWAS summary data and eQTL data, where GWAS and eQTL studies used different datasets or population. These methods include SMR, PrediXcan, Sherlock, COLOC, eCAVIAR, etc, and they can be categorized into two broad classes [27]. (1) Class 1 includes TWAS, MetaXcan and SMR, which are tests for significant genetic correlation between cis expression and GWAS. (2) Class 2 includes COLOC and eCAVIAR, which are estimations (rather than tests) of the posterior probability of colocalization, where colocalization is defined as shared causal variant(s) between the expression and GWAS. It has been reported that the results of these two classes of methods do not fully overlap [28]. Possible explanations are different assumption, algorithm, power, etc. Herein, we took a common strategy [28] and report all gene pinpoint by COLOC and/or MetaXcan, as they are potentially on the causal pathway of BW, CO and CBMI. In S3 Table, we report the LD between the lead eQTL and the lead GWAS SNP near the gene locus. There are six genes coded by a single BW GWAS locus with lead SNP rs2421016 (chr10:124167512) (Fig 2A), and we found that four out of the six genes were co-localized with BW (Table 1): PLEKHA1 (Pleckstrin Homology Domain Containing A1), HTRA1 (HtrA Serine Peptidase 1), ARMS2 (Age-Related Maculopathy Susceptibility 2), and BTBD16 (BTB Domain Containing 16). The eQTLs of these four genes also overlap with BW GWAS peaks (Fig 2B–2E). In particular, the shape of the PLEKHA1 eQTL overlaps well with the BW GWAS peak (Fig 2A and 2B), where the top eSNPs include the GWAS lead SNP (purple dots) and the SNPs in strong LD with the lead SNPs (red and orange dots). In COLOC analysis, PLEKHA1 showed PP.H4 much larger than PP.H3, indicating BW and PLEKHA1 expression levels share a single causal SNP. For HTRA1, ARMS2 and BTBD16 genes, the GWAS lead SNP and its high LD proxy SNPs (purple, red and orange dots) only showed association with transcription at moderate significance, while the top eSNPs are SNPs with moderate LD with the GWAS lead SNP (green and light blue dots), suggesting multiple SNPs influence HTRA1, ARMS2 and BTBD16 expression levels and some of these SNPs also impact BW. Of relevance, a study in a murine model found Htra1 is associated with placenta development [29]. In humans, upregulation of ARMS2 was observed in placenta of growth-restricted infants [29], and perturbations in ARMS2 may result in dysfunction of the extracellular matrix; suggesting that up-regulation of ARMS2 forms part of an important survival mechanism to compensate for placental growth discordance [29]. Furthermore, various GWAS have shown this locus (10q26.13) is associated with multiple diseases/traits, including age-related macular degeneration [30], type 2 diabetes [31,32], migraine [33] and height [34], suggesting that the versatile functions of these genes may affect many biological processes. In this paper, we report BW, CO, and CBMI GWAS signals are more profoundly enriched in placenta eQTLs comparing to adult tissue eQTLs (ie. STARNET). The differences are not driven by a few loci but rather polygenic. After strict LD pruning (Materials and methods), which eliminated the eSNPs that are in high LD, the conclusions remain unchanged that GWAS signals are more profoundly enriched in eQTLs of placenta than adult tissues (S1 Fig). Two-sample K-S test (on pruned SNP lists) showed such difference in enrichment magnitude were statistically significant (S4 Table). Such results suggest placenta could be a relevant tissue type controlling these neonatal traits. As a future direction, it would be important to investigate whether eQTLs of infant or pediatric tissues (e.g. skeleton muscle) substantially enrich for BW, CO, CBMI GWAS loci and compare to the results of placenta. The developmental origins of health and disease (DOHaD) hypothesis emphasizes the role of the in utero environment on shaping the developmental trajectory of the fetus and, thereby, subsequent health outcomes across the lifespan. Our findings showcasing the influence of genetically controlled in utero expression patterns linked to BW, CO and CBMI also support this hypothesis. One of our key observations is that placental eQTLs are enriched to a greater extent for CO/CBMI GWAS hits than eQTLs from other tissues, including adipose and liver (Fig 1B and 1C). This observation strongly suggests that the genomic control and transcription activity of these genes in the placenta has a profound and long lasting impact on postnatal development and human health. In addition, the transcripts potentially mediating the genetic control of BW and CO can be influenced by environmental factors, and our results shed light on environmental health and nutritional elements that are likely involved in the childhood obesity epidemic [4]. Moreover, our analysis also revealed the genes/transcripts likely to mediate the effects of genetic polymorphisms on BW and CO/CBMI, as well as the directionality of the mediation effect; thus providing prioritized targets for experimental follow-up, intervention and prevention in the fight against the current obesity epidemic. All GWAS summary level data are available from third party sources and no additional ethics approval is needed. The RICHS eQTL data is available in supplementary files, and the study protocol was reviewed and approved by the Office of Human Research Protections registered Institutional Review Boards [10]. We retrieved summary level data of a genome-wide association study (GWAS) meta-analysis of birth weight in 153,781 individuals, where multiple births were excluded. Further, extreme values (< 2.5 kg or > 4.5 kg) were excluded as implausible for live term births before 1970 [12]. Only GWAS meta-analysis summary level data of the Caucasian part of study (N = 143,677) is used in this paper. We retrieved summary level data from the Early Growth Genetics Consortium. In the Childhood obesity meta-analysis of childhood BMI in subjects of European ancestry, the BMI was measured between 2 and 10 years. In case of multiple births (twins, triplets), only one child was included. The GWAS meta-analysis included 20 studies with a sample size of 35,668 [14]. In Childhood obesity meta-analysis of childhood obesity in subjects of European ancestry, cases were defined as BMI>95th percentile; and controls defined as BMI<50th percentile. The meta-analysis was performed on 14 studies consisting of 5,530 and 8,318 controls [13]. We retrieved summary level data on adult BMI trait from the GIANT Consortium [22]. This adult BMI study is GWAS meta-analysis of 82 cohorts of 236,231 subjects in total. The summary level data was used in COLOC analysis and compared to the results of childhood BMI. The Rhode Island Child Health Study (RICHS)[10] consists of singleton, term infants (≥37 weeks gestation) born without serious pregnancy complications or congenital or chromosomal abnormalities. Birth weight of the RICHS subjects (newborns) was recorded and placenta tissue collected for RNA and DNA extraction [11]. Four full thickness biopsies free of maternal decidua were taken, one from each of four quadrants around the placenta, within 2 cm of the cord insertion site, and immediately placed in RNALater preservative. After at least 24 hours, these samples were snap frozen in liquid nitrogen and homogenized together into a powder, and aliquots of that homogenized sample were used with the intent to reduce variation based on location. RNA and DNA were profiled using whole transcriptome RNAseq and Illumina MegaEX SNP array, respectively [11]. After stringent QC, a subset of the RICHS placenta collections (N = 150) with high quality RNAseq and whole-genome genotyping data were used for generating eQTLs. The data is primarily of Caucasian ancestry (77.3%) and the race composition is presented in S5 Table. Placenta eQTLs were previously reported and are publically available [11]. We compared placenta tissue eQTLs to those from adult tissues in the Stockholm-Tartu Atherosclerosis Reverse Networks Engineering Task study (STARNET) [21]. The genotyping and RNAseq of RICHS placenta and STARNET samples were conducted under similar conditions. We surveyed seven tissues: blood, atherosclerotic-lesion-free internal mammary artery (MAM), atherosclerotic aortic root (AOR), subcutaneous fat (SF), visceral abdominal fat (VAF), skeletal muscle (SKLM), and liver (LIV). We randomly selected 150 samples from each STARNET tissues (the same sample size as RICHS placenta set) and computed 10% FDR eQTLs for integration with BW, CO and CBMI GWAS data. We are interested in evaluating the significance level of eSNPs’ enrichment for small GWAS pvalues, which is visualized in Fig 1. We (1) identified the shared SNPs that are in the GWAS study (e.g. BW GWAS) and are also eSNPs (≤10% FDR) in the tissue of interest; (2) conducted LD pruning on the SNP list using a rather stringent threshold (ie, r2≤0.2), which capped the r2 among the pruned SNPs at 0.2; (3) performed Kolmogorov-Smirnov test (K-S test) between the pvalues of pruned SNPs and null distribution. Similar approach was applied to compare magnitude of enrichment between eSNPs of two tissues, where we performed two-sample K-S test on the pruned SNP lists derived from the two tissues (S4 Table). We firstly define “intervals” around GWAS association signals. GWAS SNP of pvalue <1e-3 were selected, and an interval is formed around each SNP by extending 200kb both upstream and downstream. Then we merge overlapping intervals into larger intervals until the remaining intervals no longer overlap. Colocalization analysis was performed within each interval using COLOC version 2.3–6 in R[17]. This method assesses whether two association signals, GWAS summary statistics and eQTL statistics, are consistent with shared causal variant(s) [17]. Default priors of the software were used: p1, prior probability a SNP is associated with trait 1, default: 1e-4; p2, prior probability a SNP is associated with trait 2, default: 1e-4; p12, prior probability a SNP is associated with both traits, default: 1e-5. Intervals were created around each GWAS SNP (with GWAS pvalue<1e-3) ± 200Kb, and overlapping intervals were merged into one interval. Afterwards, COLOC analysis was conducted for each interval, where only eSNPs and transcripts located within the interval entered analysis. In total, five hypotheses were evaluated. H0: No association with either disease risk (i.e., trait 1) or placenta gene expression (i.e., trait 2); H1: Association with trait 1, not with trait 2; H2: Association with trait 2, not with trait 1; H3: Association with trait 1 and trait 2, multiple independent SNPs influencing the two traits; H4: Association with trait 1 and trait 2, one shared SNP. Genes that demonstrated a high posterior probability of hypothesis 4 (PP.H4 >75%) indicate the disease risk and placenta gene expression were controlled by the same genetic variant. LD Score Regression [23] version 1.0.0 was employed to estimate the heritability and genetic correlation of BW, CO and CBMI traits. The European subjects in the 1000Genome data set were used as an LDscore reference. We applied MetaXcan [19] to integrate GWAS summary level data and sample level placenta eQTL data (genotype and gene expression) to identify genes underlying BW, CO and CBMI traits. To reduce the multiple testing burden, we limited the test to genes influenced by placenta eSNPs (≤10% FDR), and with at least one of the genes’ eSNP having a pvalue ≤ 0.01 in the GWAS of interest. For BW, 487 genes were evaluated; for CO, 163 genes were evaluated; and for CBMI, 169 genes were evaluated. Bonferroni correction was applied to adjust pvalues. We took advantage of the RICHS data which measured both birth weight and placenta gene expression levels in the same set of subjects, and validated significant MetaXcan findings (ie, association between BW and imputed placenta transcription levels). Briefly, generalized linear models (glm) were performed on transcripts potentially responsible in controlling BW (identified in MetaXan, Table 2).

10.1371/journal.pcbi.1005001

Isofunctional Protein Subfamily Detection Using Data Integration and Spectral Clustering

As increasingly more genomes are sequenced, the vast majority of proteins may only be annotated computationally, given experimental investigation is extremely costly. This highlights the need for computational methods to determine protein functions quickly and reliably. We believe dividing a protein family into subtypes which share specific functions uncommon to the whole family reduces the function annotation problem’s complexity. Hence, this work’s purpose is to detect isofunctional subfamilies inside a family of unknown function, while identifying differentiating residues. Similarity between protein pairs according to various properties is interpreted as functional similarity evidence. Data are integrated using genetic programming and provided to a spectral clustering algorithm, which creates clusters of similar proteins. The proposed framework was applied to well-known protein families and to a family of unknown function, then compared to ASMC. Results showed our fully automated technique obtained better clusters than ASMC for two families, besides equivalent results for other two, including one whose clusters were manually defined. Clusters produced by our framework showed great correspondence with the known subfamilies, besides being more contrasting than those produced by ASMC. Additionally, for the families whose specificity determining positions are known, such residues were among those our technique considered most important to differentiate a given group. When run with the crotonase and enolase SFLD superfamilies, the results showed great agreement with this gold-standard. Best results consistently involved multiple data types, thus confirming our hypothesis that similarities according to different knowledge domains may be used as functional similarity evidence. Our main contributions are the proposed strategy for selecting and integrating data types, along with the ability to work with noisy and incomplete data; domain knowledge usage for detecting subfamilies in a family with different specificities, thus reducing the complexity of the experimental function characterization problem; and the identification of residues responsible for specificity.

The knowledge of protein functions is central for understanding life at a molecular level and has huge biochemical and pharmaceutical implications. However, despite best research efforts, a substantial and ever-increasing number of proteins predicted by genome sequencing projects still lack functional annotations. Computational methods are required to determine protein functions quickly and reliably since experimental investigation is difficult and costly. Considering literature shows combining various types of information is crucial for functionally annotating proteins, such methods must be able to integrate data from different sources which may be scattered, non-standardized, incomplete, and noisy. Many protein families are composed of proteins with different folds and functions. In such cases, the division into subtypes which share specific functions uncommon to the family as a whole may lead to important information about the function and structure of a related protein of unknown function, as well as about the functional diversification acquired by the family during evolution. This work’s purpose is to automatically detect isofunctional subfamilies in a protein family of unknown function, as well as identify residues responsible for differentiation. We integrate data and then provide it to a clustering algorithm, which creates clusters of similar proteins we found correspond to same-specificity subfamilies.

Despite the best research efforts, a substantial and ever-increasing amount of predicted proteins still lack functional annotation [1]. Indeed, the unprecedented increase in the number of new protein sequences being produced by genomics and proteomics projects, as well as the copious amounts of structures for proteins of unknown functions being solved by structural genomics, directly highlight the need for computational methods to determine, quickly and accurately, the molecular and cellular functions of such proteins, given that experimental investigation is difficult, costly, and time-consuming [2, 3]. As the number of sequenced genomes rapidly increases, the vast majority of gene products may only be annotated computationally [4]. However, no high-throughput approaches currently exist capable of revealing the function of every hypothetical gene in the already sequenced genomes. This goal can only be reached per the efforts of several experimental, structural, and computational biologists [5]. The work presented herein is a computational effort aiming to take a step toward that goal. The commonest protein function annotation approach is homology-based annotation transfer, which assumes proteins sufficiently alike in sequence and structure perform similar functions [3]. Such methods have various limitations due to this assumption [6]. On account of protein function plasticity and of the intrinsic imprecision in related databases, various aspects of function cannot be accurately transferred between similar sequences indiscriminately [7]. In fact, homology-based annotation transfer methods are considered one of the main sources of annotation errors due to an excessively liberal application of function inheritance [3], which fails when similar proteins cannot be identified or when they, too, lack reliable annotations [8, 9]. Moreover, such methods also fail for proteins that have the same function despite being different in sequence and structure (i.e., convergent evolution) [10], and also for those which are sequentially and/or structurally similar yet functionally diverged during evolution [9]. Automatic protein function annotation methods depend on a correlation between functional and sequential or structural similarity measures [11], the simplest of which explores global sequence similarity. Other measures commonly employed in the literature are local sequence motifs, global and local structural similarities, and 3D templates. Since such similarity measures focus on different protein features, one may expect they yield better functional annotations when combined [11]. In fact, literature shows using a single data type (e.g., sequence similarity) is insufficient to precisely annotate protein functions due to the immense amount of factors involved in determining a function, and to the consequent complexity of the automatic annotation problem [4, 7–10, 12–20]. A combined approach is usually more powerful than its individual components [3], so blending various data types is crucial in order to transfer annotations more reliably [21]. This attests to the great importance and need for automatic function analysis methods capable of integrating various data types. Increasing the difficulty, one ought to consider attributing a function to a protein family is further complicated by the fact that many families are composed of proteins with multiple folds and/or functions. In such cases, determining possible subfamilies may lead to important information about a related protein’s function and structure, as well as about the functional diversification acquired by the family during evolution [7]. Therefore, a family of homologous proteins may be divided into subtypes which share specific functions uncommon to the family as a whole [22]. We believe determining such subfamilies to be a first step toward reducing the protein function annotation problem’s complexity. Hence, this work’s purpose is the detection of isofunctional subfamilies in a protein family of unknown function, along with the identification of residues responsible for subfamily differentiation. Various methods have been proposed to identify amino acid conservation patterns which distinguish subgroups in a protein family [22–29]. In general, such methods have the considerable disadvantage that subfamilies must be known a priori. Apart from the scarcity of experimental information about subfamilies, this requirement is prohibitive when working with protein families of unknown function. To the best of our knowledge, a single method in the literature is similar to ours in that it first attempts to identify subfamilies in a Pfam [30] family through clustering and, then, to detect specificity determining residues which characterize them: Active Sites Modeling and Clustering (ASMC) [31], which clusters proteins according exclusively to active site composition. Simply put, given a Pfam family, ASMC first performs homology-based structural modeling of its members with a reference structure, later superposing such models to the structure in order to identify residues aligned to its active site. As a result, it builds a multiple sequence alignment (MSA) that represents the active site composition for each protein in the family. This MSA is then subjected to a hierarchical clustering, generating a tree whose nodes are protein groups and whose levels represent successive subdivisions of the family: the root of the tree has all proteins in the same group, whereas the leaves represent singleton clusters. Afterward, the authors manually cut this tree in order to obtain clusters which they find most interesting. The reported number of clusters in the family is, thus, manually defined. Finally, the authors perform a statistical significance analysis to determine the active site positions which were most important to differentiate among groups. Considering the various challenges to automatic function annotation may be extended to the problem of detecting subfamilies, in order to overcome the previously mentioned major obstacles faced by homology-based methods, we adopt an approach that integrates various data types. For this purpose, the similarity between protein pairs according to different knowledge domains is interpreted as evidence, albeit weak, of functional similarity. We integrate such data types using genetic programming and, afterward, provide it as input to a spectral clustering algorithm Our main goal is to propose a strategy for selecting and combining pieces of functional similarity evidence between protein pairs, and to analyze the manner in which integrating information from different knowledge domains is capable of directing a clustering process to detect, in a protein family of unknown function, isofunctional subfamilies, along with the residues that differentiate them. This goal was successfully achieved. The proposed framework’s capability of using diverse data types, even if incomplete or uncertain, is of remarkable importance for application scenarios such as this, since data from biological experiments are naturally imprecise, mainly due to the dynamic nature of the phenomena investigated as well as to experiment interpretation errors [4], and certain types of information are relatively scarce. Additionally, protein function is determined by various factors, and the complementarity of the different data sources allows for the algorithm to work with as much information as possible. Our main contributions are the proposed strategy for selecting and integrating various data types, along with the ability to work with noisy and incomplete data; the possibility of using domain knowledge for detecting isofunctional subfamilies in a protein family with different specificities or even of unknown function, thus reducing the complexity of the experimental function characterization problem; and the identification of residues responsible for specificity. The proposed framework consists of five main steps, namely definition of the protein set to be studied, collection of pieces of functional similarity evidence, data integration, clustering, and quality evaluation. Once a Pfam family of interest is defined, a filtering process is applied to obtain the protein set to be studied. First, we collect the family’s full sequence alignment from Pfam and extract the UniProt identifiers, together with the subsequences which contain the domain that characterizes the family. The protein set is later filtered by subsequence size, eliminating those whose lengths differ more than a standard deviation from the family average, as done by ASMC [31]. Afterward, we collect the structures associated to the family from PDB, separate the chains, and select the reference structures to be used as templates for structurally modeling the family sequences using Modeller [32], and also to search for pockets that are possible active sites using Fpocket [33]. We prioritize structures obtained by X-ray crystallography, with high resolution and which contain ligands. Next, the protein set is further filtered according to similarity with the reference structures, eliminating those with less than 30% identity to all structures, which is the minimum level accepted by Modeller for creating a structural model. For each sequence, we choose the model with the smallest energy, as done in [1] and [31]. By the end of this filtering process, the database contains the UniProt identifiers, amino acid subsequences containing the family domain, and structural models for all remaining proteins. The steps taken to collect and apply the various data types interpreted as evidence of functional similarity are described next. Any data that may be used to compare proteins pairs can be added to the process. Given the data integration method (i.e., genetic programming) is capable of learning which data types contribute to achieving good clusterings and filter out those of little use, even data types unlikely to be related to functional similarity were included. Each column in the database corresponds to one of the aforementioned similarity matrices, each of which is normalized to [0, 1], or to [-1, 1] in case negative values exist. The data types for which smaller values indicate greater similarity, as is the case for those involving differences or distances, have their intervals reversed. Hence, all matrices may be interpreted in the same way: the higher the value, the larger the similarity between that protein pair according to the corresponding data type. In order to combine such primary similarity matrices into a single matrix to be provided as input to the clustering algorithm, we use genetic programming (GP). GP is a natural computing technique that automatically solves problems without the user having to know or specify the form of the solution. Basically, in each generation, a population of individuals, each of which represents a combination of data types in this work, is stochastically transformed into a presumably better population [47]. The execution ends when a maximum number of generations is reached or when some other stopping criterion is met. Such transformations are accomplished by genetic operators of crossover, reproduction, and mutation, which recombine parts of individuals from one population to create individuals for the next [48]. Crossover works by randomly selecting parts of two individuals and switching them. For mutation, a random part of a single individual is replaced by new code, whereas in reproduction, an individual is selected and copied into the next generation [47]. Genetic operators are usually mutually exclusive, and their probability of application is called the operator rate. Individuals are selected to undergo such operations according to their fitness value. Thus, fitter individuals are more likely to be selected to “breed”, producing new individuals for the next generation. This work’s GP system was implemented using the lil-gp library [48] in C. Starting from random data combinations, it learns, over generations, which ones yield better protein clusters. Each primary matrix is depicted by a variable, so individuals represent equations that combine different matrices through addition. For each individual in the population, the GP system calculates the final similarity matrix by applying its equation to each protein pair, and subsequently runs the spectral clustering algorithm with the calculated matrix, returning the quality of the yielded clustering as the individual’s fitness value. By evolving a population of equations that combine the various data sources, apart from the actual clusters generated, results will allow to check which types of information are most useful to discriminate among groups in a protein family. Eq 1 shows an example of individual which calculates the similarity sij between each protein pair (i, j) by adding the number of InterPro annotations they have in common, their conserved neighborhood score in STRING, and three times the TM-score of their structural alignments. Clustering is a data mining technique which consists in dividing a set of objects into natural clusters, each of which represents a significant subpopulation, so that objects in the same group are very similar to each other, while different from those in other clusters [49]. In this work, we consider partitional clustering methods, in which none of K clusters are empty and each object belongs to a single cluster [50]. Among the various algorithms presented in the literature, we opted for employing spectral clustering, since it is capable of solving very complex problems such as the case that, when plotted, the objects from each cluster are positioned in intertwined spirals, which cannot be separated by something simple as a line or a curve. Such an algorithm was necessary for our application scenario since families of homologous proteins can rarely be separated into subfamilies easily. Spectral clustering uses the eigenvectors and eigenvalues of the similarity matrix to reduce the number of dimensions before performing clustering in the reduced space. First, a similarity graph is built from the inputted similarity matrix. Next, its Laplacian matrix is calculated, along with its eigenvectors and eigenvalues. The eigenvectors corresponding to the K smallest eigenvalues are taken, each as a dimension in the new data set representation. This change in representation from the original space to a K-dimensional space accentuates the cluster properties in the data, so that clusters may be trivially detected in the new representation [51], so much so that a simple clustering algorithm like K-Means may be used. In this work, given the similarity matrix calculated by a GP system individual, we define the adjacency matrix of the totally connected similarity graph, and calculate its normalized asymmetric Laplacian matrix, its eigenvalues and eigenvectors, taking the K first eigenvectors. This new N×K matrix, where N is the number of proteins in the family and K is the desired number of clusters, is then provided as input to the K-Means clustering algorithm. Given our goal of detecting isofunctional subfamilies in a protein family, and considering each cluster is described by an active site composition-based profile, a quality measure which numerically reflects the differences among cluster profiles is required. We consider pointwise mutual information (PMI) [52], which measures the amount of information the occurrence of a specific value x contributes to making the correct classification of an object relative to cluster y [53]. The PMI is a measure of how much the event co-occurrence probability (p(x, y)) differs from expected based on the individual event probabilities and on the independence assumption (p(x)p(y)), and is calculated by Eq 2 [54]. If there exists a genuine association between the values, then p(x, y) ≫ p(x)p(y) and, consequently, PMI(x, y) ≫ 0. If no interesting relationship exists, then p(x, y) ≈ p(x)p(y) and PMI(x, y) ≈ 0. Finally, if x and y are in complementary distributions, then p(x, y) ≪ p(x)p(y), hence PMI(x, y) ≪ 0. PMI is “pointwise” because it is calculated for two values x and y, whereas mutual information (MI) is calculated for two variables X and Y, and corresponds to the expected PMI over all possible values, i.e., MI(X, Y) = ∑x∑y p(x, y)PMI(x, y) [54]. MI measures the information dependence or overlap between two random variables, reaching maximum value when the variables are perfectly correlated [54, 55]. We consider a cluster to be interesting when it has (almost) exclusive residues for specific active site positions. Hence, we compare each cluster to the union of the others. For each position pi, residue rk’s importance for cluster cj is measured by PMIpi(cj, rk), whereas its importance in the union of the remaining clusters (c j ¯) is calculated by P M I p i ( c j ¯ , r k ). This yields MIpi(cj, rk), calculated by Eq 3, in which the ppi(cj, rk) and p p i ( c j ¯ , r k ) probabilities are estimated by residue rk’s frequency in cluster cj and in the other clusters at position pi. Because PMIpi(cj, rk) and P M I p i ( c j ¯ , r k ) values have opposite signs, and considering that we only deem important to a cluster those residues more frequent in it than in the other clusters, only residues for which PMIpi(cj, rk) > 0 are considered. In case the residue also occurs in other clusters, then P M I p i ( c j ¯ , r k ) < 0, and their addition will reduce rk’s importance for cluster cj. Finally, if PMIpi(cj, rk) ≤ 0, we consider MIpi(cj, rk) = 0. For a given cluster, there might be multiple residues in a specific active site position. Hence, considering fk as residue rk’s frequency in cluster cj, we have that MIpi(cj) = ∑k fk MIpi(cj, rk). Gaps are not considered in this calculation. Finally, the quality measure for the clustering as a whole is the overall average, calculated by Eq 4, where P is the total number of positions, and C is the number of clusters. The GP system uses this as fitness function. Thus, it attempts to maximize the mutual information between the active site residues and the clusters, which is equivalent to searching for clusters that present characteristic active site compositions. When a ground-truth exists such as the SFLD family classification, external validation measures may be used to calculate a clustering’s agreement with it. Pairwise measures consider the cluster labels and ground-truth classifications over all pairs of objects. For an object pair with the same ground-truth classification, the objects may be attributed to a same (true positive—TP) or different (false negative—FN) clusters. Analogously, a pair with different ground-truth classifications, may be assigned to a same (false positive—FP) or different (true negative—TN) clusters. The precision (P) for a given clustering is the percentage of object pairs that are in a same cluster and actually have the same ground-truth classification (P = T P T P + F P), while the recall (R) is the fraction of pairs with the same ground-truth classification that were assigned to a same cluster (R = T P T P + F N). The F1 score tries to balance the precision and recall values by computing their harmonic mean, and is calculated as F 1 = 2 × P × R P + R. The Rand index measures the fraction of true positives and negatives over all object pairs, and is defined as R a n d = T P + T N T P + F P + F N + T N. It is symmetric in terms of true positives and negatives, and measures the fraction of pairs where the clustering and the ground-truth classification agree. The Rand index has a value between 0 and 1, with 0 indicating complete disagreement and 1 indicating the clusters are exactly the same as the ground-truth classification. The Jaccard coefficient measures the fraction of true positives when ignoring the true negatives. It is defined as J a c c a r d = T P T P + F P + F N. Since it ignores true negatives, it is asymmetric in terms of the true positives and negatives. Thus, it emphasizes the similarity in terms of the object pairs that belong together in both the clustering and the ground-truth, but discounts the pairs that do not belong together [49]. The larger the values for these measures, the better the agreement of the clustering with the ground-truth classification. Additionally, the variation of information measures the amount of information not shared between the clustering and the ground-truth, and is calculated as VI = H(S) + H(S′) − 2I(S, S′), where H is the entropy of a data partition, and I is the mutual information between two partitions of the same data. Lastly, the edit distance is defined as the minimum number of split or merge operations required to transform the clustering into the ground-truth classification, where a split or merge affecting multiple objects is considered one operation. The edit distance between the ground-truth classification and a clustering, with class k and cluster k’, respectively, is calculated as Edit = 2(∑rk,k′) − K − K′, where rk,k’ equals 1 if class k and cluster k’ have items in common, and zero otherwise. K is the number of classes, while K’ is the number of clusters [64]. The smaller the values for the variation of information and the edit distance, the more similar the clustering is to the ground-truth classification. Since clustering is independent from supervision data such as class labels, the proposed framework may be applied to any protein family, even those of unknown function, as shown in this work. However, in order to evaluate our technique’s performance and to facilitate its comparison with similar literature method ASMC [31], we applied our technique to the same well-known families studied by its authors: nucleotidyl cyclases (Pfam family PF00211), serine proteases (PF00089), and protein kinases (PF00069 and PF07714). The same protein sets and subfamily labels were used, except for the removal of proteins which had since become obsolete in UniProt. We observed ASMC is unstable in terms of the clusters it produces, since this minor update to the protein sets caused the algorithm to yield clusters extremely different from those presented in [31] using the same parameter values. Given the purpose of detecting isofunctional subfamilies in protein families of unknown function, a fourth case study was performed on Pfam family DUF849, to which ASMC has also been applied in [1]. For comparison purposes, we employed the same structural models as in [1] and [31], which were obtained using the template structures listed in Table 2 along with their catalytic residues according to the Catalytic Site Atlas (CSA) [56]. In order to evaluate our technique’s performance against a gold standard, case studies were also performed on the crotonase and enolase superfamilies of the Structure Function Linkage Database (SFLD) [57], which hierarchically divides superfamilies into subgroups and families. The structural templates used for modeling the sequences and their respective catalytic residues according to the CSA are presented in Table 3. The following subsections show results for these six case studies using the experiment configurations described in the S1 Text. The protein sets studied for each family are presented in the S2 Text. The MI values for the best results found in each of five runs of the experiments are presented in the S3 Text. It is noticeable in both [31] and [1] that ASMC is usually employed to provide an initial hierarchical clustering of the protein family, which afterward is manually altered in order to obtain clusters the authors consider most interesting. This manipulation allows for different hierarchy levels to be considered for each tree branch, thus distorting the algorithm’s output. However, in order to compare ASMC to the proposed technique, for protein families nucleotidyl cyclases, serine proteases, and protein kinases, ASMC’s clustering step was rerun with the updated protein sets. For all case studies except for the DUF849 family, the trees produced by ASMC were cut at the first two levels, thus defining the number of clusters in a per-level basis. For the DUF849 family, our results were compared to the seven groups manually produced in [1] by the manipulation of ASMC’s output. As previously discussed, the quality of the resulting clusters, used as fitness function by the GP system, is measured by the MI value, calculated by Eq 4. The larger the value, the better the clustering. Table 4 summarizes the differences among MI values for the clusterings produced by our framework and by ASMC for the same numbers of clusters. The primary similarity matrices, which are combined by the GP system to produce the final matrices provided as input to the spectral clustering algorithm, are denoted by their identifiers previously listed in Table 1. The logos which represent each cluster’s active site composition profile were generated by WebLogo [58]. Their color scheme represents amino acid chemical features: green for polar residues, purple for neutral, blue for basic, red for acidic, and black for hydrophobic. Each column in the logo corresponds to a position in the putative active site, and narrower columns denote the occurrence of gaps. A residue’s height in the logo is proportional to its frequency in the corresponding cluster. Nucleotidyl cyclases are a family of cytosolic or membrane-attached domains that catalyze the transformation of a nucleotide triphosphate into a cyclic nucleotide monophosphate [25]. These proteins have fundamental roles in a wide range of cellular processes, and two functional subfamilies exist, namely adenylate cyclases, which act on ATP to form cAMP, and guanylate cyclases, which catalyze the conversion of GTP to cGMP [59]. Mutations of only two residues (Glu-Lys and Cys-Asp) are sufficient to completely alter the specificity from GTP to ATP [25]. After removing, from the original set, 75 proteins that became obsolete in UniProt, 461 remained in this family, of which 186 are labeled as adenylate cyclases and 275, as guanylate cyclases, according to the labels employed in [31]. Thus, the GP system was run to divide this family into two clusters. However, with the same parameter values used in [31], ASMC produced, for the updated protein set, a hierarchical clustering whose first level divided the family into three clusters, and whose second level divided it into six. Hence, in order to compare results, the GP system was also run with three and six clusters. Table 5 presents the data combinations produced by the GP system which yielded the best results for the nucleotidyl cyclases in five runs for each considered number of clusters. Since the MI is based on active site composition, it was expected that the related similarity matrices would be involved in the best results. Other data types which stood out were the global and local sequence alignments scores (seqAliG and seqAliL), the structural alignment identities (strAliId), and the differences in aliphatic residue content (difAliphRes). One may observe a large amount of data types was required by the GP system to partition the family into two clusters. This Pfam family, defined by the presence of a conserved domain of unknown function, was studied in [1] because it contains the Kce protein, which was of interest to the authors for they had previously discovered an initial association between it and a formerly orphan enzyme activity. Such activity is involved in the lysine fermentation pathway and catalyzes the condensation of β-keto-5-amino-hexanoate (KAH) and acetyl-CoA to produce aminobutyryl-CoA and acetoacetate. Since the DUF849 proteins do not all come from organisms capable of fermenting lysine, this suggests there are various biochemical reactions catalyzed by different family members [1]. Hence, the authors considered DUF849 a good case study for discovering new activities in a protein family of unknown function, and named the family “BKACE”, which stands for β-keto acid cleavage enzyme. The main result presented in [1] is a division of the set of 725 proteins into seven groups, obtained by manually altering ASMC’s hierarchical clustering. This manipulation allows to consider different hierarchy levels for each tree branch, thus distorting the algorithm’s output to build clusters the authors consider most interesting. Group logos are presented in Fig 3. According to the authors, G3 presents five subgroups, and there is high correlation among the distribution of the proteins in the seven clusters and the nature of the compounds they transform, as presented in Table 7. Enzymatic activity distribution in these groups, however, is not as clear as depicted, since there are proteins which showed activity for substrates related to other clusters. Considering the substrates for which activities were tested in [1], S1 Table shows the number of times an activity was detected for each group, considering two repetitions for each test. The distribution, among the manually produced clusters, of the number of enzymes considered active for each substrate, shows the complexity of clustering this family into isofunctional subfamilies due to the promiscuity it presents. The manually defined groups have MI = 14.05. For comparison, the GP system was run to divide the DUF849 family into seven clusters. The best result has MI = 36.51 and uses equation 2ASid + csmDist + 2neighborhood + 2seqAliL + strAliId. Cluster logos are presented in Fig 4, while the residues that most distinguish each cluster are listed in S2 Table. Because this is a protein family of unknown function, result comparison is complicated. However, there is substantial correspondence between the two clusterings: Despite being completely automatic, our framework was still able to obtain clusters very similar to those manually produced in [1], even better concentrating the enzymes active for 4-hydroxybenzoylacetate. The only problem in this case study was the GP system considered it to be more relevant, in terms of active site composition, to break group G7 into two relatively uniform clusters, while the authors in [1] opted not to divide this heterogeneous group since the corresponding enzymes are inactive for all tested substrates. Thus, we have successfully demonstrated the proposed framework’s ability and utility for detecting isofunctional subfamilies in families of unknown function. Protein kinases are enzymes that modify the functions of other proteins by adding phosphate groups usually removed from ATP, covalently binding them to the side chains of Ser, Thr, or Tyr residues [25]. They are one of the largest and most functionally diverse protein families, responsible for controlling the majority of biochemical pathways, performing key roles in regulating metabolic processes, cell differentiation, and proliferation of diverse cell types [60]. The main division in protein kinases is between Ser/Thr and Tyr kinases: Ser and Thr are similar in size and shape, while the reaction chemistry and substrate size are substantially different for Tyr [25]. The majority of kinases act upon Ser or Thr, while others are specific to Tyr, and some act upon all three. It is known that some positions confer specificity, such as subdomain VI, in which consensus sequence RDLKPEN is usually found in Ser/Thr kinases, while RDLAARN is typical of Tyr kinases [25]. After removing from the protein set used in [31] 314 proteins that became obsolete in UniProt, 3,087 remained in this family, of which 2,044 are labeled as Ser/Thr kinases and 1,043 as Tyr kinases, according to the labels employed in [31], in which a subgroup of 235 Tyr kinases labeled as Epidermal Growth Factor Receptors (EGFRs) was also reported. Using the same parameter values applied in [31] for the original protein set, ASMC produced, for the updated set, a hierarchical clustering with three and seven clusters in its first two levels. Given two main subfamilies exist, our framework was applied to divide the family into two, three, and seven clusters. Table 8 presents the data combinations produced by the GP system which yielded the best results for protein kinases. Yet again, the presence of active site-related data is noticeable as expected due to the quality measure employed. Other outstanding data types were structural alignment identities (strAliId) and GO term similarities (go). Proteases are a large enzyme family involved in peptide bond hydrolysis. Almost a third of all proteases are serine proteases, whose name derives from the Ser residue at the active site [61]. Serine proteases are involved in a huge number of biological processes, such as digestion, homeostasis, apoptosis, signal transduction, reproduction, immune response, and blood coagulation [61, 62]. They present a catalytic triad composed of a Ser, an Asp, and a His [31], whose 3D arrangement allows for moving protons in and out of the active site. All serine proteases act through a similar catalytic mechanism, but have different cleavage preferences due to active site changes [25, 31]. For chymotrypsins, the active site is lined with hydrophobic residues, so proteins containing hydrophobic residues such as Leu or Ile form strong bonds in the correct orientation for the triad to act. The cavity in trypsins contains a negatively charged Asp, so their substrates must have a specifically positioned positively charged residue such as Lys or Arg. In turn, elastases have smaller cavities, so only proteins containing short-chained residues such as Gly or Ala can be acted upon [62]. After removing 140 sequences that became obsolete in UniProt since being used in [31], 1,533 proteins remained, of which 43 are labeled as elastases, 26 as chymotrypsins, and 1,464 as trypsins, according to the subfamily labels employed in [31], in which a subgroup of 13 trypsins were found to be kallikreins. When ASMC is run on the updated protein set with the same parameter values used for the original set, the family is not divided. Hence, the main parameter (-C 0.25) was reduced in 0.05 decrements until a value which divided the family was found: 0.15, which yielded a hierarchical clustering with four clusters in its first level, and eleven in its second. For comparison purposes, the proposed framework was run to divide the family into four and eleven clusters. Table 10 shows the data combinations that yielded the best results for the serine proteases in five runs for each considered number of clusters. Again, the active site-based similarity matrices showed strong utility, as expected. The other data type that stood out was the global sequence alignment score (seqAliG). The crotonase superfamily enzymes catalyze a wide range of metabolic reactions. Some have been shown to display dehalogenase, hydratase, and isomerase activities, while others have been implicated in carbon-carbon bond formation and cleavage, as well as the hydrolysis of thioesters [63]. After applying the filtering process described in the “Protein family definition” subsection to the 7,908 crotonase superfamily proteins with known families in the SFLD [57], 2,694 proteins remained, distributed among twelve families, all of which are in the crotonase like subgroup. The superfamily distribution is presented in Table 11. The active site compositions were extracted from structurally aligning the models against reference structure 1MJ3’s active site. Given twelve families exist, the proposed framework was applied to divide the family into twelve clusters. The best result is obtained with equation ASid + seqAliG + strAliId, which yielded a clustering with MI = 52.18. Cluster logos and compositions in terms of SFLD family labels are presented in the S11 Text. The distribution of crotonase families among clusters is presented in Table 12. In comparison with SFLD’s family classification, the clustering produced by the GP system for dividing the crotonase superfamily into twelve clusters presents a Rand index of 0.80, a Jaccard coefficient of 0.44, 93.84% precision, 45.06% recall, an F1 score of 0.61, a variation of information of 0.80, and an edit distance of 26. These values indicate the clustering produced by the GP system is in high agreement with the SFLD family classification, yet this agreement is more related to precision (i.e., pairs that are in the same cluster and actually have the same classification) than to recall (i.e., pairs with the same classification that are actually in the same cluster). Given the enoyl-CoA hydratase, which accounts for 55.94% of the crotonase superfamily, had elements assigned to four different clusters, this greatly impacted the recall. Hence, the results suggest more clusters are required to properly separate the families, due to the existing variation among enoyl-CoA hydratases. When the SCI-PHY classification method was applied to the crotonase superfamily in [64], the authors achieved a variation of information of 1.05, and an edit distance of 32. Although different protein sets were considered for each technique, this suggests the proposed framework outperforms SCI-PHY. Unfortunately, we were unable to properly compare the techniques on a same data set because the studied protein set is not presented in [64]. Enolase superfamily enzymes catalyze the abstraction of the α-proton of a carboxylic acid to form an enolic intermediate. This is mediated by conserved active site residues. Reactions catalyzed by these enzymes include racemization, β-elimination of water and of ammonia, and cycloisomerization. These enzymes have two structural domains: a N-terminal capping domain and a C-terminal TIM beta/alpha-barrel domain, both of which are required for function [65]. After applying the filtering process described in the “Protein family definition” subsection to the 31,182 enolase superfamily proteins with known families in the SFLD [57], 4,791 proteins remained, distributed among six subgroups and twelve families. The superfamily distribution is presented in Table 13. The active site compositions were extracted from structurally aligning the sequence models against reference structure 1MDR’s active site, chosen according to the process described in the Methods section. Given twelve families exist, the proposed framework was applied to divide the family into twelve clusters. The best result is obtained with equation 3APid + go + 2seqAliG, which yielded a clustering with MI = 98.18. Cluster logos and compositions in terms of SFLD family labels are presented in the S12 Text. The distribution of enolase families among clusters is presented in Table 14. Although nine of twelve clusters are pure (i.e., contain a single family), the mixture in clusters IX, XI and XII shows that more clusters are required in order to properly separate the families. As was the case with the serine proteases and crotonases, this is likely caused by family imbalance: the enolase family accounts for 52% of the protein set, and variation among the enolases may dominate the smaller families. Interestingly, despite the mixture in Cluster IX, all three families are in the muconate cycloisomerase subgroup of the enolase superfamily. In comparison with SFLD’s family classification, the clustering presents a Rand Index of 0.87, a Jaccard coefficient of 0.62, 87.30% precision, 67.83% recall, an F1 measure of 0.76, a variation of information of 0.84, and an edit distance of 34. These values reflect the great agreement of the clustering produced by the GP system with the SFLD family classification. The somewhat low recall, however, further suggests more clusters are required to properly separate the families, likely due to the existing variation among the dominating enolase family. When applying the SCI-PHY classification method to the enolase superfamily in [64], the authors achieved a variation of information of 1.37, and an edit distance of 70. Although the protein sets are different, this suggests the proposed framework outperforms SCI-PHY. However, experiments on a same dataset would be required to properly compare the techniques. Unfortunately, we were unable to perform such comparison since the studied protein set is not presented in [64]. The case study with nucleotidyl cyclases showed our technique successfully separated the proteins into its two known subfamilies, whereas ASMC prioritized subgroups of adenylate cyclases, while the majority of the family was put in the same cluster. Only in the second hierarchy level did ASMC create a guanylate cyclase-specific cluster, but at that point it had fragmented the adenylate cyclases into five subgroups, even though there wasn’t much variability in this subfamily. Another successful case study was with protein kinases, for which our framework yielded clusters whose correspondence with the known subfamilies was almost complete, despite some mistakes in the clustering of less than 1.5% of the family proteins. Meanwhile, ASMC ended up prioritizing a cluster of proteins containing multiple gaps, which are unrelated to function, and, even after increasing the number of clusters, it was unable to separate the two subfamilies. The serine protease case study showed the tremendous imbalance between subfamilies to be a challenge for both techniques, since the substantial variability among trypsins, which account for 94.7% of the family, lead the methods to find trypsin subgroups more easily than the small subfamilies. Thus, both techniques were only able to find clusters specific to kallikreins and chymotrypsins after breaking trypsins into several subgroups. This is a data scarcity issue, which we are unable to tackle given we work with complete Pfam families. Nevertheless, the results showed this is not an issue for the MI measure, since the elastase, chymotrypsin, and kallikrein subfamilies could all be considered undersampled in comparison with the trypsin subfamily, and the proposed framework was still able to isolate the elastase subfamily even when considering four clusters, and, eventually, produced chymotrypsin and kallikrein-specific clusters when a larger number of clusters was considered. The subgroups found by our technique were shown to be relevant, since it found a cluster containing exclusively prothrombins, along with a kallikrein cluster larger than the one found by ASMC. Furthermore, some association exists among the trypsin subclusters and the proteins’ species of origin, although the larger subclusters present mixtures, as shown in the S13 Text. A larger number of clusters is required in order to create subdivisions for specific phylogenetic clades. However, as presented in the S14 Text, the analysis of the Enzyme Commission number distribution among clusters has shown the existence of subclusters for the protein kinase and serine protease families is justified by there actually existing more specific classifications than those reflected by the subfamily labels employed in [31]. Thus, the clusters generated by the proposed framework are in accordance with the existing EC number annotations, although experiments with larger numbers of clusters are required in order to create EC number-specific clusters. However, the literature has shown that the EC system is unsuited for use as a ground truth classification due to the annotation errors caused by automatic annotation transference [66–69]. The case studies with well known protein families showed our technique produces clusters that are in better agreement with their division into subfamilies than those produced by ASMC. Furthermore, when there are more groups than subfamilies, our technique tends to produce clusters which are more different from each other than ASMC, which tends to subdivide clusters that are already relatively uniform. A fourth case study involved the DUF849 protein family of unknown function, in which case we compared our technique’s results to the groups defined in [1] by manually altering ASMC’s hierarchical clustering. This proved to be a challenging family, due to the observed promiscuity. Still, the clustering produced by our totally automatic technique showed tremendous correspondence with the manually determined groups, which attests to the proposed framework’s utility and capacity for detecting possibly isofunctional subfamilies, even in families of unknown function. The last two case studies on SFLD’s crotonase and enolase superfamilies were performed in order to analyze the proposed framework’s performance against this gold-standard. For both superfamilies, the GP system was able to create clusterings in great agreement with the ground-truth family classification, seemingly outperforming the SCI-PHY classification method presented in [64]. Unfortunately, a proper comparison of our technique to SCI-PHY was precluded due to the studied protein set not being listed in [64]. As was the case with the serine proteases, the results suggest more clusters are required in order to properly separate the families for both superfamilies, likely due to the existing family imbalance and variation among the dominating families, which impaired a perfect distribution of the twelve families that exist in each superfamily into twelve clusters. Given our goal of finding isofunctional subfamilies, we consider a cluster to be interesting when it contains residues which are (almost) exclusive to its proteins for the different active site positions. Results showed our mutual information-based cluster quality measure successfully reflects this goal, since using it as the GP system’s fitness function caused contrasting clusters to be found. The better agreement of the clusters produced by our technique with the existing subfamilies was also reflected in the larger MI values. Unfortunately, the manner in which the measure is calculated, which involves partial values for each residue, each position, and each cluster, prevents its use for finding the ideal number of clusters in a protein family, since the value decreases as the number of clusters increases. However, such partial values allow us to numerically evaluate what residues, in which positions, most differentiate a cluster from the others. In fact, for the families whose specificity determining positions are known, such residues were in accordance with those considered by our technique as the most important to distinguish a given cluster. It is standard practice in the related literature to perform multiple clusterings of a data set with different numbers of clusters and then choose the “optimal” number of clusters as the one with the best value for a given clustering quality measure. Many different such measures have been tested in this work in order to tackle the problem of identifying the ideal number of clusters in a protein family, such as the internal cluster validation measures Silhouette Coefficient, BetaCV, Normalized Cut, Dunn Index, (Pointwise) Mutual Information among clusters, Relative Entropy, Log-likelihood, as well as countless variations of such measures. However, this has proven to be a major challenge. Since we work with Pfam families, every protein has some degree of similarity with all other proteins in the family. So the clustering quality measure tends to be best when (almost) all proteins are put in a single cluster, thus yielding clusterings in which one cluster consists of the bulk of the family, while the others contain very few proteins. Although subfamilies are known to exist, the so far tested quality measures do not reflect this. Thus, despite our best efforts, we have been unable to find a measure that would allow us to compare clusterings with different numbers of clusters in order to determine the ideal number of clusters in a Pfam family. Hence, at this time, we use the MI to compare clusterings with the same number of clusters, and visually and manually inspect the cluster logos and compositions in order to compare clusterings with different numbers of clusters, as done for ASMC in [1] and [31]. Considering the GP system is capable of learning which data combinations produce good clusterings and filter out those of little use, even data types unlikely to be related to functional similarity were included in this work. Since the cluster quality measure is based on the active site, it was expected that data derived from it would be among the combinations that produced the best clusterings. The active site identities (ASid) are present in the data combinations for all families, while the scores (ASscr) are included for the serine proteases and nucleotidyl cyclases. However, the association of active site data with other data types contributed to improved results. The ones which stood out were the the sequence alignment scores (seqAliG and seqAliL), present for all families except the serine proteases; the GO term similarities (go), present for all but the DUF849 family and crotonase superfamily; the structural alignment identities (strAliId), present for all families except for serine proteases and enolases; and the genomic context data (cooccurrence, coexpression, neighborhood), present for the nucleotidyl cyclases, protein kinases and the DUF849 family. These data are commonly employed, separately, by homology-based function annotation methods, so their presence among the best data combinations is due to the correspondence, although imperfect, of the similarity according to such data types with the functional similarity. Interestingly, our GP system was able to find similar clusterings with very different data combinations. This is likely due to the fact that the studied data types are not independent. Their redundancy made it virtually impossible to reach a conclusion about the semantics of the obtained equations. A correlation analysis is presented in the S15 Text, proving the existence of redundancy among some of the data types used as functional similarity evidence. Examples of highly correlated data types are the global and local sequence alignment scores (seqAliG and seqAliL, 0.96 correlation), the structural alignment scores and sizes (strAliScr and strAliSize, 0.82 correlation), and the active site identities and scores (ASid and ASscr, 0.79 correlation). Such correlations were to be expected given the data type pairs originate from the same sources. However, most data types are highly diversified, given they present low correlation to the others, which indicates they add important information to the clustering process. Indeed, the results showed an overall tendency that using more data types leads to better clusters. Since the best results involved the combination of multiple data types, this confirms our initial hypothesis that the similarity between proteins according to different knowledge domains may be interpreted as evidence of functional similarity. In summary, results showed the proposed framework, which is fully automated, obtained better clusterings than ASMC for nucleotidyl cyclases and protein kinases, in addition to equivalent results for serine proteases and the DUF849 family, whose clustering was defined with manual intervention. In general, the clusters produced by our technique showed considerable correspondence with the known subfamilies and were more relevant than those produced by ASMC, given they show more contrasting differences among each other, whereas ASMC tends to subdivide clusters which are already uniform in comparison to the others. Furthermore, we observed ASMC is unstable in regards to the generated clusters, since the removal of a small number of proteins which became obsolete lead the algorithm to produce extremely different clusters with the same parameter values. Lastly, the crotonase and enolase case studies showed the proposed framework’s ability to create clusters in agreement with a gold-standard classification. In addition, the hierarchical clustering algorithm employed by ASMC prevents occasional errors during the process from being repaired, since once a node in the hierarchy is subdivided, it is not possible for a protein to switch tree branches. Hence, if a subdivision is erroneously made, the error will be propagated to the following hierarchy levels and will never be fixed. The partitional clustering employed by our technique, on the other hand, allows proteins to migrate to a group that becomes more suitable as the number of clusters increases, which is equivalent to switching branches in ASMC’s hierarchy to repair an error. Lastly, although structural information, which is very scarce relative to other data such as sequence information, is a central part of the proposed framework, results showed it is applicable even if a structure is only available for one subfamily. In the serine protease case study, for example, kallikrein and prothrombin clusters were found even though the family sequences were not modeled against structures from these subfamilies. Yet, the proposed framework was still able to detect clusters specific to these subfamilies. Additionally, for the crotonase and enolase superfamilies, active site compositions were extracted from structural alignments of the models against reference structures for the entire crotonase-like subgroup and for the mandelate racemase subgroup, respectively. Still, the GP system was able to find family-specific clusters. Thus, only one structure is required in order to apply the proposed framework to a given protein family. In conclusion, the results presented herein have proven the proposed technique is useful and capable of detecting isofunctional subfamilies in protein families, even for those of unknown function. Hence, it may be widely applied to other protein families for which at least one reference structure is known, as well as altered to include different data types, even if available only for a subset of the studied family, as was the case for the genomic context-based data in this work. This type of framework, which integrates information from diverse and possibly incomplete sources, is of considerable interest for an application scenario such as ours, given that a protein’s molecular function is determined by numerous factors, and that the complementarity of the various data sources allows for the algorithms to work with as much information as possible. A deeper investigation is required into the semantics of the data combinations produced by the Genetic Programming (GP) system. A network of dependent or synonym variables should aid in better comprehending the redundancy among the data types employed as functional similarity evidence. Eliminating redundant data types from the GP system might ease the semantic analysis of the obtained data combinations, as well as improve the quality of the generated clusters. This requires experiments with different subsets of the studied data types. Additionally, we need to investigate the use of phylogenetic information as functional similarity evidence, given that proteins from a same species should subdivide a cluster into different, yet related, functions. This is extremely hampered, however, by data scarcity and the existing imbalance of known proteins among the species of origin. Considering the difficulty in dividing the serine protease family, as well as the crotonase and enolase superfamilies, due to (sub)family imbalance, it would be interesting to apply sampling methods to families in which this occurs, in order to evaluate the technique’s performance in more well-balanced databases. Furthermore, due to the added complexity for promiscuous protein families, we need to investigate the possibility of adapting the proposed framework to using fuzzy clustering algorithms. Unlike partitional clustering, fuzzy clustering would output, for each protein, a level of membership in each cluster. Thus, a protein which performs multiple functions could belong, at the same time, to different clusters. Lastly, further efforts are required in the pursuit of a clustering quality measure appropriate for this application scenario that will enable comparing clusterings with different numbers of clusters in order to determine the ideal number of clusters in a protein family.

10.1371/journal.ppat.1002851

Clonality Despite Sex: The Evolution of Host-Associated Sexual Neighborhoods in the Pathogenic Fungus Penicillium marneffei

Molecular genetic approaches typically detect recombination in microbes regardless of assumed asexuality. However, genetic data have shown the AIDS-associated pathogen Penicillium marneffei to have extensive spatial genetic structure at local and regional scales, and although there has been some genetic evidence that a sexual cycle is possible, this haploid fungus is thought to be genetically, as well as morphologically, asexual in nature because of its highly clonal population structure. Here we use comparative genomics, experimental mixed-genotype infections, and population genetic data to elucidate the role of recombination in natural populations of P. marneffei. Genome wide comparisons reveal that all the genes required for meiosis are present in P. marneffei, mating type genes are arranged in a similar manner to that found in other heterothallic fungi, and there is evidence of a putatively meiosis-specific mutational process. Experiments suggest that recombination between isolates of compatible mating types may occur during mammal infection. Population genetic data from 34 isolates from bamboo rats in India, Thailand and Vietnam, and 273 isolates from humans in China, India, Thailand, and Vietnam show that recombination is most likely to occur across spatially and genetically limited distances in natural populations resulting in highly clonal population structure yet sexually reproducing populations. Predicted distributions of three different spatial genetic clusters within P. marneffei overlap with three different bamboo rat host distributions suggesting that recombination within hosts may act to maintain population barriers within P. marneffei.

Fungal pathogen populations show patterns ranging from globally recombining to endemic and clonal. Among the most genetically and spatially restricted fungi is the highly clonal pathogen Penicillium marneffei, an endemic AIDS-associated pathogen in Southeast Asia. Previous studies have shown that P. marneffei has a pattern of extreme clonality despite the ability to disperse across wide distances and the presence of mating type genes that are required for sexual recombination. In this study we used genetic markers, comparative genomics, experimental data, and spatial models to determine the influence of sex on P. marneffei populations, and we found that although there was substantial evidence of sexual recombination, most of the recombination in natural populations was limited to sexual neighborhoods, amongst genetically similar and spatially close individuals. Based on the results of experiments and spatial models we found support for sex occurring in bamboo rats that are known to harbor P. marneffei and the pathogens sexual neighborhoods. Our study suggests that the high levels of effective clonality and endemicity found in P. marneffei may have more to do with specific host interactions than with an innate inability to generate population genetic diversity through sexual recombination.

Hypotheses of globally continuous populations and strict clonality in putatively asexual microbial pathogens are rarely supported [1], [2], [3], [4], [5]. Instead, genetic approaches detect recombination in microbes regardless of assumed asexuality, and pathogens are surprisingly promiscuous despite strong population genetic structure [6], [7], [8]. In some eukaryotic pathogens spatial structuring is readily attributable to dispersal limitations [9], [10], but many fungal pathogens of humans display extensive spatial population genetic structure despite their ability to disperse via aerosolized spores [11], [12], [13], [14], [15]. Examples that cause extensive morbidity include Cryptococcus neoformans, Coccidioides sp., Histoplasma capsulatum and Penicillium marneffei. These fungi are maintained in natural environmental reservoirs that might contribute to structured populations via local adaptation, and they are thought to be largely clonal. However, fungi have many different mating systems that encompass asexual propagation t7hrough multiple forms of sexual and parasexual recombination, and clonal structure may be arrived at via very different mechanisms [16], [17]. Evidence suggests that population structure in fungal pathogens is strongly influenced by host distributions and extrinsic geographic boundaries [12], [18], [19], [20], [21]. Therefore, the interplay between mating systems, population structure, and host adaptation is a central question underpinning the evolutionary epidemiology of fungal pathogens. Heterothallic mating systems in fungi require physical contact between two isolates containing opposite mating types at the mating-type locus (MAT) in order to undergo sexual reproduction. If no mating partners are present, then sexual reproduction does not occur and the fungus reproduces asexually (but see Lin et al. [22] for evidence of same-sex mating in the otherwise heterothallic fungus Cryptococcus neoformans). In this case, the relative capacities of fungal lineages to disperse and co-occupy environmental niches can drive population-level recombination rates. If strains of opposite mating type do not equally penetrate environments then species recombination rates may be reduced to levels nearing complete asexuality [23]. Previously, it has been shown that the HIV-associated emerging pathogen Penicillium marneffei shows extensive spatial genetic structure at local and regional scales across Thailand [XPATH ERROR: unknown variable "start2".], [24]. Although there has been some genetic evidence that a sexual cycle is possible in P. marneffei, this haploid fungus is thought to be genetically, as well as morphologically, asexual within these populations [25]. In this study, we use comparative genomics, experimental approaches, and population genetic data to identify the role of sexual recombination in maintaining spatial and genetic structure in this infection. We attempt to answer 4 specific questions: 1) Does the P. marneffei genome show evidence of sex? 2) How are populations of P. marneffei genetically structured? 3) Can population structure be reconciled with sex? 4) Do spatial or host factors correlate with population structure and sex? We use comparative genomics to identify genes linked to mating and genomic signatures of mutation bias associated with meiosis, and we experimentally detect recombination in vivo. We expand our collection of population genetic data across southeast Asia to include mating type data and 34 isolates from bamboo rats in India, Thailand and Vietnam, and 273 isolates from humans in China, India, Thailand, and Vietnam. Together these data form a mosaic that reveals some physical and genetic underpinnings of mating in P. marneffei that are linked to patterns of genetic diversity across its known endemic range. We used 84 sexual cycle genes (Table S1) to blast against the NCBI genome sequences NZ_AAHF00000000 (A. fumigatus), NZ_ABAR00000000 (P. marneffei), and NZ_ABAS00000000 (T. stipitatus). We screened transposon families for substitution bias by making BLAST based alignments to determine the dominant form of a functional integrase gene in each family. We counted the type of substitution based on differences of alleles as low as 70% identical to the dominant intact type. We compared gene sequences of the genomic region between slaB and apnB (the genes that flank the MAT idiomorph in related fungi) of strains FRR2161 and FRR3842. We acquired 307 isolates of P. marneffei from humans and vertebrate hosts (bamboo rats), covering the known global range of the fungus. Our study obtained 273 epidemiologically unlinked human isolates of P. marneffei from HIV-AIDS patients covering the time-period 1959 to 2005. Of these isolates, 258 were georeferenced to either the broad geographical region of collection or the patients home address. The remaining 15 isolates were recovered from patients whose infections were diagnosed in non-endemic regions, and no accurate geographical origin could be assigned. In addition to human isolates of P. marneffei, we obtained 34 isolates from the bamboo rats species Rhizomys pruinosis (n = 3), R. sumatrensis (n = 13), R. sinensis (n = 1) and Cannomys badius (n = 17). We also include the type isolate for P. marneffei Segretain et al. ATCC 18224, CBS 388.87, isolated from R. sinensis in 1959 [26]. All isolates were cultured on Sabouraud's agar and DNA extracted as previously described [27]. Subsequently, isolates were genotyped at 21 microsatellite loci using the methods described in Fisher et al. [11], [27]. The presence within each of isolate of the MAT1-1 α box and MAT1-2 high mobility group idiomorphs was determined using the PCR protocol detailed by Woo et al. [25]. Genotypes were analysed using GenAlex 6.0 [28] to determine allelic diversity, genotypic diversity and spatial correlation across regions and the global distribution of P. marneffei. We used the package adegenet and its dependencies in R to conduct spatial PCA and DAPC analyses [29]. To compare our inferred results against a model of a single continuous population structured by a dingle dispersal kernel and mutation rate we used the coalescent based program IBDsim [30]. Additional distribution data for bamboo rats were collected from specimen databases AMNH, FMNH, NMNH, and the GBIF. We used the bioclim layers 1–21 at 30 sec from the world clim database in MAXENT to generate predicted distributions for bamboo rats and P. marneffei genetic clusters. We measured distributional overlap using Schoener's D and a resampling approach [31]. We compared the relative overlap of genetic clusters to host distributions by generating null distributions of D based on resampling of R. sumatrensis and Cannomys (Text S1). Possible parental distances were compared to null distributions generated by choosing isolates randomly that met the genetic distance criteria from the population that met the parental criteria. Five co-housed outbred CD-1 male mice (16–18 g) were inoculated intranasally with 107 spores suspended in 40 µl of PBS. Conidia from two isolates, PM9, a MAT 1–2 isolate from Thailand, and the type strain ATCC18824 (FR2161) were mixed in a 1∶1 ratio to form the inoculum (S9). Serial dilutions of homogenized saline samples were plated (no later than 6 hours after they were removed from the mice) on Sabouraud agar. Colonies were counted after 4 days in 27°C. Individual colonies used for DNA extraction and subsequent genotyping as before [11]. Isolate genotypes were compared to the initial genotypes of the inoculum and genotypes differing from inoculum were confirmed via DNA sequencing. All the clinical studies from which isolates are available were approved by the Wellcome Trust ethics committees at the study sites, in the UK and by the regulatory authorities of the countries involved. All patients or their next of kin gave written informed consent and all patient data are anonymised. This work strictly complied with the animal regulations and guidelines under UK law and was approved by Imperial College's Ethical Review Process (ERP) Committee and the British Home Office. All murine work was carried out in a Biosafety level 3 secure animal facility under licensed approval from the British Home Office. Sexual reproduction leaves an imprint on fungal genomes by maintaining genes required for mating and by generating patterns of mutation and recombination restricted to meiotic processes [32], [33], [34], [35], [36]. Successful mating in fungi requires that a genome contains a functioning series of interconnected genetic pathways [37]. Using a comparative genomic approach we assessed the presence and functionality in P. marneffei of genes known to be involved in sexual development in fungi. First, comparing between strains FRR2161 and FRR3842 revealed that the region between genes slaB and apnB resembled other fungal mating type idiomorphs. A region of complete dissimilarity was flanked by regions that were nearly identical between the strains (Fig. 1A and B). We found homologs for nearly all of the genes needed for a complete sexual cycle in yeast to be present and putatively functional (Table S1). Those genes not detected in P. marneffei were also not detected in Talaromyces stipitatus, a fungus with a complete sexual cycle, and most genes that were absent in those two fungi were also missing in the recently demonstrated heterothallic fungus Aspergillus fumigatus. Although these sex-related genes may be conserved to function in processes other than mating, their presence suggests that P. marneffei has preserved the ability to complete a sexual cycle. We detected another genomic signature of a functional sexual cycle, a type of mutation bias associated with meiosis. Repeat induced point mutation (RIP), a process by which some fungi silence genes involved in mobile genetic element function by preferentially mutating repeated sequences within their genomes, is associated with meiosis [38]. This process results in skewed base pair distributions due to the induced mutations. Using an approach similar to that of Clutterbuck [39], we found evidence of an excess of sliding windows with zero AG and CT dinucleotides and mutation bias in P. marneffei transposon family Ty-1 with a skew towards G to A and C to T transitions (Fig. 2). We also detected a putatively functional RID gene (Locus ID PMAA079888), the only conserved gene so far implicated in RIP [40], [41], [42]. Although the RID gene and the observed mutation bias can be explained by several factors including those acting during mitosis, they point towards a RIP or RIP-like process that is generally considered a feature of sexually reproducing fungi and an overall genomic pattern consistent with sex. Although these genomic signatures could represent relics from a sexual past rather than ongoing sexual recombination within P. marneffei, in the related human pathogenic fungus Aspergillus fumigatus, the discovery of mating type genes and evidence of RIP heralded the eventual description of a full sexual life cycle [35], [43], [44]. Microsatellite allelic diversity was high overall and within localities (Table 1). With the exception of Thai Central and Thai South, populations assigned a priori by locality were significantly differentiated from one another by Wright's FST [45]; this metric ascertains the proportion of genetic variance among geographical regions relative to the total variance. FST values near zero mean that populations are not distinct and variation is shared equally within and between them, while higher FST values mean that more genetic differences occur between populations compared to those within populations. The China and Taiwan populations were most different from the other a priori populations (Table 2). Phylogenetic analysis revealed associations between sampling area and the occurrence of phylogenetic clustering (Fig. 3). Using discriminant analysis of principal components (DAPC) to identify genetic clusters [46], [47], we assigned individuals to 3 clusters based on the Bayesian information criterion. The clusters show some spatial association. Cluster 1 is composed mostly of isolates from central and southern Thailand, Cluster 2 of isolates from China, and Cluster 3 of isolates from northern Thailand (Fig. 3, Figure S1). As expected, we observed a strong pattern of spatial genetic correlation (r2 = 0.41, p<0.01). We also detected significant ‘global’ genetic structure (positive correlation between spatial and genetic distance) but no ‘local’ genetic structure (negative correlation) using spatial principal coordinate analysis [46], [47]. To test for a homogenous neutral process of genetic differentiation we used a spatially explicit coalescent-based simulation of isolation by distance generated with IBDsim [30] to simulate a uniform dispersal/mutation process across our exact sampling scheme. This uniform genetic structure was then compare against our recovered spatial genetic pattern. By controlling for the spatial distribution of our sample sites we are able to determine if the apparent genetic clustering is simply an artifactual product of clustered sampling and a single uniform process of genetic clustering. Because the hypothesised parameter space is nearly infinite, we concentrated on dispersal scenarios that most closely resembled the spatial genetic correlation present in our data, namely, the strength of spatial genetic correlation at the smallest spatial scale and the decay rate of the correlation. Although the simulated datasets largely overlapped with our recovered data we observed important departures between the two. The simulated datasets had a single peak in spatial genetic correlation at the smallest spatial scale and a decay in correlation dependent on the dispersal kernel, a feature of all single population isolation-by-distance models, but the observed pattern had additional peaks in certain distance classes that disrupted the uniform decay (Fig S2 and S3). One peak was composed of distances between individuals belonging to the outer edges of clusters 1 and 3. Another major peak comes at the spatial scale where the outer edges of Clusters 1 and 3 contact with Cluster 2. These results differ from previous results that observed different rates of decay for spatial genetic correlation [24], a feature that probably owes to the limited geographic scope and power of the earlier study. Our data now suggest that the observed genetic clusters are not the result of a process of uniform decay with geographic distance, and that other factors are also driving the heterogeneity observed in our dataset. Linkage disequilibrium was high throughout the sample with an overall of 0.113 (Table S2). We determined the relative frequency of mutation to recombination using the single locus variant approach applied by Fisher et al. [24] and found a mutation to recombination frequency of 0.083 suggesting that mutation is up to 12 times less frequent than recombination across the whole population. When restricted to only bamboo rat isolates, all single locus variants would be due to recombination, while for human-only isolates the ratios are unchanged in comparison to the entire dataset. Average fungal microsatellite mutation rates have been inferred from between 2.80×10−6 and 2.50×10−5 mutations per generation [48], making the inferred recombination rate in P. marneffei between 3×10−5 and 2.50×10−4, a rate about half that observed in wild yeast [49]. This approach only detects single locus recombination events, which may be a minority in eukaryotic populations, while it should detect virtually all mutations that have not otherwise been masked by recombination. However, the method could be strongly biased towards inferring recombination due to convergent mutations in microsatellite length. The measure of minimal recombination (RM), which represents the minimum number of recombination events necessary to explain alleles failing the four gamete test [50] given the arrangement of the alleles in a contig, showed that recombination did occur within contigs (Figure S4). Complete clonality and complete panmixia are rejected for P. marneffei, but similar to previous results the inferred levels of clonality remain among the highest observed for fungi [24]. To explain the high level of clonal structure either recombination must be rare or it must occur largely between closely related individuals. The entire sample population of P. marneffei showed a distribution of mating types that was significantly skewed (p = 0.02 or p = 0.04 when clone corrected) from a ratio of 1∶1 in favour of an overabundance of MAT1-1 alleles, but some local populations were skewed towards MAT1-2 alleles (Table 1). Two of the genetic clusters inferred by DAPC were skewed towards more MAT1-1 alleles, but MAT genes within the central cluster did not differ from a 1∶1 ratio (Table 3). As predicted in work prior to the discovery of MAT loci in P. marneffei, highly skewed MAT ratios would be expected in a predominately asexual population [24]. On one hand, in the absence of sex and selection, MAT genes at an initial frequency of 0.5 are expected to be fixed in a population on average by ln 2(Ne) generations. Alternatively, in a completely sexual population without selection associated with a mating type, MAT alleles would be maintained at frequencies near 0.5 with very limited variance because all individuals in each generation will possess MAT alleles according to a binomial distribution, and there is no opportunity for drift beyond a single generation. MAT allele counts can be used to represent the reduction in effective population size caused by drift in MAT ratios [51], [52], but this assumes a fully sexual population. When sex is limited, the average allele frequencies for populations that do not lose sex and become fixed remain 0.5, but the variance in MAT allele frequency depends on population size and the frequency of sex. Based only on the differences in MAT allele frequencies between clusters and an intrinsic restriction on sex, P. marneffei would have an intrinsic upper bound of sexual recombination frequency at less than 4.5% given a modest population size of 1000. This small level of sex could explain the highly skewed ratio of MAT alleles in Cluster 3 and still accommodate the 1∶1 ratio in Cluster 1 while avoiding any fixation of MAT alleles. However, if the intrinsic sexual recombination rate explained the distribution of MAT alleles it would predict equal frequencies of clone detection across populations. Instead, percent clonality tracks the MAT allele skew, suggesting that sexual recombination in Cluster 3 is reduced relative to Cluster 1 (Table 3). We do note, however, that MAT allele frequencies would not be informative about where sex occurs if unisexual mating occurs in P. marneffei as is known in C. neoformans [22]. Given that recombination occurs in P. marneffei, we wanted to determine the geographic and genetic scope of sex. Out of 43 clonal groups inferred with EBURST [53], six contained both MAT alleles, and four otherwise genetically identical multi-locus microsatellite types contained both MAT alleles (Fig. 4). Otherwise genetically identical isolates that differ only at mating type have also been detected in Cryptococcus gattii populations [54], [55], [56]. Clones with both MAT alleles represent the smallest possible genetic scale of sex, and are unequivocal evidence for recombination. To detect the more divergent recombination events we defined putative recombinants as any genotype that had no unique alleles, yet differed from the most similar genotype for at least three loci. Putative parents or ancestral parents were defined as all isolates that together could complete the multi-locus microsatellite type (MLMT) of the recombinant genotype (Figure S5). This allows us to compare between observed distances of maximal observed recombination against a null hypothesis that any two isolates could recombine. We identified 11 potential recombinants with 54 possible parent genotype combinations. Geographic distance between putative parents was shorter, 382 km, and genetic similarity higher, 60.06% identical, than random potential parents drawn from the entire population, 675 km (p = 0.005) and 49.29% identical (p = 0.025) respectively. The scale of recombination determines the efficacy of adaptation and the adaptive potential of populations. Although recombination across large distances allows generation of greater genetic diversity and more rapid spread of advantageous alleles, it disrupts locally advantageous combinations reducing local ecological genetic correlation. When sex is limited to small geographic distances it can reinforce local adaptation, and when limited to smaller genetic distances can reinforce genomic coadaptation. Together these effects can promote ecological speciation [57], [58], [59]. When ecological adaptation acts to reinforce genetic differentiation, strong correlations between key ecological factors and population distributions will exist [60], [61], [62], [63], [64]. To assess the possibility that ecological adaptation drives population differentiation in P. marneffei we used MAXENT [64] to predict overlap between the ecological niches of the genetic clusters (Fig. 5). Cluster 1, the cluster with the ratio of MAT alleles nearest 1∶1, had the widest predicted range and overlapped with the entire predicted range of Cluster 3, including the predicted range that was not sampled in Myanmar. Cluster 1, Cluster 2, and Cluster 3 isolates are all found in bamboo rats, but 16 of 17 samples from Cannomys badius and 13 of 14 unique genotypes were from cluster 1 and distributed among India, Thai Central, Thai North, and Thai South sampling localities. None of the Cluster 1 isolates were among the 13 recovered from Rhizomys sumatrensis, which were all in Cluster 3. Cannomys badius is relatively more abundant in the western portion of the range of P. marneffei. The predicted distributions of bamboo rats were similar to the IUCN species ranges and had overlap with P. marneffei distribution. Although our spatial sample of Cluster 2 was geographically restricted it was entirely within the distribution of Rhizomys sinensis, a species that has been shown to consistently harbour clinically relevant P. marneffei [65]. The distributions for R. sumatrensis overlapped with Cannomys and Clusters 1 and 3 (Fig. 5). However, using ENMTools [66] to account for sampling error we found that Cluster 1 predicted distributions overlapped more with the Cannomys distribution than R. sumatrensis distribution, and Cluster 3 similarly overlapped more with R. sumatrensis than Cannomys distribution (S8). This observed range overlap supports a host specific effect on P. marneffei population structure. Hosts may structure populations of pathogenic fungi in many ways, including by providing an environment in which recombination can occur and by acting as a selective filter on population genetic diversity [67], [68]. We used a murine inhalation model of co-infection with genetically distinct strains to investigate the effect of host infection on P. marneffei (Text S2). Isolates of different mating types were used for experimental co-infection of 5 mice. Subsequent culture after 15 days from the livers showed a strong bias towards recovery of the MAT1-1 genotype for each of the mice. However, in two mice, genotypes of 4 isolates recovered from co-infections also revealed infrequent transfer of alleles between isolates of different mating type and genetic cluster (Table 4), suggesting that recombination may be possible across genetic barriers if multiple strains are within a host. In a smaller but similar in vitro experiment we did not observe significant bias towards MAT1-1, and from our scan of partial genotypes we did not recover any recombinants (S9, Table 4). We do not rule out regular recombination outside of hosts, but in the context of our spatial genetic evidence, the result of experimental infections indicate that hosts may play an important role in the development of sexual neighborhoods in populations of P. marneffei. However, the evolution of that role may involve restricted mating with or without host adaptation and remains to be explored. Asexual spores are common in vitro and likely a feature of natural P. marneffei populations, but sexual recombination may be an unexpectedly common occurrence in natural populations. The evidence supports the occurrence of recombination and perhaps even frequent sex, yet the natural populations remain strongly clonal and spatially structured. Although many mycologists might perceive this as a paradox because clonality is usually used as a proxy for asexuality, many fungi, including key pathogens, also employ same clone mating or sibling mating [7], [16], [69]. Three key hypotheses could explain the perceived clonality in P. marneffei; 1) Spatially restricted dispersal keeps individuals in contact with only closely related individuals; 2) Genetic incompatibility between dissimilar individuals restricts sex to genetically similar individuals; 3) Local adaptation restricts the ability of dissimilar genotypes to penetrate habitats ensuring mating between genetically similar individuals. All three are likely to be partially correct. Although the genetic evidence shows spatial limitations to effective dispersal, the physical dispersal of airborne conidia is not likely to be a limiting factor, and four genetically identical clones are dispersed across distances over 800 km. We have little information about the effect of genetic similarity on mating success in P. marneffei, but genetic restrictions on successful recombination are present in some plant pathogens [63], [70], [71] and should not be completely discounted. Local adaptation is not fully supported by ecological niche models that show overlap between distinct genetic clusters, but there is limited evidence of host specialisation. A key question unanswered in all of these hypotheses is why have sex at all? Previous work has focused on the consequences of selectively neutral loss of sex in P. marneffei [72], but the persistence of a sexual cycle in P. marneffei despite abundant asexual reproduction in the lab suggests that there is a selective advantage for sex not associated with the advantages of greater adaptive potential provided by outcrossing. One major consequence of sexual clonality is release from Muller's Ratchet compared to asexuality [51]. Large numbers of haploid offspring and wide dispersal maximize environmental exposure of genets and increase the efficiency of purging deleterious alleles [73], [74]. However, the sexual process itself may also reduce the accumulation of deleterious mutations independent of recombinational effects [75]. Among the close relatives of P. marneffei in the subgenus Biverticillium, outcrossing has not been shown to occur, but self-fertility is common [76], [77], and the distribution of mating systems in the subgenus suggests that inbreeding may not reduce the evolutionary longevity of this group [78]. Another compelling scenario favouring sex recognises the opportunity presented by mating itself for dramatic shifts in morphology and physiology. A predominant view in the fungal literature is that sex occurs in otherwise mostly clonal fungi in response to stressful conditions [16], [79]. Sexually produced spores are often viable for long periods of time and are resistant to extreme environmental conditions [78], [80], [81], [82]. Regardless of the costs and benefits of recombination, P. marneffei might withstand stress by mating when it would otherwise not survive. If that were the case, recombination in P. marneffei might be clustered in space or time where or when stress occurs. Unfortunately, little is known about the natural ecology of P. marneffei, and any conditions that might allow mating to occur are unknown. Isolates are commonly recovered from bamboo rats, yet the epizoology of the fungus is poorly known including unknown routes of infection and unknown course and outcomes of the zoonosis. In other dimorphic fungi including Histoplasma and Blastomyces, mating in natural populations is also poorly known, but in these species it mating has long been studied outside of the host and at lower temperatures in vitro [83], [84]. Nevertheless, the association with small mammals may be the best starting point for a search for the natural sexual niche of P. marneffei. Cryptic mating and inbreeding in P. marneffei has some parallels with other fungal and non-fungal eukaryotic pathogens [7]. There is growing support for high inbreeding in addition to asexual reproduction in Leishmania brasiliensis [85]. Experimental data support a role for within-vector recombination and it is thought that this restricted recombination in Leishmania results in sexual neighbourhoods of pathogen genotypes with high differentiation at multiple spatial scales [86]. In the malaria parasite Plasmodium falciparum, high inbreeding has been linked to faster emergence of drug resistance in some low infection intensity regions, but inbreeding has also emerged as a general property of P. falciparum populations regardless of infection intensity [87], [88]. In Toxoplasma gondii and Sarcocystis neurona, ‘clonal’ emergence is enabled by high rates of selfing and results in spatially structured populations [89], [90]. The most common fungal infection of humans, Candida albicans, also undergoes same-sex mating that can facilitate inbreeding and the ability to accelerate evolution via sex within clonal populations [91]. In C. neoformans, multiple ecological niches where recombination between strains with opposite mating type occurs have been found [92], [93]. However, most mating in C. neoformans globally is likely to occur between strains of the same mating type, and although there is as yet no indication of what factors alter the probability of this kind of inbreeding across natural populations, it may have facilitated the global emergence of a single mating type and highly clonal populations [15], [22], [94], [95]. Although some of the mechanisms underlying these inbreeding eukaryotic pathogens remain mysterious and likely differ between organisms, there is an emerging consistent pattern of clonality resulting from inbreeding rather than strictly asexual propagation even in the absence of a recognized sexual stage.

10.1371/journal.pcbi.1006196

Simulations to benchmark time-varying connectivity methods for fMRI

There is a current interest in quantifying time-varying connectivity (TVC) based on neuroimaging data such as fMRI. Many methods have been proposed, and are being applied, revealing new insight into the brain’s dynamics. However, given that the ground truth for TVC in the brain is unknown, many concerns remain regarding the accuracy of proposed estimates. Since there exist many TVC methods it is difficult to assess differences in time-varying connectivity between studies. In this paper, we present tvc_benchmarker, which is a Python package containing four simulations to test TVC methods. Here, we evaluate five different methods that together represent a wide spectrum of current approaches to estimating TVC (sliding window, tapered sliding window, multiplication of temporal derivatives, spatial distance and jackknife correlation). These simulations were designed to test each method’s ability to track changes in covariance over time, which is a key property in TVC analysis. We found that all tested methods correlated positively with each other, but there were large differences in the strength of the correlations between methods. To facilitate comparisons with future TVC methods, we propose that the described simulations can act as benchmark tests for evaluation of methods. Using tvc_benchmarker researchers can easily add, compare and submit their own TVC methods to evaluate its performance.

Time-varying connectivity attempts to quantify the fluctuating covariance relationship between two or more regions through time. In recent years, it has become popular to do this with fMRI neuroimaging data. There have been many methods proposed to quantify time-varying connectivity, but very few attempts to systematically compare them. In this paper, we present tvc_benchmarker, which is a python package that consists of four simulations. The parameters of the data are justified on fMRI signal properties. Five different methods are evaluated in this paper, but other researchers can use tvc_benchmarker to evaluate their methodologies and their results can be submitted to be included in future reports. Methods are evaluated on their ability to track a fluctuating covariance parameter between time series. Of the evaluated methods, the jackknife correlation method performed the best at tracking a fluctuating covariance parameter in these four simulations.

Time-varying connectivity (TVC) is being applied to an increasing number of topics studying the brain’s networks. Topics that have been explored with TVC include development [1], various pathologies [2, 3], affect [4], attention [5], levels of consciousness [6], and temporal properties of the brain’s networks [7–9]. There are many concerns raised regarding methodological issues. These issues span biased variance [10, 11], movement artefacts [12], and appropriate statistics [13, 14]. Methods used to derive TVC estimates are as diverse as its range of applications. Examples of different methods include: the sliding window method, sometimes tapered [15], multiplication of temporal derivatives [16], methods using Euclidean distance between spatial configurations [8], k-means clustering methods [7, 17], eigenconnectivities [18], point process methods [19, 20], Kalman filters [21, 22], flexible least squares [23], temporal ICA [24], sliding window ICA [25], dynamic conditional correlation [26], phase differences [27] wavelet coherence [4], hidden Markov models [28], and variational Bayes hidden Markov models [29]. This list of TVC methods is not exhaustive, and even more methods can be found in the literature. While these methods and their applications may offer new insights into the functions of the brain and cognition, it becomes difficult to compare results when different studies use different methods to estimate brain dynamics. Each method is often introduced and evaluated by the authors’ own simulations, empirical demonstrations, and/or theoretical arguments. However, apparent differences in time-varying connectivity in different studies may have been influenced, or even caused, by differences in the underlying methodology used to derive connectivity estimates. In order to maximize reproducibility of reported findings, it is important that comparisons of proposed TVC methods can be made with a common set of simulations. To this end, we have developed four simulations that aim to show how well results from different TVC methods correlate with each other and evaluate their performance of tracking time varying covariance. The proposed methods and simulations are included in the Python package tvc_benchmarker, (available at www.github.com/wiheto/tvc_benchmarker). Researchers can evaluate their own TVC methods in tvc_benchmarker. The software also allows for new methods to be submitted to us for inclusion in future reports. Here we demonstrate the functionality and results obtained by tvc_benchmarker by evaluating the performance of the following five methods: sliding window (SW), tapered sliding window (TSW), spatial distance (SD), jackknife correlation (JC), and multiplication of temporal derivatives (MTD). All methods for TVC derivation were implemented in Teneto v0.2.7b [8]. Bayesian statistics for evaluating performance of TVC methods were calculated in PyMC3 V3.1 [30], simulations and analysis were done using Numpy V1.13.1 [31], Scipy V0.19.1 [32], and Pandas V0.19.2. Matplotlib V2.0.2 [33] and Seaborn V0.7.1 [34] were used for figure creation. As discussed in the introduction, the list of published TVC methods that are designed to be applied to fMRI imaging data is long. In an ideal world all methods will be contrasted under the same conditions such that an evaluation of those methods that give appropriate results can be performed. However, it was not our intention to provide a complete comparison of all published methods. Instead we have made all simulation tools freely available so that researchers can evaluate their own TVC methods. Before describing the simulations and the results, we provide a brief overview of the five methods that are evaluated in this article. This section provides an overview of the simulations that are conducted and the general methodology used. See each simulation’s subsection in the results section for full details of each simulation. To compare accuracy and performance for the five TVC methods, we performed four different simulations. The first simulation investigated the similarity of the different TVC methods by correlating their respective connectivity estimates. The second simulation targeted how well the different methods were able to track a fluctuating covariance parameter. The third simulation tested how robust the estimated fluctuating covariance is when the mean of the time series fluctuates, mimicking the haemodynamic response function. The forth simulation considered whether TVC methods can accurately track abrupt changes in covariance. All simulations considered two time series each consisting of 10,000 samples generated from multivariate Gaussian distributions. At each time point, the covariance between the time series could vary (see below). A full account of all model assumptions made as well as a justification for our model parameter settings for the four simulations models used in the present study are given in S2 Appendix. Simulations 2, 3, and 4 all consisted of a fluctuating covariance parameter (rt) that was used to generate the covariance between the time series. TVC methods were evaluated based on their ability to track the rt parameter. How rt was generated could vary for different simulations. In simulation 2, rt varied throughout the time course based on a normal distribution. The simulation was run multiple times allowing for different autocorrelation of rt through time. In simulation 3, rt varied in the same way as simulation 2 but it was applied to time series that had a non-stationary mean that mimicked a HRF. This simulation was also run multiple times with different autocorrelations. In simulation 4, rt varied based on two different “states” that lasted for varying amounts of time. This method was run two times when states could be short (2-6 time points long) or long (20-60 time points long). By evaluating the correlation of different TVC methods with each simulation’s rt, we can evaluate which time varying properties a method is sensitive to. Simulation 1-3 have all their parameters justified on empirical data in S2 Appendix. Simulation 4 has its state lengths based on what has been identified by different TVC studies. It is important to stress that these different state lengths may have been identified due to the methods which were used and may not reflect real dynamic properties. In principle, it is possible to simply correlate the results from the different TVC methods with the rt values of each simulation to statistically evaluate their performance. However given the inherent, but known, uncertainty in rt, we deemed it was appropriate to create a statistical model which accounts for this uncertainty. Thus, for each TVC method, a Bayesian statistical model was created to evaluate the relationship between the TVC estimate and the signal covariance. The Bayesian model aims to predict y, which is the vector of the known sampled covariances (i.e. rt) with x, which is the connectivity estimate for each TVC model. All TVC estimates and the values of rt were standardized prior to calculating the models with a mean of zero and standard deviation of one. This was done to facilitate the interpretation of the posterior distribution parameter β. The different TVC methods vary in the number of time points estimated (e.g. the beginning and end of the time series cannot be estimated with the sliding window method). In order to facilitate model comparison between methods, we restrained the simulations to include only the time points that had estimates from all TVC methods (i.e the limit was set by the SW and TSW methods which can estimate the covariance for 9,972 out of 10,000 time points). The statistical models were estimated through 5,500 draws from a Markov Chain Monte Carlo (MCMC) with a No-U-Turn Sampler [38] sampler implemented in pymc3. The first 500 samples were burned. The statistical models for the different TVC methods can be contrasted in two ways: (1) model comparison by examining the model fit; (2) by comparing the posterior distribution of β for the different TVC methods. To evaluate the model fit, the Watanabe-Akaike information criterion (WAIC, [39]) was used. The posterior distribution of β illustrates the size and uncertainty of the relationship between x and y. To aid the interpretation of these results for readers unfamiliar with Bayesian statistics, the mode of the distribution corresponds approximately to a maximum-likelihood estimated β value in a linear regression (if uniform priors are used for the parameters the posterior mode and the maximum-likelihood estimator would have been exactly the same). In simulation 1, the different TVC estimates are compared with each other to evaluate how similar these estimates are. To do this, a Spearman correlation is used to evaluate the relationship. The first simulation aimed to quantify the similarity of the different TVC time series estimates. If two TVC methods are strongly correlated, this is a positive sign that they are estimating similar aspects of the evolving relationship between time series. A negative correlation between two methods would suggest that they do not capture the same dynamics of the signal. In this simulation we created two time series (X), each consisting of 10,000 time points in length. The time series were constructed by: X t = α X t - 1 + ϵ (7) The autocorrelation with lag of 1 is determined by αXt−1 and the covariance at t is determined by ϵ. ϵ was sampled from a multivariate Gaussian distribution (N): ϵ ∼ N ( μ , Σ ) (8) where μ is the mean and Σ being the covariance matrix of the multivariate Gaussian distribution. Both time series were set to have a mean of 0, variance of 1 and a covariance of 0.5. In summary: μ = 0 , 0 Σ = ( 1 0 . 5 0 . 5 1 ) (9) The autoregressive parameter α controls the size of the autocorrelation in relation to the preceding time point (i.e. the proportion of the previous time point that is kept). Here, it was set to 0.8 which was deemed to be an appropriate degree of autocorrelation for BOLD time series (see S2 Appendix). A portion of the two simulated time series is found in Fig 2A together with the plots of their respective autocorrelation (Fig 2B and 2C) and a plot of the correlation between the two time series (Fig 2D). The resulting connectivity time series for the different TVC methods when applied to the simulated data is shown in Fig 3. From Fig 3, several qualitative observations can be made about the methods. Firstly, there was a very strong similarity between the SD and JC methods, despite the fact that they consist of quite different assumptions. Further, the SD, JC, and MTD methods were all able to capture considerably quicker transitions than the SW and TSW methods. The long window lengths (SW-29 and TSW-29) were smoother than the SW-15 and TSW-15 methods. Finally, the variance of the JC method was considerably smaller than all other methods, illustrating the variance compression as previously discussed. To assess the degree of similarity of the estimates of functional connectivity time series obtained from all TVC methods, a Spearman correlation was computed for each TVC method pairing (Fig 4). The connectivity time series estimates from all methods correlated positively with each other (Fig 4). Some methods showed strikingly strong correlations (SD & JC: 0.976; SW-15 & TSW-15: 0.999; SW-29 & TSW-29: 0.978). Between the different window lengths the correlation was slightly smaller (SW: 0.644; TSW: 0.755). The lowest correlation was found between the JC and MTD methods (ρ = 0.138). The results from Simulation 1 showed that the connectivity estimates provided by the tested methods are, to a varying extent, correlated positively with each other. It also illustrated how the different methods differ in their resulting smoothness of the connectivity time series. The results from this simulation cannot validate whether any TVC method is superior to any other, it merely highlights which methods produce similar connectivity time series. In Simulation 1, it was not possible to evaluate how well the different TVC methods perform. To evaluate the performance, the simulated data must change its covariance over time and how this changes must be known beforehand. The aim of this simulation was to see how well the derived TVC estimates can infer the covariance that the data was sampled from when the covariance is fluctuating. Two time series were generated (X). Each time point t is sampled from a multivariate Gaussian distribution: X t ∼ N ( μ , Σ t ) (10) where the covariance matrix was defined as: Σ t = ( σ r t r t σ ) (11) and where the variance, σ = 1, was set to 1. At each time point, rt was sampled from another Gaussian distribution: r t ∼ N ( μ r , σ r ) (12) The mean of the time series (μ) was set to 0, the mean of the covariance (μr) was set to 0.2. The simulation was run three times where the parameter for the variance of the fluctuating covariance (σr) was set to three different values {0.08, 0.1, 0.12}. This ensured that the different TVC methods are robust to different variances of connectivity changes. The covariance at time (rt) was sampled from a Gaussian distribution. Each time point received a new value of rt. This allowed us to compare each TVC method’s connectivity estimate in relation to the time varying covariance parameter rt. Note, that at each time point the relationship between the two time series is dictated by a single realization from a Gaussian distribution where rt is the covariance. Thus, we should not expect the connectivity estimate from any method to correlate perfectly with rt. However, it is possible to compare which method correlate better or worse with rt to evaluate the overall performance. The above model will have a temporally fluctuating covariance. It fails to include any autocorrelation in the time series. Not accounting for this may bias the results for some of the tested methods that utilize nearby temporal points to assist estimating the covariance. Merely adding an autocorrelation, like in Simulation 1, will also increase the covariance between the two time series and this will not be tracked by rt. To account for this, we placed a 1-lag autoregressive model for the fluctuating covariance at rt: r t = α r t - 1 + ϵ (13) ϵ ∼ N ( μ r , σ r ) (14) Where α is the autocorrelation parameter. The values for μr and σr were the same as above. When t = 1, ϵ was set to 0. This revised formulation of our simulation model allowed for the covariance to fluctuate, but with an added autocorrelation on the covariance parameter. In simulation 2, three different settings of the parameter α were used (α = 0, 0.25, 0.5). When α = 0 it is equivalent to the original model outlined above with no autocorrelation. With an increased α it entails a greater influence of the covariance from t − 1 in sampling the covariance at t. α = 0.5 is reasonable given highly correlated BOLD time series. An α = 0 is more to be expected when time series are less correlated. 10,000 time points were sampled for each of the three different settings of the autocorrelation parameter. See also S2 Appendix for a justification of the parameter settings chosen here based on empirical fMRI data. Simulation 2 was run with 9 different simulation parameter combinations: three different values of α and three different values of σr. A sample of time series generated with the model using different settings for the autocorrelation parameter α is shown in Fig 5A, 5D and 5G. Due to the varying degree of autocorrelation, the mean covariance for time series changes as a function of α, but rt still depicts a Gaussian distribution (Fig 5B, 5E and 5H). The degree of crosscorrelation between the two time series followed the specified α parameter for the autocorrelation of the covariances (Fig 5C, 5F and 5I). The results from Simulation 2 are shown in Tables 1–3 (for σr = 0.1) and Tables A-F in S3 Appendix (for σr = 0.08 and 0.12). The JC method had the lowest WAIC score for all settings of α, followed by the SD method. The MTD method came in third place for all but one parameter configurations. All WAIC values, their standard error and Δ WAIC scores are shown in Tables 1–3. The posterior distribution of the β parameter for each of the TVC methods for all parameter choices are shown in Fig 6 when σr = 0.1 (for other values of σr see Figs A-B in S3 Appendix). Larger values in the β distribution for a method (i.e. correlating more with rt) conforms with the best fitting models (i.e. lower WAIC score). The SW-15, SW-29, TSW-15, TSW-29 and MTD methods performed equally poor when α = 0, and all improved as α increased. The MTD method improved the most as the α value increased, followed by the TSW-15 and SW-15 methods. SD and JC showed the best performance, with similar posterior distributions of β, although the JC was always slightly higher. There was little difference between the methods when changing the variance of the fluctuating covariance (σr) (See S3 Appendix). The β values do however scale when σr changes. When σr is smaller, β values decrease due to there being more uncertainty when sampling each realization from similar distributions. At times parts of the posterior distributions of the SW, TSW and MTD methods were below 0 to the extent that they would be not classed as “significant”. For example, these methods performed worst when σr = 0.08 and α = 0. Here the percentage of the posterior distribution above 0 was: SW-15: 80%, SW-29: 47%, TSW-15: 84%, TSW-29: 54%, MTD: 89%. The JC and SD methods always had the entire posterior distributions above 0. In sum, the JC method, followed closely by the SD method, showed the best performance in terms of tracking a fluctuating covariance between two time series as performed in Simulation 2. The MTD method ranked in third place when there is a higher crosscorrelation between the time series present. The SW and TSW methods showed the worst performance, both in the WAIC score and posterior distributions of β. The aim of Simulation 3 was to examine the behaviour of different TVC methods when there were non-stationarities present in the data. A typical scenario when this will occur is in a TVC analysis in task fMRI. Simulation 3 is identical in structure to Simulation 2 apart from the following two changes: (1) A non-stationarity, aimed to mimic the occurrence of an event related haemodynamic response function (HRF). Specifically μ, which was set to 0 for both time series in Simulation 2, received a different value at each t (see next paragraph). (2) σr was set to 0.1 instead of varying across multiple values. This is because Simulation 2 showed no large differences when varying σr. μt was set, for both time series, according to the value of a simulated HRF, that was twenty time points in length and repeated throughout the simulation. The HRF was simulated, with a TR of 2, using the canonical HRF function as implemented in SPM12 using the default parameters [40]. This HRF, which has a length of 17 time points, was padded with an additional 3 zeros. The amplitude of the normalized HRF was multiplied by 10 to have a high amplitude fluctuations compared to the rest of the data. μt is thus the padded HRF repeated throughout the entire simulated time series. This represents a time series that includes 250 “trials” that each lasts 40 seconds. This simulation helps illustrate how well TVC methods could be implemented in task based fMRI. Examples of the time series generated using different autocorrelation are shown in Fig 7. The results from Simulation 3 are shown in Fig 8 (posterior distributions of β) and Tables 4–6 (model fit) which evaluated each TVC’s method performance at tracking the fluctuating covariance (rt). Results were similar with Simulation 2. In the case when the autocorrelation of the covariance was 0, the SW, TSW and MTD methods performed quite poorly, but again all improved to varying degrees as this increased. The longer windows (SW-29 and TSW-29) methods were generally the worst method, followed by shorter sliding window methods (SW-15 and TSW-15). The MTD method came in third place. The JC method has the best performance, followed closely by the SD method, in all parameter conditions. When α = 0, some methods had only portions of their posterior distribution above 0 (SW-15: 73%, SW:-29: 30%, TSW-15: 78%, TSW-29: 65%, MTD: 84%). The JC and SD methods had 100% of their distributions above 0 for all parameter conditions. In sum, the results from Simulations 2 and 3 suggests that the JC method has the best performance in terms of detecting fluctuations in covariance compared to the other four TVC methods. This result also holds when a non-stationary event related haemodynamic response was added to the mean of the time series. Simulation 4 aimed to test how sensitive different TVC methods are to large and sudden changes in covariance (i.e. changes in “brain state”) that previously have been postulated to exist in fMRI data (e.g. [11, 15, 17]). We here start in a similar fashion as we did in Simulation 2 where samples for the two time series are drawn from a multivariate Gaussian distribution X t ∼ N ( μ t , Σ t ) (15) Σ t = ( σ r t r t σ ) (16) Similar to simulation 2, we set μt = 0 and σ = 1. The covariance parameter rt was sampled from a Gaussian distribution where the mean was shifted r t ∼ N ( μ state t , σ r ) (17) and where σr = 1. At each state transition, μ state t was randomly chosen from a set M (M = {0.2, 0.6}). The duration of each state was randomly sampled from L. Two different scenarios for state transitions were simulated. In the fast transition condition L = {2, 3, 4, 5, 6} and in the slow transition condition L = {20, 30, 40, 50, 60}. These values correspond to the number of time points a “state” lasts. Beginning at t = 1, μ state t to μ state t + l was randomly sampled from M where l was sampled from L. This procedure was continued until Xt was 10,000 samples long. These choices for brain state changes provide time scales of state transitions between 40-120 seconds (slow condition) or 4-12 seconds (fast condition) in simulated fMRI data with a TR of 2 (Fig 9A and 9D). The statistical model for evaluating the different TVC methods performance was the same as Simulation 2 and 3. A summary of data generated in Simulation 4 is shown in Fig 9. The results from Simulation 4 are shown in Fig 10 and Tables 7 and 8. In the quick transition condition, the JC and the SD showed the best performance for both the WAIC scores and the posterior distribution of β (Fig 10A; Table 7). This was followed by the SW-15 and TSW-15 methods. In the slow transition condition the two sliding window methods outperformed the other methods (Fig 10B; Table 8), with the longer windows (TSW-29 and SW-29) being outperforming the shorter windows. The JC and SD methods perform similarly for both conditions. Thus, when there are shifts in covariance that occur relatively slowly, the sliding window methods are sensitive at tracking these changes. All methods had 100% of their posterior distributions above 0. In this study we have developed four simulations to test the performance of different proposed time-varying connectivity methods. The first simulation showed which methods yield similar connectivity time series. Notably, all methods correlated positively with each other, but to a varying degree. The second simulation generated data in which the autocorrelated covariance between simulated time series varied in time. In this case, the JC method, followed closely by the SD method, showed the best performance. In the third simulation, the generated time series contained a non-stationary mean related to haemodynamic responses. Again, our simulations suggested that the JC method performed best. The fourth simulation included nonlinear shifts in covariance (in an attempt to simulate brain state shifts). When the states changes were quick, the JC method performed best. When the state changes were slow, the TSW (followed by the SW) performed best. In a previous simulation that evaluated the sliding window method, the sensitivity of the SW and TSW methods was found to be good at detecting state shifts [41]. Here, at least when the transitions are slow, we found similar results. The sliding window methods is optimal if there are slow state changes. However it is unclear if “state changes” are the best yardstick for time-varying connectivity. In particular, non-stationarities in time-varying connectivity have been attributed to spurious sources such as movement [12]. Given the unknowns of the “true” connectivity, methods which are robust over conditions are more likely the safer options—in this case the JC or SD method performed similarly in both conditions. However, as mentioned in the methods section, the SD method tested here is the bivariate version of the method and not the multivariate version previously proposed in [8] (see also S1 Appendix for more the relationship between these methods). Overall the jackknife correlation method performed the best across all simulations. We have shown it to be robust to numerous changes in parameters. However, the JC method is not without some considerations. First, it introduces variance compression that reduces the absolute variance, while preserving the relative variance within the time series. This variance compression also scales with the length of the time series. The consequence of this is that direct comparisons of the TVC variance between cohorts/conditions become hard to interpret as time-varying fluctuations, especially when the length of the data varies. However, this is the case for most methods and it should be remembered that the variance is proportional to the static functional connectivity [7, 9, 10]. Simply put, the JC method (like all other methods) should not be used for a direct contrast of the variance of TVC time series. Second, the JC method sensitivity means that noise will be carried over per time point instead of being smeared out over multiple time points. This is actually beneficial as it allows for further processing steps to be applied that aim to remove any remaining noise (e.g. motion) which cannot be done when the noise has been smeared across the connectivity time series (e.g. in windowed methods). The simulations and results presented in this study should not be taken as an exhaustive and complete assessment of all aspects of a given method to conduct TVC. Rather, the four simulations described here represents a subset of possible scenarios in terms of different methodological characteristics that might be of interest. The current four simulations are marked tvc_benchmarker simulation routine V1.0. If modifications or additional scenarios are considered to be improvements to the current simulations, these will get an updated version number. Many additional simulations could be conceived on top of this original routine. For example, one could include multiple time series, adding movement type artifacts, adding frequency relevant characteristics, a stationary global signal etc. These have not been included here, as the focus in these simulations was to primarily assess tracking of a fluctuating covariance. Input from researchers about appropriate additions to the simulations is welcome. We encourage researchers designing TVC methods to benchmark their own results with tvc_benchmarker (www.github.com/wiheto/tvc_benchmarker). Researchers need only to write a Python function for their method and use it as an input for tvc_benchmarker.run_simulations() and their method will be compared to the TVC methods presented in this paper (see online documentation). Functions can then be submitted through the function tvc_benchmarker.send_method(). All valid methods submitted will be released in summaries of the submitted benchmarked results so that researchers can contrast the performance of different methodologies.

10.1371/journal.pntd.0004015

Distribution of Peripheral Memory T Follicular Helper Cells in Patients with Schistosomiasis Japonica

Schistosomiasis is a helminthic disease that affects more than 200 million people. An effective vaccine would be a major step towards eliminating the disease. Studies suggest that T follicular helper (Tfh) cells provide help to B cells to generate the long-term humoral immunity, which would be a crucial component of successful vaccines. Thus, understanding the biological characteristics of Tfh cells in patients with schistosomiasis, which has never been explored, is essential for vaccine design. In this study, we investigated the biological characteristics of peripheral memory Tfh cells in schistosomiasis patients by flow cytometry. Our data showed that the frequencies of total and activated peripheral memory Tfh cells in patients were significantly increased during Schistosoma japonicum infection. Moreover, Tfh2 cells, which were reported to be a specific subpopulation to facilitate the generation of protective antibodies, were increased more greatly than other subpopulations of total peripheral memory Tfh cells in patients with schistosomiasis japonica. More importantly, our result showed significant correlations of the percentage of Tfh2 cells with both the frequency of plasma cells and the level of IgG antibody. In addition, our results showed that the percentage of T follicular regulatory (Tfr) cells was also increased in patients with schistosomiasis. Our report is the first characterization of peripheral memory Tfh cells in schistosomasis patients, which not only provides potential targets to improve immune response to vaccination, but also is important for the development of vaccination strategies to control schistosomiasis.

Schistosomiasis affects more than 200 million people worldwide and causes more than 280,000 deaths per year. Current control strategies are based on chemotherapy, but recurrent reinfection of people living in endemic areas makes researchers search for an effective vaccine to provide long-term protection against schistosomiasis. The generation of long-lived high-affinity antibodies after vaccination is a pivotal step for anti-schistosome vaccine to eliminate schistosomiasis. Considering it is well-known that Tfh cells are specialized effector CD4+ T cells that provide help for germinal center (GC) formation and induce GC B cells to develop protective antibody responses, understanding the biology of Tfh cells in schistosomiasis patients is fundamental for vaccine strategy development. Here, for the first time, we documented increased frequencies of total and activated peripheral memory Tfh cells in schistosomiasis patients. Furthermore, we showed that Tfh2 cells were a major contributor to increased frequency of peripheral memory Tfh cells in patients with schistosomiasis japonica. More importantly, we found the significant correlations of the percentage of Tfh2 cells with both the frequency of plasma cells and the level of total IgG antibody in schistosomiasis patients.

Schistosomiasis remains a major public health problem in many developing countries. Estimates place the current number of infections at approximately 200 million people, with another 600 million considered at risk [1]. Although praziquantel remains highly effective in schistosomiasis treatment, it provides only short-term protection and does not block disease transmission or reinfection [2]. Furthermore, drug resistance and decreased susceptibility to praziquantel may occur with long-term use of the drug [3]. Thus, an effective vaccine against schistosome infection would be a major step towards eliminating this devastating and widespread tropical parasitic disease. An effective anti-schistosome vaccine would immensely reduce the morbidity associated with schistosomiasis through induced immune responses leading to decrease in parasite load and reduced egg production [4,5]. The antibody dependent cell mediated cytotoxicity (ADCC) of effector immune cells such as eosinophils and macrophages has been suggested as one of the most important mechanisms of anti-schistosome vaccine-mediated protection [6–8]. Thus, the generation of long-term humoral immunity is a crucial component of successful vaccines. Interactions between T cells and B cells in germinal centers (GCs) are reported to be required for the generation of long-term humoral immunity [9]. Recent studies reveal that in GCs, a specialized subset of CD4+ T cells called T follicular helper (Tfh) cells, provide help to B cells to undergo proliferation, isotype switching and somatic hypermutation, resulting in long-lasting antibody (Ab) responses [10–12]. Thus, understanding the biological characteristics of Tfh cells in schistosomiasis patients is one of central issues to develop the vaccination strategies to control schistosomiasis. In this study, we for the first time explored the characteristics of peripheral memory Tfh cells in patients with schistosomiasis japonica, which provides a better understanding of the role of Tfh cells in schistosomiasis and contributes to the development of the future vaccination strategies in schistosomiasis. Ethical clearance for this study was obtained from the Institutional Review Board of Nanjing Medical University, Nanjing, China (Permit Number: 2014NMUIEC001). The aims and objectives of the study were explained to each participant and written informed consent was obtained. All personal identifiers of the study notes and tapes were kept confidential and destroyed once the study was completed. The study was conducted on a total of 100 subjects, and all subjects were from a village in Chizhou City, Anhui province. The subjects included 50 healthy adult controls, 50 patients with schistosomiasis japonica by egg detection using the Kato-Katz method with duplicate examination of 3 consecutive stool specimens obtained from each individual [13]. The healthy controls did not display a history, laboratory or clinical signs of schistosomal infection, did not suffer from coinfections with HBV or HCV, and did not use medication two weeks before blood collection. Human peripheral blood mononuclear cells (PBMCs) were collected into sodium heparin tubes (BD Biosciences, San Diego, CA) and purified by Ficoll-paque plus (GE healthcare, Sweden) density gradient centrifugation. Cells recovered from the gradient interface were washed twice, and stained for 30 min at 4°C with the following antibodies: CD3-FITC (clone HIT3a), CD4-Percp-Cy5.5 (clone RPA-T4), CD45RA-APC-H7 (clone HI100), CXCR5-Alexa Fluor 647 (clone RF8B2), PD-1-PE-Cy7 (clone EH12.1), ICOS-PE (clone DX29), CCR6-PE (clone 11A9), CXCR3-PE-Cy7 (clone 1C6), CD27-APC-H7 (clone M-T271), CD38-PE (clone HIT2), CD86-PE-Cy7 (clone 2331), CD19-APC (clone HIB19), all from BD Biosciences. In brief, for total or activated peripheral memory Tfh surface marker analysis, cells were incubated with CD3-FITC, CD4-Percp-Cy5.5, CD45RA-APC-H7, CXCR5-Alexa Fluor 647, PD-1-PE-Cy7, ICOS-PE. For Th1, Th2, or Th17 surface marker analysis, cells were incubated with CD3-FITC, CD4-Percp-Cy5.5, CCR6-PE, and CXCR3-PE-Cy7. For Tfh1, Tfh2, or Tfh17 surface marker analysis, cells were incubated with CD3-FITC, CD4-Percp-Cy5.5, CD45RA-APC-H7, CXCR5-Alexa Fluor 647, CCR6-PE, and CXCR3-PE-Cy7. For circulating B cell surface marker analysis, cells were stained with CD3-FITC, CD4-Percp-Cy5.5, CD19-APC, CD27-APC-H7, CD38-PE, and CD86-PE-Cy7. The samples were fixed with 1% paraformaldehyde/PBS. Cells acquisition was performed using a FACSVerse cytometer (Lasers: 488 and 633; Mirrors: 507 LP, 560 LP, 665 LP, 752 LP, 660/10, and 752 LP; Filters: 488/15, 527/32, 568/42, 700/54, 783/56, 660/10 and 783/56, BD Biosciences). Data were analyzed with FlowJo (Tree Star, version 10.0.7). To evaluate the percentages of GATA-3+ Tfh and Tfr cells, PBMCs were stained with CD3-FITC, CD4-Percp-Cy5.5, CD45RA-APC-H7, CXCR5-Alexa Fluor 647. Then, the cells were further intracellular stained with GATA-3-PE (clone L50-823, BD Biosciences) or FOXP3-PE (clone PCH101, eBioscience, San Diego, CA) after they were permeabilized with cold Fix/Perm Buffer (eBioscience). The samples were fixed with 1% paraformaldehyde/PBS. Cells acquisition was performed using a FACSVerse cytometer (BD Biosciences). Data were analyzed with FlowJo (Tree Star, version 10.0.7). To quantify the total serum IgG and IgE levels two commercial kit (Bethyl, Texas, USA) with established protocols from the manufacturer was used. Briefly, 96-well plates (Nunc MaxiSorp) were coated with 1 μg/well of capture antibody for IgE (catalog number A80-108A, Bethyl) or IgG (catalog number A80-104A, Bethyl) in Coating Buffer (0.05M carbonate-bicarbonate, pH 9.6) for 1 h at 25°C and blocked for 30 min with 200 μL/well Blocking Solution (50 mM Tris, 0.14 M NaCl, 1% BSA, pH 8.0). Between each step, the plates were washed 5 times with Wash Solution (50 mM Tris, 0.14 M NaCl, 0.05% Tween 20, pH 8.0). The serum from each patient was diluted 1:2 or 1:20,000 for IgE or IgG in Sample/Conjugate Diluent Solution (50 mM Tris, 0.14M NaCl, 0.05% Tween 20, 1% BSA), and 100 μL/well was added to the plates. Known concentration of purified human IgE (catalog number RC80-108-6, Bethyl) or IgG (catalog number RS10-110-4, Bethyl) was added to each plate to obtain a standard curve. Serum samples and standards were incubated for 1 h at 25°C. IgE or IgG bound to the plates was detected by the addition of HRP-anti-human IgE (catalog number A80-108P, Bethyl) or HRP-anti-human IgG (catalog number A80-104P, Bethyl) at 1:75,000 or 1:200,000 dilution in Sample/Conjugate Diluent Solution, followed by the addition of Substrate Solution (catalog number E102, Bethyl). After 15 min, the reaction was stopped with 100 μL of 0.18M sulfuric acid solution. Absorbance was measured at 450 nm using an automated ELISA reader (BioTek Synergy HT, Texas). For each patient, the amount of total IgE or IgG was quantified in triplicate. All data were analyzed using SPSS software (IBM, version 22). Significant Differences between specimens were determined by using Student’s t test or Mann-Whitney U test. Correlations were determined by Spearman’s ranking. The differences at p<0.05 were considered to be statistically significant. Both our previous study [14] and other literature [15] described the increased frequency of Tfh cells in mice with schistosome infection. However, whether Tfh cells increase in percentage in schistosomiasis patients remains unknown. To study the biological characteristics of Tfh cells in patients with schistosomiasis, a total of 50 patients and 50 healthy controls were recruited. There was no statistically significant difference in the distribution of age or gender between patients and healthy controls (Table 1). The frequency of CD4+ T cells among total lymphocytes was comparable between patients and healthy controls, although the percentage of T cells was slightly lower in patients with schistosomiasis (Fig 1B and 1C). Next, we compared the frequencies of total peripheral memory Tfh cells (CXCR5+CD45RA-CD4+ T cells) [16] and activated peripheral memory Tfh cells (PD-1+ICOS+Tfh cells) [17,18] among CD4+ T or total peripheral memory Tfh cells in 50 patients with schistosomiasis japonica to 50 healthy controls. Results showed the increased frequencies of total and activated peripheral memory Tfh cells in patients with schistosomiasis (Fig 1A and 1D–1H). In addition, we found that almost all of the CXCR5+CD45RA-CD4+ T cells expressed high level of GATA-3 (Fig 1I). Evidence supports that peripheral memory Tfh cells in human can be subdivided into three major subsets with distinguished biological functions according to expression of CXCR3 and CCR6, Tfh1 (CXCR3+CCR6-CD45RA-CXCR5+CD3+CD4+ cells), Tfh2 (CXCR3-CCR6-CD45RA-CXCR5+CD3+CD4+ cells), and Tfh17 (CXCR3-CCR6+CD45RA-CXCR5+CD3+CD4+ cells) [19]. We then determined the distribution of peripheral memory Tfh-cell subsets in healthy controls and schistosomiasis patients. Results showed that Tfh2 cells were a predominant subset of peripheral memory Tfh cells, and accounted for more than 50% of total peripheral memory Tfh cells in patients with schistosomiasis (Fig 2A and 2B). In addition, the percentage of total Th2 cells, which include Tfh2 cells, was greater in schistosomiasis patients than those in healthy controls (Fig 2C). Furthermore, we found that the percentage of Tfh2 cells, rather than that of Tfh17 or Tfh1, is significantly increased in schistosomiasis patients (Fig 2D–2G). Given that Tfr cells were identified as a Treg cell subset specialized for suppressing B and Tfh cells [20–22], we next investigated the biological characteristics of circulating Tfr cells in schistosomiasis patients. Results showed that most of circulating Tfr are CD45RA- cells (Fig 2H), which is consistent with the previous observation that circulating Tfr cells have memory-like properties [22]. Furthermore, we found that the percentage of circulating Tfr cells within CD4+ T cells (Fig 2I and 2J) or total Tfh cells (Fig 2K) was significantly increased in schistosomiasis patients. Given the roles of Tfh cells in providing help to B cells, we next characterized the frequencies of different subsets of B cells by flow cytometry analysis. As shown in Fig 3, the percentages of CD27+CD19+CD3-CD4- memory B cells [23–25], CD86+CD19+CD3-CD4- activated B cells [26–28], and CD38++CD19+CD3-CD4- plasma cells [29–31] in schistosomiasis patients were significantly greater than those in the HCs, although the percentage of total B cells was slightly decreased in patients with schistosomiasis. In contrast, the percentage of CD27-CD19+CD3-CD4- naïve B cells [32] in patients was less than that in the HCs (Fig 3C). Furthermore, the percentage of Tfh2 cells was moderately correlated with the percentage of plasma cells (rs = .362, p = .01) in schistosomiasis patients, and its correlation with the percentage of activated B cells (rs = .211, p = .141) followed the same trend but it did not reach statistical significance (Fig 3I and 3J). Given that Tfh2 cells are efficient at promoting IgG and IgE secretion [33], we next determined whether the frequency of Tfh2 cells was associated with the levels of total IgG and IgE antibodies in schistosomiasis patients. Results showed the increased concentrations of total IgG [HC vs. Sj: mean = 359.7, 95% confidence interval (95% CI) = 343.6–375.7 vs. mean = 417.9, 95% CI = 398.2–437.5; p<0.001] and IgE antibodies (HC vs. Sj: mean = 10.6, 95% CI = 9.5–11.7 vs. mean = 38.6, 95%CI = 30.5–46.7; p<0.001) in schistosomiasis patients (Fig 4A and 4C). Furthermore, a striking correlation between the percentage of Tfh2 cells and the level of total IgG (rs = .425, p = .002) was observed in schistosomiasis patients (Fig 4B). Although there was a tendency of correlation between the percentage of Tfh2 and the level of total IgE (rs = .173, p = .229) in schistosomiasis patients, it did not reach statistical significance (Fig 4D). Prevention and control of schistosomiasis demands an effective vaccine. T follicular helper cells have a pivotal role in the generation of the long-term humoral immunity and are proved to be one of crucial contributors of successful vaccines. However, the lack of knowledge about Tfh cells in schistosomiasis patients limits the ability to develop successful anti-schistosome vaccinations. Here, we characterized the distribution of peripheral memory Tfh cells in patients with schistosomiasis japonica. Our study significantly extends our understanding of Tfh cells in patients with schistosomiasis, which is helpful for vaccine design for the prevention of schistosome infection. Although the phenotypes of Bona fide Tfh cells in GCs are easy to be analyzed by flow cytometry, it is not only difficult to get lymph nodes from schistosomiasis patients only, but from humans in general. Fortunately, studies in humans showed that peripheral memory Tfh cells share functional properties with bona fide Tfh cells in secondary lymphoid organs [17,18,33–36], indicating the analysis of peripheral memory Tfh cells by flow cytometry is an alternative approach to study the biological characteristics of bona fide Tfh cells in human. Here, for the first time we revealed that the percentages of total and activated peripheral memory Tfh cells were significantly increased in schistosomiasis patients. These findings are in accordance with our previous observation that Tfh cells are substantially increased in schistosome-infected mice [14]. Human peripheral memory Tfh cells can be divided into three major subsets with distinguished functions according to the analysis of CXCR3 and CCR6 expression, including in CXCR3+CCR6- Tfh1 cells, CXCR3-CCR6- Tfh2 cells and CXCR3-CCR6+ Tfh17 cells [19]. Results from staphylococcal enterotoxin B in vitro coculture experiment suggested that human blood Tfh2 and Tfh17,but not Tfh1, cells can help naive B cells to produce immunoglobulins via producing interleukin-21 [33]. More specifically, it suggests that Tfh2 cells are considered to be efficient at promoting IgG and IgE secretion, whereas Tfh17 cells promote IgG and IgA secretion [33]. Our data showed that the significant increase of Tfh2 cells is a major contributor to the increased frequency of total peripheral memory Tfh cells in patients with schistosomiasis japonica. These findings nicely connect to our observation that the levels of memory B cells, activated B cells, plasma cells as well as total IgG and IgE responses were considerably increased in patients with schistosomiasis. More importantly, both IgG and IgE antibodies have been reported to be essential components of protective immunity, and involved in the ADCC of eosinophils and macrophages against schistosome larvae [37–39], which is considered as one of the most important means of anti-schistosome vaccine-mediated protection [6–8]. Thus, Tfh2 cells, a predominant subset of peripheral memory Tfh cells in schistosomiasis patients, might be considered as a potential target to improve IgG and IgE responses to vaccination. However, Tfh2 cells secrete Th2 cytokines, i.e., IL-4, IL-5, and IL-13 [33]. Given that the prolonged excessive production of Th2 cytokines contributes to the development of hepatic fibrosis and chronic morbidity in schistosomiasis [40] and the progression of Th2-mediated pathology in some diseases, such as asthma and other infectious diseases caused by extracellular parasites [41], it is very important for us to consider adverse effect of anti-schistosome vaccines on triggering Th2-mediated inflammation responses, particularly on liver pathology in patients with a history of infection with schistosoma japonica, when we manipulate Tfh2 cells to enhance IgG and IgE responses to vaccination. Strikingly, we found that the frequency of circulating Tfr cells, which play a crucial role in GC responses by limiting Tfh and GC B cell numbers as well as plasma cells differentiation [20], was significantly increased in schistosomiasis patients. In summary, our study for the first time described the distribution of peripheral memory Tfh cells, circulating Tfr cells, and B cells in patients with schistosomiasis japonica, which provides us a better understanding of the biological characteristics of these cells in patients with schistosomiasis.

10.1371/journal.pcbi.1003946

The Amino Acid Alphabet and the Architecture of the Protein Sequence-Structure Map. I. Binary Alphabets

The correspondence between protein sequences and structures, or sequence-structure map, relates to fundamental aspects of structural, evolutionary and synthetic biology. The specifics of the mapping, such as the fraction of accessible sequences and structures, or the sequences' ability to fold fast, are dictated by the type of interactions between the monomers that compose the sequences. The set of possible interactions between monomers is encapsulated by the potential energy function. In this study, I explore the impact of the relative forces of the potential on the architecture of the sequence-structure map. My observations rely on simple exact models of proteins and random samples of the space of potential energy functions of binary alphabets. I adopt a graph perspective and study the distribution of viable sequences and the structures they produce, as networks of sequences connected by point mutations. I observe that the relative proportion of attractive, neutral and repulsive forces defines types of potentials, that induce sequence-structure maps of vastly different architectures. I characterize the properties underlying these differences and relate them to the structure of the potential. Among these properties are the expected number and relative distribution of sequences associated to specific structures and the diversity of structures as a function of sequence divergence. I study the types of binary potentials observed in natural amino acids and show that there is a strong bias towards only some types of potentials, a bias that seems to characterize the folding code of natural proteins. I discuss implications of these observations for the architecture of the sequence-structure map of natural proteins, the construction of random libraries of peptides, and the early evolution of the natural amino acid alphabet.

If we were to design a proteome, what types and what proportion of amino acids would we use in order to optimize properties such as the diversity of sequences and structures, their robustness to mutations, or their ability to fold efficiently? Here, I use simple models to study the sequence-structure map of proteins from a design and evolutionary perspective. These models can be used to explore all sequences and structures, as a function of the types of interactions encoded by the sequence. I study the range of possible binary interactions between monomers, which include natural and artificial amino acids. The results indicate that different amino acid compositions induce vastly different sequences-structure maps. I classify and study the properties of these maps and relate their features back to the type of energy interactions. I compare these observations to the types of interactions observed in natural amino acids. My observations provide insights for our current view of the sequence-structure map of natural proteins, guiding principles for the construction of random libraries of peptides, and suggests constraints for the early evolution of the natural amino acid alphabet.

The implications of understanding the properties and organization of the sequence-structure map of proteins are broad, they range from explaining the diversity of known protein folds in the context of cellular physiology and their evolution [1], synthesize molecules of biomedical or industrial interest [2], to engineer polymers [3] and proteomes de novo. From an evolutionary standpoint the relation between sequence and structure is a particular case of a more general problem known as the genotype-phenotype map (GP map) [4]. According to the GP map framework, protein sequences correspond to genotypes and structures to phenotypes [5]. By using a measure of distance (e.g. the number of point mutations necessary to transform one genotype into another), sequences can be thought as part of a space of genotypes [6]. A graph theoretic representation of genotype space provides a quantitative, unifying framework to explore different properties of the sequence-structure relation, while considering these properties on a broader evolutionary perspective. In the following, I refer to this detailed characterization of the sequence-structure map, as its architecture. The study of the sequence-structure map of proteins unifies three research programs. First, the structural biologist's, seeking to understand the limits of structural diversity and its relation to sequences in the context of a universe of folds [7]. Second, the evolutionary biologist's program, focused on the role of selection versus neutral forces shaping the architecture of the map [8], [9], as well as on the nature and role of mutational mechanisms on the origin and evolution of biomolecules [10]. And third, the protein engineer and synthetic biologist's, interested on identifying regions of genotype and phenotype space, amenable to in vitro search and design [2]. Simple models of polymers, so called protein lattices or simple exact models (SEMs) [11] have been used extensively to explore the sequence-structure relation of proteins. These models were originally developed to study the dynamics of polymers by modeling key thermodynamic properties that govern folding [12]. They consist of short sequences (e.g. 12 to 36 mers), composed of a limited alphabet size, usually 2 to 20 monomers. Sequences are folded onto a lattice of fixed dimensionality (i.e. 2 or 3-dimensional) and geometry (e.g. square, cubic, FCC, etc). The most common SEM is the HP model, consisting of only 2 monomers (i.e. H, hydrophobic and P, polar). In the HP model only H-H contacts contribute to the stability of the conformation [12]. Their main limitation relates to finite size effects. That is, artifacts arising as a consequence of the model's geometry, dimensionality, and polymer length; which introduce biases on the relation between surface versus core residues and long-range interactions [13], [14]. These limitations have been proven detrimental to the study of folding kinetics and the cooperative two-state transition of globular proteins, for which the use of detailed atomistic models is advised [15]. Despite these limitations and their simplified representation of the geometric complexity of protein structures, SEMs have been instrumental in understanding a variety of aspects of protein biology [13]. They have been used to study theories and mechanisms of protein folding [13]; the distribution of sequences versus structures [16]; determinants of folding kinetics [17]; protein design [14]; recombination [18]; protein-protein interactions [19]; misfolding and aggregation [20]; the study of energy functions and their performance [21]; comparative modeling [22]; neutral networks and innovation [23] and protein evolution [11], [24], among others. In contrast to the study of natural proteins, SEMs can be used to fully characterize the sequence-structure map, that is, the relation of all possible sequences to all possible structures. Their strength relies on the characterization of large number of sequences and conformations, and therefore on the study of phenomena for which the statistics predominate over the details of folding [25]. A first relevant property of the sequence-structure map of proteins is that not all possible sequences are equally likely to encode a structure. Different criteria has been employed to decide on the propensity of a sequence to fold. In general, these criteria consider key thermodynamic determinants that distinguish between the stability of a sequence across conformations. For instance, the total stability of a sequence on its native conformation (E) [26], the energy difference (i.e. energy gap) between E and the next stable conformation(s) [27], or the deviation of E from the ensemble of all possible conformations (i.e. foldability [28], see below). Although all these criteria are approximations to the propensity of a sequence to fold, the degeneracy (g) of a sequence have proven a useful proxy to distinguish between foldable and random polypeptides. Degeneracy corresponds to the total number of conformations that a given sequence stabilizes at its minimum observed energy. Under this criterion, a sequence is considered foldable if it is non-degenerate (i.e. g = 1). In the case of SEMs, the stability of a protein sequence, folded onto a given conformation, can be approximated by the strength of the interactions between non-adjacent residues along the peptide chain. These interactions are encapsulated by a potential energy function, or simply, potential. The derivation and performance of potentials have been the subject of a long research tradition [29]. The most successful potentials are the result of statistical approximations that derive propensities of interactions between monomers from a large set of protein crystal structures (i.e. knowledge-based or statistical potentials). The physical interpretation of these forms of energy functions, however, remains a subject of debate [30]. One of the reasons is that, statistical potentials ignore much of the details of the interactions between residues in proteins. A major distinction between statistical potentials is the use of different reference states. The study of a diverse set of statistical energy functions derived using different reference states shows that most of them describe two putative stages during folding [30]. On the one hand, some potentials characterize the hydrophobic collapse of globular proteins [29], [31]–[34]. On the other hand, they might reflect subtle differences among residue interactions at the native or near-native state [29], [35]–[37]. Similar to the approximation employed by SEMs, statistical potentials have been successfully used to score the stability of protein crystal structures, and protein models, by only considering the pairwise interactions of amino acids [38]. Similar to the concept of degeneracy, one may consider the fraction of conformational space encoded by non-degenerate sequences, or encodability [39]. Both, non-degeneracy and encodability are closely related properties. They depend on the amino acid alphabet size and composition, which in turn defines the potential. In 1996, Chan and Dill [39], studied the impact of properties of the potential on degeneracy and encodability. They explored the role of repulsive interactions and correlations between energy values on well-known binary potentials and showed that the nature of the potential affects the sequence-structure map and, in doing so, it is as important as the size of the alphabet. Specifically, they studied the HP model and a modified version, the AB model; and showed that repulsive interactions reduce the average sequence degeneracy and consequently, increase the fraction of foldable sequences and encodable structures. While non-degeneracy and encodability describe the fraction of accessible sequences and structures, a full description of the sequence-structure map should also account for the relative use and distribution of sequences and structures in genotype and phenotype spaces. The language of graphs has been used to represent and study the distribution of sequences in genotype space [6], [40]. According to this paradigm, groups of non-degenerate sequences that fold onto the same structure and can be connected to each other by single point mutations, are known as neutral networks [5]. The size of neutral networks has consequences for the evolution of phenotypes. Arguably, sequences that are part of a large neutral network can undergo a considerable number of mutations while still preserving their phenotype. These phenotypes are found more frequently by a random search on genotype space and because of their robustness to mutations, represent good candidates for protein design experiments [41]. Following Maynard-Smith's concept of protein space [6], Lipman and Wilbur used the HP model to explore the existence and general statistics of neutral networks [40]. They observed that sequences folding onto the same conformation, map to nearby regions of genotype space and can be reached from various mutational paths. Subsequent studies, inspired by analysis of the RNA GP map, used SEMs to analyze the distribution of neutral networks in sequence space. These studies showed that neutral networks of the HP model distribute on isolated regions of genotype space, with unfrequent mutational paths between networks [42]. Other studies have explored the distribution of genotypes' stabilities in neutral networks [42], [43]. They showed that neutral networks have a funnel-like organization, where the most stable sequence usually corresponds to the network's ‘average’, or consensus sequence. The relation between structural stability and consensus sequence has been explored experimentally [44]. These authors have also compared the neutral networks between the HP and AB models. They demonstrated that features of the potential impact the number, size and longest paths of these networks [25], [43]. While sequences of a neutral network use nearby regions of genotype space, sequences that preserve the same phenotype may also occupy divergent regions of genotype space. These type of sequences, that belong to disconnected neutral networks in genotype space, are called neutral set [42]. Neutral sets are usually characterized by their size, in number of sequences, or designability [16]. Li et al (1996) used two and three-dimensional SEMs to show that designability distribute slightly less than exponential over conformations [16]. In other words, most conformations are associated to a single or few sequences, while few conformations use a large fraction of the available space of genotypes. At the time, this was a remarkable observation, because it recovered the biased distribution of the number of sequences per structure observed from very sparse natural samples [45]. Since then, two related hypothesis have been proposed to explain the origin of the vast differences on the designability of protein structures. One hypothesis relies on the requirement of structural stability [46]. Structural stability correlates closely with the total number of contacts of a conformation (or compactness). Since the contribution to the total energy of a sequence folded onto a conformation is given by the number of contacts between residues, the larger the number of contacts, the more stabilized a conformation can get and consequently, the larger the sequence variability. In other words, compact conformations are intrinsically designable. A second hypothesis concentrates on the propensity of sequences to fold fast [27]. Folding can be seen as a competition of a sequence for conformations. The diversity and stability of conformations surrounding the native structure is a measure of a sequence's ability to fold efficiently. This property is called foldability [28]. Different theoretical formalisms have been proposed to quantify it. Intuitively, these formalisms consider the energy gap, or difference in stability between the sequence folded onto its native structure and the stability at the next(s) most stable conformation(s). In other words, foldability is a measure of the steepness of the energy landscape surrounding the native structure. The concept of foldability does not aim to provide mechanistic details on the protein folding path, but simply identify important energetic features that distinguish natural proteins from random polymers [47]. Similar concepts rely on the same principle, such as the comparison of conditions for folding versus the conditions for chain collapse [48], or the principle of minimal frustration [27]. Theory based mainly on the random energy model and extensive simulation studies, have demonstrated the practical value of this idea. Other studies have also shown that this criterion alone, does not fully address the degrees of kinetic and thermodynamics complexity of natural proteins [15]. However, in the context of simple exact models, as it been studied before, the concept of foldability remains a good approximation as to how protein-like a polymer is [49], and as a requirement for protein design [14]. Designability and foldability capture different aspects of the sequence and structural constraints imposed on folding. Govidarajan and Goldstein showed that conformations have different foldabilities and that optimally foldable conformations are also highly designable [28], [47]. Buchler and Goldstein [50] used 25 mer, a two-dimensional, maximally compact SEM, to explore the distribution of designabilities under a range of amino acid alphabets and foldability requirements. They observed that, under these large variety of parameters, the distribution of designabilities remain strongly biased across conformations. This finding let them to suggest that designability is a general property of the protein GP map. The distribution of designability across structures, however, is highly dependent on the size of the amino acids alphabet, as is the identity of the most designable structures [50]. From an evolutionary standpoint the designability of a network of sequences, as well as their foldability, are important determinants of the mutational robustness of a phenotype [51]. In addition to the properties of isolated networks of sequences, a full description of the protein sequence-structure map should account for the distribution of neutral networks across genotype space relative to other networks and to the phenotypes that they map onto. Similar to the concept of designability, in revealing aspect of the mutational robustness of a phenotype, a sequence's accessibility to different phenotypes is a property of evolutionary relevance. This is because, the larger the phenotypic diversity in a neighborhood of sequence space; the larger the capacity of a sequence to innovate upon mutation [52], [53]. Because the amino acid alphabet, and therefore the potential energy function, impacts the fraction of foldable sequences and the encodability of phenotypes, arguably, it may affect the relative distribution of phenotypes respect to other phenotypes across sequence space, and consequently, impact both, the map's constraints on the accessibility to new phenotypes, as well as, its general ability to innovate through mutation. Recent advances in the de novo design and synthesis of polymers [3], the synthesis and manipulation of entire chromosomes [55], as well as, the introduction of new amino acids into the genetic code [54]; has opened new perspectives and challenges that touch upon these ideas. If we were, for instance, to choose the monomers to engineer a proteome, what types and proportion of interactions would we include in order to optimize mutational robustness, the fraction of accessible genotypes and phenotypes, and/or their foldability? This question suggests a sequence-structure map problem, that is not concerned with the mechanisms of folding, but with predicting the architecture of the map, given the composition of the amino acid alphabet. Similar questions exist in the field of protein design [2]. The construction of large random libraries of polypeptides used in in vitro search studies, would benefit of understanding what number and types of natural or artificial amino acids may promote sequence andor structural diversity [56], [57]. Yet another significant area of research relates to the origin and establishment of the early genetic code [58]. What is the minimal number and types of amino acids that allow the synthesis of a primordial, protein-like sequence-structure map of proteins? [58], [59]. This is a question that has haunted a wide variety of research fields since the late 60's [60], and for which there are partial theoretical and empirical insights [14], [61]. Although a thorough exploration of the myriads of factors involved in the early evolution of the genetic code is beyond the scope of the present study, an understanding of the relation between amino acid composition and the sequence-structure map, might provide indirect evidence on fundamental constraints that affected the establishment of the primordial amino acid alphabet of proteins. In this work, I study the impact of the potential energy function on the architecture of the protein sequence-structure map. I use SEMs, sample the space of possible binary potentials, and study the properties of the maps they induce. I analyse properties such as non-degeneracy, encodability, designability and foldability, the connectivity and relative distribution of neutral networks, as well as the overall phenotypic diversity of the sequence-structure maps induced by these potentials. I study the types of binary potentials present in natural amino acids and compare them to a random sample of the space of potentials. A detailed exploration of these properties may first, provide an alternative view of the sequence-structure map of natural proteins; second, help to explore the limits imposed by the architecture of the sequence-structure map on the evolution of proteins; and finally, may provide insights on the construction of random libraries of peptides and the large-scale design of sequence-structure maps with desired properties. A simple exact model (SEM) consists of three main parameters: sequence length (), monomer alphabet () and the potential (). Genotype space (), is composed of sequences. Where . (, is the cardinality of the set ). The dimension of , , is defined as the total number of single point mutant neighbors of a given sequence, as: . For 2, is called a generalized hypercube (). A sequence , is composed of monomers . A hamming distance metric, , over , defines a n-cube or hypercube, where (,), corresponds to the number of point mutations needed to transform genotype into [62]. Similarly, the space of phenotypes, , corresponds to the set of all possible conformations. The enumerable conformational space is independent of and growths exponentially as a function of . The potential energy function, , specifies the energy associated to the interaction between monomers and . The total stability () of a sequence , folded onto conformation , is defined as:(1) The function , adopts a value of 1 if monomers at positions and are in contact and non-adjacent along the chain, 0 otherwise. The degeneracy () of sequence corresponds to the number of conformations adopted at ( = min()). According to the thermodynamic hypothesis of protein folding, a sequence folds onto a conformation , if and only if, is non-degenerate on (i.e.  = 1). In that case, is called the native structure of . The k- neighborhood of a sequence is defined as the set of sequences at a hamming distance equal or lower than k, respect to (). The number of sequences of a k-neighborhood increases as . For and  = 18; 1, 3, and 5-neighborhood contain 18; 987; and 12,615 sequences, respectively. In order to quantify the relative distribution of phenotypes across sequence space, I consider the phenotypic diversity of a k-neighborhood centered at a sequence (). is simply defined as the set of phenotypes encoded by sequences in the k-neighborhood of . for small k values, informs on the fraction of immediate accessible phenotypes, those expected to be available after few point mutations; whereas larger k values, tell us about the overall diversity of phenotypes across sequence space. By applying Eq. 1 over all sequences in , a given potential , induces the folding (i.e. mapping) of a set of non-degenerate sequences (), which represents a fraction of genotype space ( ); into the set , a fraction of phenotype space ( ). We say that is the accessible conformational space induced by the potential on . The total fraction of non-degenerate sequences induced by , is called non- degeneracy ( =  = ). Similarly, encodability can be defined as:  = . The non-degenerate fraction of sequence space induced by , can be treated as a network of genotypes () (Figure 1). Sequences are nodes, and edges are formed between pairs of sequences that differ in one point mutation (h(,) = 1). When two nodes in can be connected by a series of single point mutations, we say there is a mutational path () between them. The diameter of a graph corresponds to its largest . A connected component of a genotype network, or genotype component (), is a subset of nodes in , for which there is at least one between every possible pair of sequences (,). A genotype network can be composed of one or more than one genotype component (); and the total number of sequences in the network is the sum of the number of sequences in each component (). Note that represents the set of genotype components, represents the number of genotype components, whereas , the size, in number of sequences, of genotype component . For instance, in Figure 1, is composed of three : . (I drop the subscript G, since all genotype components are necessarily part of ). The distinction of genotypes according to the phenotypes they map onto, induces subgraphs, whose properties have important consequences for the architecture of the map and can be characterized quantitatively in terms of the statistics of their expected size, diameter and distances. Sequences that fold onto the same phenotype are called neutral sets () and are, by definition, subsets of the genotype network (). Note that the number of is equivalent to the number of accessible phenotypes (). For instance, in Figure 1, is composed of 5 , represented by different colors. Sequences are known to distribute heterogeneously over conformations and this property of a phenotype, traditionally called designability (C) [16], has important implications for evolution and design. The designability of a phenotype j is equivalent to the size, in terms of number of sequences, of the neutral set associated to phenotype j (C = ). As in the case of , a neutral set () can also be composed of more than one connected component. These connected subsets of non-degenerate sequences, that map to the same phenotype, are called neutral networks (). Note the subscript distinction. refers to genotype, as in genotype network () and genotype component (). refers to phenotype. However, instead of using phenotype network () and phenotype component (), I stick to the terms traditionally used in the literature: neutral sets and neutral networks, respectively [5], [42]. A neutral set can be composed of more than one neutral network (). refers to the neutral network in the neutral set of phenotype (); and genotype component . As is the case of , all pairs of sequences in are connected by at least one mutational path. For instance, Figure 1, shows 9 . The largest genotype component in (), is composed of 4 and 5 . Similar to the idea of designability (C), here I define a network's neutrality as the size, in number of sequences, of a neutral network, as: C = . Whereas the neutrality of a single sequence (), is defined as the fraction of mutants in the 1-neighborhood of , that are part of . A genotype component () in , can be composed of more than one neutral network (). But, note that not all neutral networks () of a given neutral set (), are necessarily part of the same genotype component (). Sequences in are the sum of sequences in neutral networks () that can be part of different genotype components: . For instance, (Fig. 1, in blue), can be expressed as  = (). (See Table S2 in Text S1 for a summary of symbols and abbreviations). A binary potential can be represented as a vector composed of 3 values, that describe 2 types of interactions (Figure 2A). First, those between the same type of monomers, or homomonomeric (i.e. , ); and second, the heteromonomeric interaction (i.e. ). The heteromonomeric interaction of a binary potential can be decomposed into ideal and excess parts [12], [30]. These parts describe the extent to which the potential favors two different hypothetical stages of the folding process. The ideal part represents an heteromonomeric interaction as in an ideal liquid. That is, as if there was no energetic contribution by the heteromonomeric interaction, and therefore it could just be approximated by the arithmetic mean of their homomonomeric values, as:  = (+)/2. In contrast, the excess part ( =  - ), aims to capture the contribution of the heteromonomeric interaction, and describe the extent to which the native conformation differs from an ideal mixture of amino acids, its additivity (). Here, I quantify the additivity of a given potential as:  = [E/E]+1 = /E. In this study I use a two-dimensional SEM of sequence length 18 mer. In the following I refer to this model as L18. The motivations for using this model, are fourfold. First, L18 represents a good compromise in relation to the number of sequences versus the number of conformations. Second, inspired by globular proteins, some previous studies assume that foldable sequences must adopt a maximum number of contacts. Because the restriction of phenotype space to maximally compact conformations introduces artifacts, as inflated values of designabilities [50], here I consider sequences folding onto any possible conformation, as long as, the thermodynamic criterion is met. Third, compared to three-dimension, two-dimension SEM show a surface-core ratio more similar to natural proteins [13]. Finally, the L18 model has been extensively used to evaluate different alphabets and potentials [42], [43], which will allows us to compare our results to previous findings. In the case of L18, is composed of 5,808,335 total conformations, and , of 262,144 sequences. Because the energy of a sequence folded onto a given conformation is here approximated by the contact of non-adjacent monomers along the chain, conformations in a lattice are usually represented as contact sets, a binary symmetric L by L matrix that describes the interactions of non-adjacent monomers [63]. Due to the larger degrees of freedom of conformations with few contacts, different conformations may correspond to the same contact set. The total 5,808,335 conformations of the L18 model, can be described by 170,670 non-redundant contact sets. Only 77,635 out of the 170,670 contact sets, are unique (ie. each one of them correspond to a single conformation) and therefore potentially encodable () under the thermodynamic hypothesis criterion [39]. The accessible conformational space of a sequence-structure map, represents a subset of the uniquely encodable set of conformations (). A sequence's foldability (), is mathematically described as the deviation of the energy minimum from the energy distribution of the ensemble of all possible conformations in [50]:(2) is the expected stability of sequence , over all possible conformations in the ensemble; , the standard deviation over the same distribution; and , the minimum observed energy of folded onto . A more negative value describes a steeper folding funnel and therefore protein-like behavior. In order to explore the impact of the potential on the architecture of the sequence-structure map of natural proteins, I concentrate on the L18 model and binary alphabets. The computational tractability of this model allows us to study exact statistics of a large sample of potentials. The potential of a binary alphabet is described by three values: , and (with  = , see Figure 2A). values are real numbers. If negative, they correspond to attractive interactions. If positive, repulsive. Neutral interactions () do not contribute to stability. Because of the symmetry () of the cube (), homomonomeric interactions ( and ) are interchangeable. In other words, if all monomers were exchanged by monomers, properties of genotype space would remain the same. The first protein lattice model ever studied was the HP model [12]. It is composed of two types of amino acids, polar (P) and hydrophobic (H). The potential is detailed in Fig. 2B. Only homomonomeric hydrophobic interactions () contribute to the stability of a folded sequence. An alternative to the HP model, the AB model, was introduced in order to explore the impact of the potential on protein design [64]. The AB potential introduces equivalent interactions between homomonomers ( =  = −1.0) and a repulsive interaction ( = 1.0), (Fig. 2C). The HP and AB potentials have been modified (the so called shifted potentials) to study explicitly the impact of repulsive interactions [39]. Figures 2D and 2E show the shifted versions of HP and AB potentials. I refer to these 4 potentials as canonical. In order to investigate the impact of the potential on the sequence-structure map, I begin our analysis by sampling the space of possible binary potentials, with −1.00, −0.75, −0.50, −0.25, 0.00, 0.25, 0.50, 0.75, 1.00; of which, canonical potentials are a small subset. Since a binary potential is composed of three values, our sample produces a total of possible . From these total possible combinations one must ignore potentials with no relative favorable interactions ( =  = ), potentials with only repulsive or neutral interactions (, , 0.0), scaled potentials of the form , , (with ), take into account the symmetry at homomonomeric interactions (i.e. ,, = ,,), and, the symmetry between homo versus heteromonomeric interactions (i.e. if  = ; then, , , , , , , ). These considerations result on a total of 245 potentials. Potentials can be represented as vectors. Due to the symmetry of the cube (), half of the space contains all possible non-redundant binary potentials. As suggested by previous studies, many properties of the potential energy function are determined by the proportion of repulsive, attractive and neutral interactions [39], [50]. I use this criterion to distinguish among 7 types of potentials, that correspond to the 6 non-redundant octants in the 3d -coordinates representation, plus any potential with at least one 0 (Table 1, Figure 3). The octant in black (Fig. 3), that corresponds to all-repulsive interactions (, , 0.0); is defined as potential type VII and, by definition, does not stabilize any conformation (see Eq. 1). The 245 potentials described above are an homogeneous sample from this space. For each of the 245 potentials I proceed as follows. I enumerate all possible sequences. I fold each sequence onto every contact set and calculate its stability and foldability using equation 1 and 2, respectively. (The raw data of the 245 sequence-structure maps studied here, is available at: www.santafe.edu/eferrada, see Table S1 in Text S1.). In order to compare different potentials and their impact on properties of the sequence-structure map, I use hierarchical clustering (see Supplementary Methods). The Jaccard similarity index, between the sets and (), (with , ; ), is defined as:  = . , measures the similarity between the sets of conformations and (with ), induced by the potentials and , respectively. Similarly, , compares sets of sequences and (with ) (see Supplementary Methods). Figure 4 presents a hierarchical clustering of phenotype space based on (and ), for all possible pair combinations of binary potentials and (, 1,…, 245). Here I arbitrarily choose to focus on , however, similar conclusions arise from the analysis of the Jaccard index on genotype space () (Figure S1). Each tip of the tree represents an independent sequence-structure map. Maps that cluster closely in this tree have similar sets of accessible phenotypes (), that is, values close to 1.0. values that compose each potential are specified on a color scale at the branches' tips, with , specified at the outermost value. Branches are colored according to the potential, as described above (Table 1, Figure 3). Green and blue stacked bars following the color-coded potentials, correspond to non-degeneracy and encodability values, respectively. Boxplots, in black, represent the distribution of foldability over all non-degenerate sequences of each map. A first observation from Figure 4 is the impact of the potential on non-degeneracy, encodability and foldability, as well as the overall consistency of these properties across potentials with similar values. The potential can induce considerable differences in non-degeneracy and foldability. Confront, for instance, potentials type IV and potentials type II (green and orange branches, respectively). A similar observation applies in the case of clustering based on . In both cases, results are independent of the clustering method (Fig. S2, and Supplementary Methods). A second general observation regards the abrupt changes in the use of phenotype space between some of the maps with potentials of the same type. While potentials type I, II and V (blue, orange and red branches, respectively) are highly clustered, potentials type III and IV (magenta and green branches, respectively), distribute across different clusters. Figure 4 also reveals that canonical potentials are part of a larger family of potentials, which represent only 3 out of the 7 different types described above (Table 1; Fig. 3 and 4). Most notably, other combinations of values, in particular, potentials type I and II; induce sequence-structure maps that are as protein-like as the HP model (see below, section Foldability). Moreover, potentials that induce similar fractions of sequences and structures, present considerable variation in their average foldability. I now turn to a closer look at these differences. Non-degeneracy () is the fraction of genotype space that yields viable, folding sequences. It ranges from 2 to 28% across maps generated by the binary potentials sampled in this study (green bars in Fig. 4). Similarly, encodability (c), the fraction of accessible phenotypes, varies from 1 to 19% (Fig. 4, blue bars). Both, and vary considerably across types of potentials (Fig. S3). and are not independent and overall, correlate positively. Their association, however, depends on the type of potential (Fig. 5). In the case of potentials type I, II and IV, an increase in leads to larger c values. Potentials type III, however, preserve similar c despite large variation in . With the exception of few potentials type I and VI, maps induced by binary potentials, use a larger fraction of sequence than structure space (dashed line, Figure 5). Two main features of the potential account for and c. First, low negative values of , that is, average attractive homomonomeric interactions (0.0), promote both increasing and (see Fig. S4). The lowest values of are observed in the case of potentials types II and III. Second, positive values seem to be sufficient to promote , but not (Fig. S4 and S5). Potentials type I and II are the only potentials with positive values. They present encodabilities that are on average one order of magnitude larger than the rest of the potentials sampled in this study. These two features provide some intuition as to why potentials type II and III reach large values of , but only type II present also large values of c (Fig. 5); whereas potentials type V show low and conserved values of and . The component of the potential does account for both 0.0 and 0.0. Therefore, and c are expected to correlate positively with (Fig. S4B, S4D). As observed before, repulsive interactions reduce the average sequence degeneracy, increasing and [39]. However, our analysis of a large sample of potentials shows that not any type of repulsive interaction possess this effect, but only the heteromonomeric component of the potential, and that the effect is favored in the context of overall attractive . Most notably, these observations suggest that, by controlling for the components of the potential, both, the fraction of sequence and structures can be increased and furthermore, optimized independently of one another. For instance, the average number of sequences per conformation (i.e. designability) can be optimized by increasing while keeping c constant, as is the case of potentials type III (i.e. increasing attractive interactions in both, homo and heteromonomeric components, Fig. 3 and 5). Although can be seen as the probability of finding a viable sequence, the distribution of sequences in genotype space is not uniform, and depends on their monomer composition. Sequences can be classified according to their composition into classes. Compositional classes correspond to the frequency of the relative fraction of monomers across non-degenerate sequences induced by a given potential. In the case of maps composed of binary potentials, compositional classes distribute binomially. If all 2 sequences in the L18 model were non-degenerate, there would be 19 compositional classes, ranging from the unique two sequences composed of only one of the two monomer types (compositional classes 0 and 18 in Fig. 6; with 0 and 100 of monomer , respectively) to 48,620 sequences composed of 50 of each monomer (compositional class 9). In order to study the distribution of non-degenerate sequences across genotype space, I compare observed versus expected frequencies of different compositional classes. Expected compositional classes are estimated for a given potential , by sampling random sequences from genotype space, assuming that every sequence is equally likely to be non-degenerate. Potentials present considerable biases toward certain compositional classes (Fig. 6). In particular, genotype spaces of potentials type I are enriched in monomers, with compositional classes near 61. In contrast, potentials type IV show significant deviations toward monomers. In addition, consistent with the abrupt transitions between similar potentials (, Fig. S1), potentials type III, IV and VI show considerable variation (error bars, Fig. 6). Deviations from expected distributions can be explained by the proportion of attractive and repulsive values at homo versus heteromonomeric interactions (Fig. 3, Table 1). In the case of perfectly symmetric interactions between homo and heteromonomers, as is the case of potentials type II and V (Fig. 3, 6), there are no major deviations toward compositional classes enriched in either of the monomers. In these two cases, the diversity of repulsive and attractive interactions do not favor any compositional class. In the case of potentials type I and IV, however, one of the homomonomeric interactions breaks the symmetry of the potential, favoring the monomer that better counteracts stability respect to , increasing the diversity of competing interactions. Thus, potentials type I favor monomers type (0 and 0); whereas potentials type IV, monomers (0 and 0). Similarly, deviations in structural space can be estimated by considering the distribution of number of contacts across the conformations induced by a potential (i.e. compactness). The distribution of expected number of contacts can be estimated by assuming that every uniquely encodable conformation is equally likely to be accessed by non-degenerate sequences. Therefore, for a given potential (), I sample conformations and calculate their number of contacts. Where , is the encodability of sequence-structure map , and , the set of uniquely encodable conformations of phenotype space (see Models). All types of potentials deviate significantly from the expected distributions and in particular, compact conformations are more underrepresented than open ones (Fig. S6). Error bars indicate that deviations from expected distributions of contacts, are more consistent across potentials type I, V and VI. This is not the case of potentials type II, III and IV (Fig. S6). Potentials type I favor structures with less number of contacts (i.e. open conformations), and types II deviate toward compact conformations. Figure S7 shows examples of the most and least common structures per type of potential. Notice the reduced number of contacts in potentials type I, even for the most common conformation (Fig. S7A). As shown before, repulsive heteromonomeric interactions (0) promote c (Fig. 3 and 5). In the case of an additional repulsive homomonomeric interaction (0 in potentials type I), the distribution of conformations shifts considerably towards open conformations (Fig. S6, I & II). A similar effect is observed by comparing potentials type III, IV and V. The addition of repulsive interactions in potentials type IV and V, have a slight impact on the unfavored open conformations observed in potentials type III (Fig. S6, III). In summary, the potential energy function affects the monomer composition of non-degenerate sequences and the compactness of conformations. The symmetry of the potential, defined as the proportion of attractive and repulsive forces in homo versus heteromonomeric interactions, favors the unbiased use of genotype space and viceversa. Moreover, the relative increase of repulsive over attractive interactions, favors open conformations. In the previous sections I observed that first, potentials vary in their propensity to induce the folding of sequences and structures. Second, potentials favor the viability of regions of sequence and structure space with biased sequence composition and compactness. Here I turn more closely to the relation between sequence and structure across maps. In particular, the relation between the number of sequences per structure, or designability (see Models). Designability (C) is known to distribute heterogeneously over conformations [16], and this property of a phenotype, has important implications for protein evolution and design. Designable structures, those that map to many sequences, are more likely to be found by a random search across genotype space and are, by definition, more resistant to mutations. In order to study C across the phenotype space of a sequence-structure map, I calculate the probability of finding, among non-degenerate sequences, a genotype that folds onto a phenotype with designability C or larger. Figure 7 shows such probabilities as logarithmic cumulative distributions for different types of potentials. As studied before, in the case of the HP model, the probability of finding a phenotype with C, distributes approximately exponential in the 2D lattice [16], [50]. I confirm this trend for potentials type I and II. Other potentials, however, deviate strongly from an exponential distribution. In the case of potentials type I and II, the probability of finding sequences that map to increasingly designable phenotypes decreases fast compared to the rest of the potentials and is similar to the HP model (black dashed line, Fig. 7). The opposite is true for potentials type III and VI. For instance, in the case of potentials type I, the probability of finding a non-degenerate sequence that maps to a phenotype with C 10, is approximately 0.05; while in the case of potentials type III, with the same probability, one finds maps with C 110 sequences. Potentials type V, on the other hand, distribute narrowly and with probability 0.05, presents neutral sets of at least 40 sequences. Different degrees of variation across potentials of the same type (e.g. contrast potentials type V and III), are the result of the differential distribution of and . For instance, potentials type III, that show increasing , while keeping relatively constant values of , present the broadest C distributions. In contrast, potentials type I, with increasing almost linearly with , the probability of finding larger C, decreases rapidly. In contrast, potentials type V present conserved values of and , which translates on narrower probability distributions (confront Fig. 5 and 7). As suggested by a previous study, the identity of designable phenotypes is largely influenced by the potential [50]. As noted above, there is considerable overlap among the phenotypes induced by the potentials studied here. In order to explore this observation further, I group potentials according to their type, rank phenotypes by designability and consider the top and bottom 1 percentiles. There are only 122 conformations (2% of the average number of conformations per potential) encoded by every potential. There are no universally designable phenotype across the potentials studied in this work. I observe that with the exception of potentials type V, the most and least designable phenotypes are unique to each type of potential. Figure S7 shows examples of these phenotypes. Recall that genotypes of the same neutral set () are not necessarily connected (see Models). Therefore, from an evolutionary standpoint, instead of and C, one should rather look at the size of neutral networks (C). The reason is that the connectivity of genotypes that are part of the same , allows them to mutate while preserving the same phenotype (). Here, the super and subscripts, stand for the neutral network of phenotype , in genotype component (see Models). Figure S8 shows the cumulative probability distribution of size, across maps. As expected, the probability of finding C, decays faster compared to C of neutral sets. This trend is particularly clear for potentials type II, III and V. In the case of potentials type III, for instance, the probability of finding neutral sets with 10 or more sequences, reaches values of 0.8; whereas finding neutral networks of similar size, only occurs at probabilities of 0.25. The trend is also evident for potentials type II and V. For instance, with probability of 0.05, one finds neutral sets of 20 and 40 sequences, respectively; whereas, with the same probability, one finds on average neutral networks of only 6 and 8 sequences, respectively. In contrast, the probability distributions of neutral networks and sets, are very similar in the case of potentials type I (see below). In both cases, with probability of 0.07, one finds cluster of sequences of approximately 10 sequences or larger. These observations suggest that, in addition to variation on the available fraction of sequences and conformations (i.e. and ), there are considerable differences in C and C across potentials. Although different types of potentials induce similar sets of phenotypes, the identity of the most and least common phenotypes vary considerably. Additionally, potentials induce differential allocation of sequences across connected components (), which suggests influences on the size and distribution of neutral sets and neutral networks across genotype space. In the next section, I explore this aspect in more detail. As described in Models, non-degenerate sequences in genotype space can be construed as graphs. In order to investigate the impact of the potential on the distribution of sequences in genotype space from a network perspective, I look at the expected size of connected components (), neutral sets () and neutral networks () across different maps. I calculate the expected size of a cluster of sequences () from a collection of sets, x, as: ; where are particular instantiations in the set x: , or . simply computes the weighted average of sequences by their corresponding component size. Because every sequence is multiplied by its component's size; is equivalent to sum over the squares of the size of each component. If we were to choose a random non-degenerate sequence, from genotype space; would represent the expected size of the genotype component to which belongs; , the expected designability of its phenotype and , the expected neutrality of the neutral network associated to the same phenotype. Figure 8 shows the distributions of and per type of potential. In order to compare maps generated by different potentials, I scale expected size by non-degeneracy (see legend of Fig. 8). Potentials type I, II and V, show genotype components that span on average 97, 99 and 93% of non-degenerate sequences, respectively (insets Fig. 8I, II, V). Note, however, that these distributions of expected size are generally due to the presence of a large genotype component. Figure S9 shows the distribution of the diameter () of genotype components per type of potential (see Models). While 60 to 90% of genotype components of potentials are composed of a single sequence ( = 0), all types of potentials show at least one large spanning genotype component ( = 18) (Fig. S9). In addition, potentials type I, II and V, as confirmed by designability of neutral networks in the previous section (Fig. S8), present small neutral networks mostly composed of 2 sequences (Fig. 8). Figure S10 shows the distribution of neutral networks diameter across potentials. Potentials type I and V do not show neutral networks with 9. Maximum diameter observed for potentials type II is 11. In contrast to potentials type I, II and V; III, IV and VI, present genotype components and neutral networks that deviate towards smaller and larger expected size, respectively (Fig. 8, S9, S10). Although giant components dominate in the case of potentials type III and IV (Fig 8), they also show cases where genotype components' expected sizes reduce to 60 and 40% of non-degenerate sequences, respectively. In both cases the expected size of neutral networks increases up to 120 and 60 sequences, respectively (without scaling by ). Potentials type IV and VI present neutral networks of diameters up to 14 and potentials type III show cases of neutral networks that cross genotype space ( = 18). Random graph theory predicts that the diameter of a neutral network (D()) is a function of the average neutrality of sequences that compose the network (see Models). The theory predicts the existence of a critical value [65]. If the average neutrality of sequences in a network of phenotype (), is larger than the critical value (), then, sequence in , percolate across genotype space and form a giant component. For a binary alphabet,  = 0.5. Overall, potentials sampled in this work show low ( = 0.33) (Figure S11); and, the maximum diameter of neutral networks for potentials type I, II, IV, V and VI; are 9, 11, 14, 9, and 14, respectively (Fig. S10). Five potentials type III, however, present at least one neutral network with D = 18. Moreover, the average neutrality of these neutral networks is 0.15–0.19. There are at least 2 reasons for this disagreement with the theory. First, random graphs may not be a good approximation for neutral networks in the L18 model and/or potentials type III. Results presented in a previous section (i.e. the use of sequence and structure space), support this hypothesis. Second, it might be the result of finite size effects in the L18 model. In order to test the second hypothesis, should be calculated at the asymptotic limit [65], an analysis that is beyond the scope of this work. As explained in Models, sets of sequences that fold onto the same conformation (i.e. neutral set) can be composed of more than one neutral network (). Similarly, genotype components can be composed of more than one neutral set () (Fig. 1). In order to explore these differences on the architecture of sequence-structure maps from a broader perspective, I look at and as a function of the number of genotype components () and number of neutral sets (), respectively (Figure 9). Each point in Figure 9 is a sequence-structure map induced by a potential of a given type (color code, Fig. 3, Table 1). The number of and , vary approximately one order of magnitude across different potentials. However, compared to , there are ten times more (Fig. 9). As the number of increases, the space gets partitioned into more components and the expected designability of phenotypes () decreases proportionately (Fig. 9A). Potentials type I show few number of components (100–400) that contain on average, a large number of neutral sets (5,000–10,000 - Fig. 9B), of small expected size (10 sequences - Fig. 9A). Similarly, potentials type II (and V), induce maps with fewer (and larger) , with relatively larger (and smaller) , respectively (Fig 9B). Potentials type I, II and V show, on average, small (i.e. low neutralities). In contrast, potentials type III and IV show genotypes components of vastly different sizes. These potentials are enriched on sequences of the same phenotype and consequently, their maps show low encodabilities (x-axis, Fig. 9B). Strikingly, the expected designability of some potentials type III, decreases almost linearly as function of the decimal logarithm of , approximately as 15% per order of magnitude (Fig. 9A). The number of decreases rapidly once encodability reaches values of 5,000 phenotypes (Fig. 9B). The ratio of the expected size of different sequence clusters shows that genotype components are approximately 1,000 to 3,000 fold larger than the expected size of an average neutral set across potentials (/1,000–3,000) (Figure 10A). Although in general the expected size of an average neutral network follows a similar proportion, potentials type II and V, show large deviations, with genotype components: / 9,000 to 12,000 fold larger than the expected size of neutral networks. These ratios are particularly well conserved across potentials type V (Fig. 10A). A similar analysis comparing the expected sizes between neutral sets and neutral networks, shows major differences across potentials (Fig. 10B). Strikingly, and as anticipated (see section the designability and neutrality of phenotypes), potentials type I show exclusively fully connected neutral sets (/ 1). In contrast, potentials type V present neutral sets on average 5 to 6 times consistently larger than the expected neutral network. With the exception of potentials type II, that shows on average 4 to 5 neutral networks per phenotype (Fig. 10B); the rest of the potentials show large variation with predominantly 1 to 2 neutral networks per phenotype. In summary, potentials type I, II and V, induce sequence-structure maps of relatively similar organizations. These potentials show large genotype components and on average few sequences per phenotype. With the exception of potentials type V, however, I and II show on average few genotype components. Potentials type I show neutral sets composed of a single neutral network and on average, 3,000 networks per genotype components. As seen before, these networks possess short diameters. Similar to potentials type I, type II show approximately 3,000 neutral sets per genotype component. These types of potentials, however, show on average, neutral sets 4 times larger than the expected size of a neutral network. Potentials type V, on the other hand, show on average, 1,800 neutral sets per genotype component and these neutral sets are consistently composed of 5.5 to 6 times more neutral networks. In contrast, potentials type III, IV and VI, induce sequence structure maps with genotype components and neutral sets of a wide variety of sizes. These potentials show long-tailed distributions of neutral networks per phenotypes, with on average 1 to 2 networks per neutral set. In addition, they show approximately 1,000 to 2,000 neutral sets per genotypes component. While potentials IV and VI reach neutral sets and networks of expected size up to 70 sequences, potentials type III shows neutral networks of up to 120 sequences. Although the distribution of sequences in genotype space, and in particular of neutral networks, informs on the abundance of phenotypes and their expected mutational robustness, it does not tell us about the mutational divergence between different phenotypes. The differential accessibility to phenotype variants across genotype space has a profound impact on the ability of sequences to produce new, unobserved phenotypes. In order to study the relative accessibility of sequences to new phenotypes, I consider the phenotypic diversity of a pair of -neighborhoods centered at and (, ) (see Models). I calculate the overall fraction of phenotypes unique to each of the two k-neighborhoods at distance , and constant , as: . measures the overall diversity of two phenotype neighborhoods as a function of their divergence in genotype space. Note that non-overlapping k-neighborhoods only occur at [8], [66]. Figure 11 presents and as a function of distance for potentials type I-VI. At very short distances (with even overlapped neighborhoods), shows 50 to 70% of unique phenotypes (Fig 11A). As expected, at short and larger , decreases as a function of (Fig. 11B). In the case of 2-neighborhoods, the fraction of unique phenotypes increases rapidly with distance and, at short , there are only slight differences between types of potentials. At the overlapping threshold of (, dashed line Fig. 11), approximately 85 to 95% of phenotypes are unique to pairs of neighborhoods. At larger distances, however, differ considerably across potentials. For instance, at , potentials type I access 2-neighborhoods with 100% new phenotypes; whereas, distant 2-neighborhoods of potentials types II, III, IV and VI, share from 10 to 15% phenotypes. This trend intensifies in the case of potentials type V, that reach similar values compared to 2-neighborhoods at short distances (Figure 11A). In the case of larger -neighborhoods, differences between potentials discussed above become more evident (Fig. 11B and S12). Strikingly, potentials type I consistently find unique phenotypes at . In stark contrast, potentials type V, recover completely the levels of observed at short distances in a fairly symmetric pattern (Fig. 11B, S12). In order to further explore these differences I look at maximal distances () between sequences that are part of the same neutral set, that is, sequences that fold onto the same phenotype. Figure S13 shows such distributions per type of potential. As expected, potentials type I show short maximal distances, with hardly larger than 7 point mutations. In contrast, all other potentials show phenotypes at varying distances and sequences at opposite sides of genotype space ( = 18). In particular, potentials type II, IV and VI show 40 to 60% of phenotypes with  = 18. Consistent with the patterns observed in Fig. 11 (and S12), potentials type III and V show on average, 70 and 97% of phenotypes with  = 18, respectively (Fig. S13). The existence of sequences at can be explained by the degree of symmetry between attractive and repulsive interactions in the potential. A sequence folds onto its native conformation by stabilizing a set of contacts (2 to 10 in the case of the L18 model). In the case of a completely symmetric potential (as type V), sequences in opposite sides of genotype space, those with every position mutated by the opposite monomer, would preserve the same fraction and type of interactions, and therefore stabilize the same phenotype. In contrast, an asymmetric potential, as potentials of type I, with a single homomonomeric attractive interaction, will populate phenotypes at biased compositional classes. As these observations predict, potentials type III, a fully symmetric potential with no competing interactions, stabilizes sequences at a varying range of distances (Fig. S13). In summary, potentials induce maps with variable degrees of phenotypic diversity and divergence between neighborhoods. At short mutational distances, there is a large fraction of phenotypic diversity and these values are consistent across types of potentials. At moderate and long distances, however, potentials differ extensively in the distribution of unique phenotypes and those differences are due to the symmetric distribution of the potential's attractive and repulsive forces in homo versus heteromonomeric energy terms. Not all non-degenerate sequences are guaranteed to fold readily onto their native conformations. The propensity of a polypeptide to fold fast is an important determinant of how protein-like is a random sequence [67]. Next, I look at the impact of the potential on foldability (), a measure of a sequence's propensity to fold (see Models). Foldability is very sensitive to parameters in the potential. As shown in Figure 4, varies extensively across sequence-structure maps, even among those induced by potentials of the same type (see Fig. S14). Min and max median values are −8.2 and −2.3, respectively (the lowest the foldability, the faster the folder - see Eq. 2). Similar values of foldability are also observed to correlate very well with the accessible set of genotypes (Figure S1). The canonical potential HP has a notorious long-tailed distribution biased towards fast folders. This is however not a peculiarity of the HP model, and similar protein-like sequence-structure maps are observed in the case of potentials type I, II and other potentials type VI (Fig. 4 and S1). In addition, variation on the foldability of maps induced by the same potential type, suggests that foldability is highly sensitive to changes on the potential (confront for instance, potentials type I or II in Fig. 4 and S14). Evidence from the theory of protein folding relates foldability to cooperativity or the non-additivity of interactions [15]. In the context of binary potentials, I measure additivity () as deviations of excess from the ideal part of the potential (see Models). Figure 12 presents the median across all non-degenerate sequences of each sequence-structure map, as a function of . In the case of a completely additive potential: (dashed lined at ). The association between and for potentials type I-V is delineated by the foldabilities of potentials type VI (black dots in Fig. 12). This is due to the fact that transitions between types of potentials occur whenever 0.0 (grey planes in Fig. 3). Sequence-structure maps that favor foldability are induced by HP-like energy functions, which include potentials type I, II and VI. Not every potential that deviates significantly from additivity ensures a foldable map. Figure 12 shows that the extent to which additivity dictates the overall of a map, is a function of the type of interactions present in the potential. For instance, potentials type II and III reach better as ; whereas in the case of potentials type V, when . In summary, our observations confirm the impact of a potential's non-additive interactions on favoring protein-like sequences. I observe that the role of non-additivity is highly dependent on the form of the potential and that different potentials can induce maps as protein-like as the canonical HP model. HP-like sequence-structure maps are particularly induced by potentials type I and II. Most notably, this analysis suggests that by controlling for the form of the potential, it is possible to design a map with a desired fraction of protein-like sequences. How random are the pairwise interactions observed in the natural amino acid alphabet? I assess this question by comparing the pairwise interactions of amino acids in the Miyazawa-Jerningan (MJ) potential (Table VI in [29]), to the unbiased random sample of potentials studied in previous sections. I start by counting all pairs of natural amino acids in the MJ potential (i.e. 190), and classify them according to the definition in Fig.3. The MJ potential presents all 7 types of potentials analyzed in this work. Because of its continuous energy values, there are only six binary potentials with neutral interactions (i.e. type VI), and for convenience, I neglect them in this analysis. According to Fig. 3, an homogeneous sample from the space of binary potentials produces potentials type I:II:III:IV:V:VII in the ratio 2∶1∶1∶2∶1∶1. The histogram in Fig. 13, shows a comparison between expected versus observed types of binary potentials in the MJ energy function. This analysis shows that natural amino acids tend to avoid purely repulsive potentials (type VII) as much as they promote HP-like potentials (type I and II). Strikingly, there is a strong overrepresentation of potentials type III, approximately equivalent to the overall underrepresentation of potentials type IV and V. In order to gain further insights on the properties of binary combinations of natural amino acids, I perform a similar analysis as the one reported in Figure 4 (see Figure S15). In contrast to Fig. 4, the clustering of sequence-structure maps of binary potentials of natural amino acids seems more homogeneous. Similar to our previous observations, non-degeneracy ranges from 2 to 34% and encodability, from 1 to 21%. Several sequence structure maps reflect good foldabilities. These maps usually involve strong interactions such as Cys. Sequence-structure maps with different degrees of protein-likeness are observed in the case of potentials type I, II, IV and VI. The purpose of this analysis is not to argue that these binary potentials reflect the architecture of the sequence-structure map of natural proteins; but to suggest that, if the combination of potentials can be approximately considered additive, then combinations of these binary interactions may indeed reflect some of the properties of sequence-structure maps induced by larger alphabets. Indeed, random libraries composed primarily of 3 amino acids, such as AEK [68] and QLR [57], can be decomposed into potentials type I, II, IV; and I, I, V; respectively. In summary, these results show that the types of binary potentials observed in an unbiased sample of the space of energy functions, are represented in the interactions of natural amino acids, as described by the MJ potential. The types of potentials overrepresented in proteinaceous amino acids are potentials characterized as HP-like. Furthermore, these analysis suggest that random libraries enriched in HP-like potentials, are likely to favor protein-like sequence-structure maps. A graph theoretic approach, inspired on the concept of genotype-phenotype map, provides a common quantitative framework to investigate the sequence-structure relation. According to this framework, viable genotypes are represented as nodes, and edges connect genotypes that differ in a single position along the sequence. The distinction of genotypes according to the phenotypes they map onto, induces subgraphs whose properties and distribution have important consequences for biology. These subgraphs can be characterized quantitatively in terms of the statistics of their expected sizes, diameters and distances. I refer to this detailed characterization of the sequence-structure map, as its architecture. In this study I showed that the potential affects the architecture of the sequence-structure map and, that its impact on some of the map's properties is highly predictable based on features of the potential. First, the balance between attractive versus repulsive interactions in the potential, affects the available fraction of sequences and structures, and also induces biases towards compositional classes and the compactness of conformations. Second, although potentials induce similar sets of phenotypes, the identity of the most and least common phenotypes, differs. Third, potentials affect both the number, expected size, and the relative distribution of genotype components, neutral sets and neutral networks. Fourth, the overall symmetry of the potential, defined as the distribution of attractive and repulsive forces in homo versus heteromonomeric interactions, predicts the phenotypic diversity of genotype neighborhoods across divergent regions in sequence space. Fifth, foldability varies considerably across both potentials of different type, and potentials of the same type that preserve similar non-degeneracies and encodabilities. I observed that the predictability of a potential's non-additive interactions on the average foldability of a sequence-structure map, depends on the type of potential. Sixth, binary potentials of natural proteins, as defined by the MJ energy function, present biases that overrepresent HP-like potentials. In order to interpret these results in the context of the sequence-structure map of real peptides, one should be aware of the limitations of SEMs and the meaning of the energy terms in the potential. In the following, I discuss these limitations and evaluate critically the results presented in this study. Previous explorations of binary alphabets showed that repulsive interactions reduce the overall degeneracy of sequences, increasing the available fraction of viable genotypes and phenotypes [39]. The results in the present study confirm this observation and by distinguishing between homo and heteromonomeric interactions, show that non-degeneracy is promoted by potentials with predominantly attractive interactions (type II and III, Fig. 3 and Table 1), whereas encodability is only promoted by a combination of attractive homomonomers and repulsive heteromonomers (type II). A second observation anticipated by SEMs is the effect of repulsive interactions on the compactness of conformations [39]. Repulsive interactions tend to induce conformations with less number of contacts. The results in the present study reveal that not every repulsive interaction induces this effect. Indeed, only a combination of repulsive interactions at both homo and heteromonomeric interactions, reduces compactness (type I). In the case of repulsive homomonomer or heteromonomer interactions only, the effect is either none or opposite, respectively (type V and II). Previous studies pointed out that the effect of repulsive interactions is due to the avoidance of local energy minima and the distinction between conformations, by the induction of larger energy gaps [39]. The results in the present work confirm this intuition. Potentials with a larger average fraction of repulsive interactions show better foldabilities (see below). Several studies using different alphabet sizes, potentials, and polymer lengths, suggest that designability arises under a large variety of parameters [50], [69], [70]. Some of these studies, using maximally compact conformations, have shown that designability is affected by the potential and that, although different potentials induce a similar set of phenotypes, the most and least common phenotypes vary considerably across potentials [50]. The results presented here confirm these observations in the L18 model, with a full enumeration of the conformational space; and show that due to the differential induction of non-degenerate sequences and encodable conformations, potentials induce maps with variable degrees of designabilities. Similarly, I showed that the neutrality of networks presents analogous trends compared to the designability calculated over entire neutral sets. I showed that their relation depends on the type of potential. In the present study I explored three additional properties of sequence-structure maps, and their dependence on the potential energy function. Firstly, by considering the expected size of genotype components, neutral sets and neutral networks; I observed that potentials induce a large variation on the relative distribution of sequences and structures in genotype space. Strikingly, there are significant differences on the number of neutral networks per phenotype and the fraction of networks per genotype component across potentials. Secondly, as a consequence of different non-degeneracies and encodabilities observed across maps, as well as the variation of expected size of neutral sets and neutral networks, sequence-structure maps show considerable differences on the phenotypic diversity at divergent distances on genotype space. Thirdly, I used previous definitions of foldability, based on the energy gap, as a proxy to estimate the extent of protein-likeness across non-degenerate sequences. I observed that not every potential is equally likely to induce good folders. Most notably, non-additive potentials induce lower values of foldability. However, this prediction depends on the type of potential. Among these, are potentials that also show optimized levels of non-degeneracy and encodability (i.e. type I and II). Altogether, these results support previous observations on the distribution of sequences across genotype space based on the HP model [42]. HP-like potentials (i.e. type I and II), show on average small neutral networks that hardly reach diameters larger than 50% of genotype space. However, in contrast to the HP model, HP-like potentials are not always isolated in genotype space, but part of genotype components of large expected sizes. In part, this is due to the symmetry of the potential, that is, the proportion of attractive and repulsive interactions on homo versus heteromonomeric interactions. In practice, a symmetric potential is one in which interactions can be realized by more than one combination of monomers (i.e. redundant). Because the chemistry of the natural amino acid alphabet is known to be redundant, these observations imply that, as long as types of amino acid interactions in the structure are preserved, neutral networks (or at least, neutral sets) are likely to extend over divergent regions of genotype space. Previous, in silico analysis, support this observation [71], as do protein design strategies based on conservation patterns of hydrophobic-polar interactions [72]. The results presented here provide a rationale based on the proportion and types of interactions resulting from the monomer composition of sequences. As shown, this rationale makes predictions on the expected phenotypic diversity and the relative distribution of clusters of sequences. In addition, this framework makes further predictions about the distribution of sequences in genotype space and the role of structural determinants of sequence variation. For instance, it predicts the existence of larger neutral networks/sets in the case of structures with high degrees of symmetry. Indeed, studies exploring structural determinants of sequence variation show that designable folds are more symmetric than expected [73], [74]. Moreover, such a framework, suggests a strategy to improve fold assignment, a common task in comparative modeling [75] and in the identification of divergent homologous sequences [76]. This and similar predictions can be tested systematically in the case of proteins with long evolutionary histories, that encompass large superfamilies spanning divergent regions of genotype space (e.g. globins [77]; -barrels [78]). In extrapolating these observations to natural polypeptides one should take into account two relevant features of the potential, and evaluate how these features scale with the size of the potential. First, as suggested by previous studies, alphabet size has a fundamental impact on the fraction and diversity of accessible phenotypes [14], [79]. The observations presented here, suggest that a more accurate definition of alphabet size should account for the number and types of non-equivalent monomeric interactions. One might consider an effective alphabet size as the total number of chemically non-redundant pairwise interactions. Such a measure should account for differences between homo versus heteromonomeric interactions, and attractive versus repulsive. This represents a natural distinction between the types of potentials analysed in this work (Fig. 3, Table 1). In the case of binary potentials, hetero versus homomonomeric interactions are in a 1∶2 ratio. In general, with an alphabet size , hetero to homomonomeric interactions are in an (-1):2 ratio. Thus, in the case of natural proteins, there are approximately 9 hetero per each homomonomeric interaction. In addition, some of the types of potentials studied here, are more diverse in terms of attractive versus repulsive forces. Overall, because of the diversity of energy values, the alphabet of HP-like potentials must present indeed, large effective sizes. A second important aspect is to what extent, potentials composed of 2, can be considered simply as the additive contribution of independent binary potentials. Observations from simulation and empirical results, suggest that some of the properties presented above, for independent pairwise potentials, may apply to sequences composed of larger alphabets. Firstly, successful energy functions used to distinguish between native and non-native conformations, are based on the additive contribution of pairwise interactions [80]. Secondly, in silico mutational studies, show that changes in stability across different types of folds, are normally distributed [81]. This observation implies that most perturbations to the stability of protein structures, are additive. Thirdly, natural amino acids, as analyzed according to the MJ potential, overrepresent binary energy functions with HP-like features, as do natural sequences (unpublished data). A three-monomer alphabet may involve up to 3 different types of potentials; and a 4- and 5-monomer alphabet could, in principle, encompass up to 6 and 10 different types of potentials, respectively. Considering this observation, it is tempting to suggest an explanation as to why random libraries of polypeptides and protein folds designed using small alphabets, favor some types of potentials. For instance, libraries composed of mainly 3 amino acids such as AEK [68] and QLR [57], present I, II, IV and I, I, V potentials; respectively. Similarly, random libraries constructed of 5 amino acids, such as VADEG, composed of potential types: I, II, III, IV; in a 1∶2∶1∶1 ratio; show high levels of solubility and evidence of secondary structure formation [82]. A related empirical observation comes from the synthesis of protein folds using reduced alphabets. Riddle et al. [83], synthesized the SH3 fold using an alphabet of size 5: AIGEK. This alphabet includes potential types: I, II, III, IV; in a 3∶4∶2∶1 ratio, respectively. Fourthly, its has been recognized that non-native interactions play an important role during folding [84]. This suggests that although dominant, HP-like interactions would not be the only force required for successful folding, and would explain the relative lower representation of other types of potentials in reduced alphabets and in natural proteins. Other types of protein sequences may serve to test this hypothesis. Indeed, disordered proteins are known to be enriched in interactions that differ considerably to those commonly found in globular proteins [85]. Two sources of bias may appear when comparing the actual natural pairwise potentials to a random sample from the space of energy values. First, the chemistry of natural amino acids might cause an overrepresentation of pairwise potentials of certain types. Such bias might be explained by either biochemical constraints on the synthesis of a limited amino acid chemistry, or by the influence of natural selection on the amino acids introduced into the genetic code. A second source of bias, due to natural selection, is the differential usage of amino acids in natural proteins. From the proteinaceous amino acid pool, natural sequences might tune their composition and favor types of interactions that promote folding. Since the MJ potential was derived from the propensity of pairwise amino acid interactions in crystal structures of proteins, it might contain a mixture of these biases. The predominance of some types of potentials in natural proteins, as well as the empirical evidence of random libraries listed above, suggest the existence of constraints on the establishment of a primordial amino acid alphabet. Studies exploring the average solubility of random libraries have demonstrated a strong variation of protein-like features in these libraries, as a function of amino acid composition. Indeed, the so called primordial amino acids, have been shown to promote solubility and the formation of secondary structure [82]. The analysis presented here can be used to fully enumerate potentials that are likely to meet these constraints. Such analysis may provide a quantitative method to test the likelihood of reduced amino acid alphabets. Conversely, conjectures about the use of larger alphabets suggest the expansion of phenotype space [61]. In a forthcoming publication I explore larger amino acid alphabets, and quantitative ways of evaluating the effect of combinations of different types of potentials on the architecture of the sequence-structure map of natural proteins.

10.1371/journal.pbio.1001301

Stitching together Multiple Data Dimensions Reveals Interacting Metabolomic and Transcriptomic Networks That Modulate Cell Regulation

Cells employ multiple levels of regulation, including transcriptional and translational regulation, that drive core biological processes and enable cells to respond to genetic and environmental changes. Small-molecule metabolites are one category of critical cellular intermediates that can influence as well as be a target of cellular regulations. Because metabolites represent the direct output of protein-mediated cellular processes, endogenous metabolite concentrations can closely reflect cellular physiological states, especially when integrated with other molecular-profiling data. Here we develop and apply a network reconstruction approach that simultaneously integrates six different types of data: endogenous metabolite concentration, RNA expression, DNA variation, DNA–protein binding, protein–metabolite interaction, and protein–protein interaction data, to construct probabilistic causal networks that elucidate the complexity of cell regulation in a segregating yeast population. Because many of the metabolites are found to be under strong genetic control, we were able to employ a causal regulator detection algorithm to identify causal regulators of the resulting network that elucidated the mechanisms by which variations in their sequence affect gene expression and metabolite concentrations. We examined all four expression quantitative trait loci (eQTL) hot spots with colocalized metabolite QTLs, two of which recapitulated known biological processes, while the other two elucidated novel putative biological mechanisms for the eQTL hot spots.

It is now possible to score variations in DNA across whole genomes, RNA levels and alternative isoforms, metabolite levels, protein levels and protein state information, protein–protein interactions, and protein–DNA interactions, in a comprehensive fashion in populations of individuals. Interactions among these molecular entities define the complex web of biological processes that give rise to all higher order phenotypes, including disease. The development of analytical approaches that simultaneously integrate different dimensions of data is essential if we are to extract the meaning from large-scale data to elucidate the complexity of living systems. Here, we use a novel Bayesian network reconstruction algorithm that simultaneously integrates DNA variation, RNA levels, metabolite levels, protein–protein interaction data, protein–DNA binding data, and protein–small-molecule interaction data to construct molecular networks in yeast. We demonstrate that these networks can be used to infer causal relationships among genes, enabling the identification of novel genes that modulate cellular regulation. We show that our network predictions either recapitulate known biology or can be prospectively validated, demonstrating a high degree of accuracy in the predicted network.

Cells are complex molecular machines that employ multiple levels of regulation that enable them to respond to genetic and environmental perturbations. Advances in biology over the past several years to elucidate the complexity of this regulation have been truly astonishing. However, despite transformative advances in technology, it remains difficult to assess where we are in our understanding of cell regulation, relative to a complete comprehension of such a process. One of the primary difficulties in our making such an assessment is that the suite of research tools available to us seldom provides insights into aspects of the overall picture of the system that are not directly measured. While different technologies provide information that our analytical tools, both algorithmic and intellectual, seek to combine into a coherent picture, one of the primary limitations of the majority of analytical tools in use today is a focus on single dimensions of data, rather than on maximally integrating data across many different dimensions simultaneously to view processes more completely, thereby achieving a greater understanding of these processes. The full suite of interacting parts in a cell over time, if they could be viewed collectively, would enable our achieving a more complete understanding of cellular processes, much in the same way we achieve understanding by watching a movie. The continuous flow of information in a movie enables our minds to exercise an array of priors that provide context and constrain the possible relationships (structures), while our internal network reconstruction engine pieces all of the information together regarding the highly complex and nonlinear relationships represented in the movie, so that in the end we are able to achieve an understanding of what is depicted at a hierarchy of levels. If instead of viewing a movie as a continuous stream of frames of coherent pixels and sound, we viewed single dimensions of the information independently, understanding would be difficult if not impossible to achieve. For example, consider viewing a movie as independent, one dimensional slices through the frames of the movie, where each slice is viewed as pixel intensities across that one dimension changing over time (like a dynamic mass spec trace). In this way it would be very difficult to understand the meaning of the movie by looking at all of the one dimensional traces independently. Despite the complexity of biological systems, even at the cellular level, research in the context of large-scale high dimensional -omics data has tended to focus on single data dimensions, whether constructing coexpression networks on the basis of gene expression data, carrying out genome-wide association analyses on the basis of DNA variation information, or constructing protein interaction networks on the basis of protein–protein interaction data. While we achieve some understanding in this way, progress is limited because none of the dimensions on their own provide a complete enough context within which to interpret results fully. This type of limitation has become apparent in genome-wide association studies (GWAS), where many hundreds of highly replicated loci have been identified and highly replicated as associated with disease; but our understanding of disease is still limited because the genetic loci do not necessarily inform on the gene affected, on how gene function is altered, or more generally, how the biological processes involving a given gene are altered [1]–[4]. It is apparent that if different biological data dimensions could be formally considered simultaneously, we would achieve a more complete understanding of biological systems [2],[3],[5]–[7]. (See the documentary film The New Biology at http://www.youtube.com/watch?v=sjTQD6E3lH4.) Therefore, to form a more complete understanding of biological systems, we must not only evolve technologies to sample systems at ever higher rates and with ever greater breadth, we must innovate methods that consider many different dimensions of information to produce more descriptive models (movies) of the system. Methods are emerging that integrate pairs of data dimensions. For example, we recently developed methods that simultaneously integrate DNA variation and RNA expression data generated in a population context to identify coherent modules of interconnected gene expression traits driven by common genetic factors [2],[8]. In addition, many groups have begun incorporating a time dimension in the context of high-dimensional molecular-profiling data to elucidate how networks can transform over time [9],[10]. Here we develop and apply a network reconstruction approach that simultaneously integrates six different types of data: endogenous metabolite concentration, RNA expression, DNA variation, DNA–protein binding, protein–metabolite interaction, and protein–protein interaction data, to construct probabilistic causal networks that elucidate the complexity of cell regulation (Figure 1). The goals of our integrative analysis are not only to find causal regulators underlying expression quantitative trait loci (eQTL) hot spots, but to uncover mechanisms by which these predicted causal regulators affect genes and metabolites whose transcriptional profiles or metabolite profiles are linked to the eQTL hot spots. We leveraged a previously described cross between laboratory (BY) and wild (RM) yeast strains (referred to here as the BXR cross) for which DNA variation and RNA expression had been assessed [11],[12], to carry out a quantitative metabolite profiling using quantitative NMR (qNMR) under the same experimental conditions as the gene expression study [12]–[14]. We demonstrate that, like transcript and protein levels, concentrations of many metabolites are strongly linked to metabolite QTLs (metQTLs). Several of the metQTLs are seen to colocalize with expression quantitative trait loci (eQTLs) previously identified in the same yeast population [13], enabling us to infer causal relationships between metabolites and expression traits [13],[14]. Then, by extending a previously described Bayesian network (BN) reconstruction algorithm [13], we constructed a probabilistic causal network by integrating metabolite levels, genotype, gene expression, transcription factor (TF) binding, and protein–protein interaction data. The resulting network not only validates the functional importance of eQTL hot spots in the BXR cross, but elucidates the mechanisms by which variation in DNA at eQTL hot spots affect gene expression. By systematically using the networks to elucidate the regulators of these eQTL hot spots, we are not only able to recapitulate known regulatory mechanisms, we are able to provide a number of novel and experimentally supported causal relationships predicted by our network, including that cellular amino acid concentrations are related to both amino acid biosynthesis pathways and amino acid degradation pathways, with VPS9 predicted and prospectively validated as a key driver of a previously identified eQTL hot spot that could not previously be well characterized. In addition, we further experimentally demonstrated that PHM7, a previously predicted and validated causal regulator for stress response genes whose expression variations are linked to the PHM7 locus on Chromosome XV, affected trehalose, a yeast metabolite product of the stress response pathway. These results combined not only help uncover the mechanisms by which gene expression profiles are regulated by metabolite profiles, but they also confirm the importance of gene expression in understanding system-wide variation linked to genetic perturbations. Given the strong genetic signal detected in the metabolite data and the coincidence of metQTL and eQTL hot spot regions, we set out to explore an integrated network analysis strategy using the gene expression profiles [11] as well as the metabolite data described above. Gene expression and metabolite traits were treated equivalently as nodes in our BN reconstruction process. As such, we modified our previously reported BN reconstruction method [13] to accommodate metabolite data, in addition to genotype, gene expression, protein interaction, and TF–DNA binding data. The KEGG biochemical pathway database [37] was used to generate structure priors between metabolites and genes encoding enzymes known to be involved in biochemical reactions in canonical pathways. Intuitively, genes encoding enzymes that directly catalyze biochemical reactions for the metabolites were assigned stronger prior probabilities of being related during network reconstruction, whereas genes that encode enzymes catalyzing downstream or upstream biochemical reactions of the metabolites were assigned weaker priors (see Methods for details). Differentially regulated genes and the structure priors for genotype, TF–DNA, and protein–protein interaction data were defined as previously described [13]. The 56 reliably quantified metabolites were included as input into the BN reconstruction program. From this probabilistic causal network we can identify subnetworks for all of the metabolites or any set of genes (see Methods for details). To assess the predictive power of this network, we examined how metabolites and gene expression traits relate to one another at the four eQTL hot spots in Table 1, providing for the possibility of elucidating regulatory mechanisms and generating testable hypotheses about novel regulatory relationships. By integrating six different fundamental types of data, including RNA expression, DNA variation, DNA–protein binding, protein–metabolite interaction, and protein–protein interaction data, with metabolite data, we constructed a BN using an approach that simultaneously considers all of these data, with the resulting network providing a number of novel insights into the mechanisms of the eQTL hot spots in a segregating yeast population (the BXR cross). Importantly, we validated the biological consequences of the transcriptional variation linked to each of the four eQTL hot spots identified in the BXR cross to which metabolite levels were also linked. Our results indicate that the incorporation of metabolite levels into the network reconstruction process significantly enhanced the utility of the network-based models [46],[47]. While the integration of metabolite abundance and gene expression traits in a genetic context have been attempted in plants [48] and mouse [49], the main distinguishing characteristic of our study is the de novo construction of a global molecular network that simultaneously incorporates many different types of information (DNA, RNA, protein, and metabolite), along with known biochemical pathways as prior information. To aid in further understanding how we integrate these data to construct probabilistic causal networks, and to enhance the ability to repeat our results, we provide as Text S1 results of an in-depth description of the construction of the URA3 subnetwork (Figure 4), using different types of data to assess the contributions of different data types to the predictive power of the network and to the identification of key modulators of important biological processes. We examined in detail all 4 eQTL hot spots that coincided with metQTLs. Our findings for eQTL hot spots 1 and 2 recapitulated well-known biological processes, and for eQTL hot spots 3 and 4 our predictions implicated novel genes as modulators of established biological processes, which we subsequently validated prospectively. Among the many predictions made by our network, we uncovered novel insights into the biological processes that in the BXR cross are responsible for variations in amino acid levels. While amino acid concentrations are known to be regulated by multiple processes (e.g., synthesis, degradation, recycle, and storage), our approach objectively identified that variations in concentrations of a number of amino acids in the BXR cross were affected by both the amino acid biosynthesis and degradation pathways. We predicted and prospectively validated VPS9 as a major driver of amino acid concentrations via the amino acid degradation pathway. These results open novel and interesting questions about the mechanism by which sequence variation at this locus affects phenotype. VPS9 is involved in vesicle-mediated vacuolar protein transport, and in Saccharomyces cerevisiae, the vacuole is the main compartment for amino acid storage, recycling, and cytosolic amino acid concentration maintenance [50]. The cellular effects of variation in VPS9 are likely mediated by differential regulation of amino acid storage in the vacuole; we speculate that such storage changes may affect cytosolic amino acid pools that in turn have downstream consequences on transcript and protein levels of amino acid pathways, as has been shown for CHA1 [40] and GCV3 [51]. However, only with enhanced screening of all molecular states of the systems can we achieve a complete understanding of these processes. Thus, while the integrated BN elucidated some of the mechanistic underpinnings of the eQTL hot spots in the BXR cross, additional information will be required to more fully understand how processes perturbed in the BXR cross lead to phenotypic changes. Despite lacking an exhaustive assessment of all molecular traits in the BXR cross, it is of particular note that the strong correlations we observed between gene expression and metabolite data may help resolve an ongoing debate regarding the functional consequences of gene expression regulation. While some reports indicate that gene expression levels and protein abundances are not well correlated [52], other reports indicate a high degree of correlation [53]. A recent proteomic study in the BXR cross demonstrated that a large number of protein levels are linked to eQTL hot spots [34], two of which (the eQTL hot spots 1 and 3) were highlighted in our present work. Metabolites are the final functional products of protein activity regulation. We showed that PHM7 not only alters expression levels of stress response genes linked to eQTL hot spot 4, but also alters the abundance of trehalose, a metabolite product of the stress response genes. Our results demonstrate that gene expression and metabolite levels are not only strongly correlated, but that a significant proportion of that covariation can be explained by common genetic control. Given that variations in protein levels can result from sequence-specific transcriptional and translational regulation or from nonsequence-specific protein degradation, the integration of gene expression and metabolic traits can help dissect the complex processes that regulate protein levels. The yeast growth conditions for metabolite profiling were the same as previously used to generate the gene expression data in the BXR cross [12]. Both gene expression and metabolite abundances are under strong genetic regulation and are linked to common eQTL hot spots (Table 1). When metabolite data were integrated with gene expression data, our resulting integrated network was able to recapitulate the mechanism of multiple known biological processes that in turn explained the connection between genes linked to the LEU2 locus and genes with Leu3 binding sites, with the metabolite 2-isopropylmalate objectively identified as the key intermediate. These results also confirmed that changes in expression of stress response genes lead to changes in stress response metabolites such as trehalose. Therefore, the integration of the gene expression and metabolite data has provided new insight into common biological processes that are perturbed by genetic variation segregating in the BXR cross. Going forward, as more technologies emerge that can generate large-scale data in different dimensions for low cost, we will achieve a more complete understanding of biological systems only if we integrate all of the information together to consider all of the different cellular components and how they interact with one another at the population level. For example, comprehensive proteomic data and protein phosphorylation data are needed and should be further integrated with other high throughput genomic and genetic data. For metabolites, their cellular abundances are not only affected by specific enzymes in related biochemical reactions, but they are also affected by proteins that bind them or transport them into different cellular compartments. Further research on how to integrate these data into networks is needed. In addition, there is an abundance of existing knowledge, such as genetic interactions and regulatory cascades, which can be converted into prior information and integrated with other data and priors. Further efforts in developing methods to integrate these diverse data and information are warranted. In more complex systems, we will need to consider the fundamental building blocks of a cell in the context of cell–cell interactions that lead to tissue-based networks, the interactions of tissues that lead to organ-based networks, and the interactions of organs in a given system to understand the physiological states of that system associated with complex phenotypes of interest, given these phenotypes emerge from this complex web of interacting networks [54]. Only by taking the full complement of raw data available on living systems can we move from the accumulation of knowledge to actual understanding, and from understanding, wisdom. Yeast parental strains BY4716 (MATα lys2Δ0) and RM11-1a (MATa leu2Δ0 ura3Δ0 HO:kan) and 111 segregants of BXR cross [11] were provided by R. Brem. Auxotrophies, mating type, and G418 resistance were confirmed for all strains to be as previously reported [12]. Cells were grown under identical conditions as previously described [12]. Strains were freshly started from freezer stocks and stored at room temperature on synthetic complete medium plates for no longer than 1 wk before each experimental run. For each run, cells from the plates were precultured in 10 ml of synthetic complete media (Table S8) at 30°C with shaking for 24 h. These cultures were then diluted into 25 ml fresh synthetic complete media to an optical density of 0.005 to 0.02. This starting density was determined from previous growth rate measurements and empirical observations such that after overnight growth at 30°C, the cultures would be exponentially growing, i.e., at a cell density of less than 2×107 cells/ml. Overnight cultures were diluted into 52 ml fresh synthetic complete medium to an optical density of 0.1, and incubated with shaking for approximately 5 h at 30°C. Starting at 3 h after dilution, optical density was monitored every 60 min. Cell suspensions were counted in a hemocytometer to obtain cell count per OD values and an estimate of cell-doubling time. Since some of the yeast strains produced flocculent cultures under these growth conditions, all cultures were diluted 5× into 0.25 ml PBS and sonicated three times on ice for 45 s using a Misonix sonicator 3000 equipped with a microprobe before optical density was determined and/or cells were counted. At an optical density of approximately 1.0, each exponentially growing culture was concentrated 10-fold by rapid centrifugation at room temperature and suspension of the cells in 5 ml of synthetic complete medium prewarmed to 30°C. These concentrated cell suspensions were then incubated at 30°C with shaking for 1 h. Metabolites were then immediately extracted from the cells in these concentrated suspensions. Yeast parental strain BY4742 (MATα his3Δ1 leu2Δ0 met15Δ0 ura3Δ0) and six deletion strains derived from it (Δtsl1::kanMX, Δnup188::kanMX, Δcac2::kanMX, Δyml096w::kanMX, Δvps9::kanMX, and Δarg81::kanMX) were provided by Elton Young's lab, Department of Biochemistry, University of Washington, from a copy of the Yeast Deletion Consortium knockout collection prepared in Stanley Fields' lab, Department of Genome Sciences, University of Washington. Cells were grown under identical conditions as the BXR cross strains in synthetic complete medium, and metabolite extracts were also obtained and further processed in identical fashion (see below). Each experiment was repeated on three different days. Yeast parental strain BY4743 (MATa/MATα his3Δ1/his3Δ1 leu2Δ0/leu2Δ0 lys2Δ0/+ met15Δ0/+ ura3Δ0/ura30) was obtained from ATCC (Manassas, Virginia), and the derived PHM7 knockout strain 31775 (phm7::KanMX/phm7::KanMX) constructed by the Yeast Deletion Project [55] was obtained from Open Biosystems (Huntsville, Alabama). Cells were grown under identical condition as the PHM7 knockout gene expression experiment [13], and metabolites were extracted as described below. Each experiment was repeated on three different days. BY4742 and Δvps9 strains (both MATα) were grown as described above and harvested by centrifugation in crushed ice when cells reached optical density of approximately 1.0. Total RNA was extracted using RNeasy mini-columns, transcribed with SMARTScribe Reverse Transcriptase (Clontech) from oligo(dT), and diluted 1,000×. Real-time PCR was run for 17 genes (including VPS9) associated with the Chromosome XIII eQTL hot spot subnetworks and ACT1 internal standard gene (Table S6) on an ABI 7900HT instrument with 2× Sensimix dT (Quantance), primers at 0.2 µmol/l, and SYBR Green reagent. Relative expression was calculated using the ΔΔCt method with ACT1 internal standards [56]. TAF9 was used to estimate the false positive rate as 0.033. Intracellular metabolites were extracted using a modification of previously described methods [31],[57]. First, all intracellular metabolic processes were rapidly quenched by pipetting each concentrated cell suspension into 20 ml of rapidly mixing 60% (v/v) methanol at −40°C. Cells were rapidly (5 min) sedimented in a centrifuge precooled to −8°C and washed twice with 20 ml of the −40°C methanol. Metabolites were then extracted with boiling 75% (v/v) ethanol at 80°C and 0.25 ml dry volume of acid-washed glass beads (Sigma G1277), by vigorous vortexing for 30 s. The cell-glass bead slurry was incubated 3 min at 80°C, vortexed 30 s, and then placed on ice for 5 min. Large cellular debris and glass beads were removed by centrifugation at 2,000 g for 5 min. The resulting ethanolic extracts were clarified by three rounds of centrifugation at 14,000 g in a microcentrifuge. The clarified metabolite extracts were stored at −80°C until drying. Extracts were dried in a Savant Speed Vac under 150 mtorr vacuum in low retention microcentrifuge tubes. Dried metabolite extracts were stored at −80°C until preparation for NMR analysis. The process of NMR spectra acquisition and quantification follows the previously outlined procedure [29]. Dry metabolite extracts were dissolved in 0.7 ml deuterated 80 mM potassium phosphate buffer (containing 2 mM DSS-d6 as an internal reference standard), and transferred to 100-mm 5-ml NMR tubes. NMR samples were stored in Varian 768AS auto-sampler at 8°C before and after NMR analyses. NMR data were acquired on the Varian 700 MHz NMR spectrometer at 25°C with one-dimensional proton pulse sequence. The water peak was suppressed by the WET pulse sequences [58]. For each sample, 512 acquisitions were acquired with 3 s of acquisition and 15 s of delay between pulses. Analyses of NMR spectra were carried out using DataChord Spectrum Miner (One Moon Scientific, Inc.). Stacked NMR spectra were referenced to DSS-d6 as 0 ppm, and peaks of each endogenous metabolite were checked against their reference spectra (about 700 common endogenous metabolites). Each metabolite usually displays multiple peaks, for example trehalose, shown in Figure S6. Overlapping peaks were quantified by peak area correction according to stoichiometric peak ratios for each metabolite. For three correlated variables , , and, there are three groups of causal/reactive relationships among them as the following:For example, the three graphs in the group G1, , describe the same set of condition independent relationship that and are independent conditioning on . The three graphs have the same probabilities and are called Markov equivalent. In an F2 cross, we can represent quantitative traits as and , and the genetic locus as . In an F2 cross experimental design, all F2 strains are under the same experimental condition. Therefore, the only source of variation in the quantitative traits and are genetic differences in , so that the relationships and are plausible. On the other hand, the genetic variation in is stable and does not change during an F2 cross experiment, so that and are not plausible. Thus in an F2 cross, only one graph in each of the three Markov equivalent groups above is plausible. We can simplify the above three groups as follows:where the genetic locus is the anchor in each causal/reactive relationship in a F2 cross. For two quantitative traits and linked to the same locus in the yeast cross, there are three basic relationships that are possible between the two traits relative to the DNA locus as described above. Either DNA variations at the locus lead to changes in trait that in turn lead to changes in trait , or variations at locus lead to changes in trait that in turn lead to changes in trait , or variations at locus independently lead to changes in traits and , as previously described [14]. Assuming standard Markov properties for these basic relationships, the joint probability distributions corresponding to these three models, respectively, are: where the final term on the right-hand side of equation M3 reflects that the correlation between and may be explained by other shared loci or common environmental influences, in addition to locus . We assume Markov equivalence between and for model M3 so that . is the genotype probability distribution for locus and is based on a previously described recombination model [59]. The random variables and are taken to be normally distributed about each genotypic mean at the common locus , so that the likelihoods corresponding to each of the joint probability distributions are then based on the normal probability density function, with mean and variance for each component given by: (1) for the mean and variance are and , (2) for the mean and variance are and , and (3) for the mean and variance are and , where represents the correlation between and , and and are the genotypic specific means for and , respectively. The mean and variance for follow similarly from that given for . From these component pieces, the likelihoods for each model are formed by multiplying the densities for each of the component pieces across all of the individuals in the population [14]. The likelihoods are then compared among the different models in order to infer the most likely of the three. Because the number of model parameters among the models differs, a penalized function of the likelihood was used to avoid the bias against parsimony. The model with the smallest value of the penalized statisticwas chosen. Here, is the maximum likelihood for the ith model, pi is the number of parameters in the ith model, and k is a constant. In this instance we took the penalized statistic to be the Bayesian Information Criteria (BIC) where k is set to , with n denoting the number of observations. BNs are directed acyclic graphs in which the edges of the graph are defined by conditional probabilities that characterize the distribution of states of each node given the state of its parents [60]. The network topology defines a partitioned joint probability distribution over all nodes in a network, such that the probability distribution of states of a node depends only on the states of its parent nodes: formally, a joint probability distribution on a set of nodes can be decomposed as , where represents the parent set of . In our networks, each node represents a quantitative trait that can be a gene or a metabolite. These conditional probabilities reflect not only relationships between genes, but also the stochastic nature of these relationships, as well as noise in the data used to reconstruct the network. Bayes formula allows us to determine the likelihood of a network model M given observed data D as a function of our prior belief that the model is correct and the probability of the observed data given the model: . The number of possible network structures grows super-exponentially with the number of nodes, so an exhaustive search of all possible structures to find the one best supported by the data is not feasible, even for a relatively small number of nodes. We employed Monte Carlo Markov Chain (MCMC) [61] simulation to identify potentially thousands of different plausible networks, which are then combined to obtain a consensus network (see below). Each reconstruction begins with a null network. Small random changes are then made to the network by flipping, adding, or deleting individual edges, ultimately accepting those changes that lead to an overall improvement in the fit of the network to the data. We assess whether a change improves the network model using the BIC [62], which avoids overfitting by imposing a cost on the addition of new parameters. This is equivalent to imposing a lower prior probability on models with larger numbers of parameters. Even though edges in BNs are directed, we can't in general infer causal relationships from the structure directly. For example, in a network with two nodes, and , the two models and have equal probability distributions as . Thus, with correlation data itself, we can't infer whether is causal for or vise versa. In the more general case, for a network with three nodes, , , and , there are multiple groups of structures that are mathematically equivalent. For example, the following three different models, , , and , are Markov equivalent (which means that they all encode for the same conditional independent relationships). In the above case, all three structures encode the same conditional independent relationship, , and are independent conditioning on , and they are mathematically equalThus, we can't infer whether is causal for or visa-versa from these types of structures. However, there is a class of structures, V-shape structures (e.g., ), which have no Markov equivalent structure. In this case, we can infer causal relationships. There are more parameters to estimate in the Mv model than M1, M2, or M3, which means a large penalty in the BIC score for the Mv model. In practice, a large sample size is needed to differentiate the Mv model from the M1, M2, or M3 models. The same 3,662 informative genes used previously [13] and 56 metabolites were included in the network reconstruction process using a BN reconstruction software program based on a previously described algorithm [63],[68] as outlined above. One thousand BNs were reconstructed using different random seeds to start the reconstruction process. From the resulting set of 1,000 networks generated by this process, edges that appeared in greater than 30% of the networks were used to define a consensus network. Our previous simulation study shows that the 30% inclusion threshold results in a stable structure and achieves the best tradeoff between precision and recall [68]. The histogram of percentage of occurrences of all potential edges shows that 30% is a reasonable cutoff threshold for inclusion (Figure S8). Edges in this consensus network were removed if (1) the edge was involved in a loop, and (2) the edge was the most weakly supported of all edges making up the loop. The genetic, TFBS, and PPI data were used to derive structure priors as previously described (details described above in Methods) [13]. Structure priors for metabolites and genes are derived from KEGG chemical reactions as described above. All data and software used to construct the BNs described herein are available at http://www.mssm.edu/research/institutes/genomics-institute/rimbanet. Subnetworks for sets of genes were constructed as follows. Genes in the input set were used as seeds and the direct neighbors of seeds were identified. Seeds and their direct neighbors define the nodes of a given subnetwork. Links between nodes in the subnetworks are the same as in the complete BN.

10.1371/journal.ppat.1005611

Antibiotic-Resistant Neisseria gonorrhoeae Spread Faster with More Treatment, Not More Sexual Partners

The sexually transmitted bacterium Neisseria gonorrhoeae has developed resistance to all antibiotic classes that have been used for treatment and strains resistant to multiple antibiotic classes have evolved. In many countries, there is only one antibiotic remaining for empirical N. gonorrhoeae treatment, and antibiotic management to counteract resistance spread is urgently needed. Understanding dynamics and drivers of resistance spread can provide an improved rationale for antibiotic management. In our study, we first used antibiotic resistance surveillance data to estimate the rates at which antibiotic-resistant N. gonorrhoeae spread in two host populations, heterosexual men (HetM) and men who have sex with men (MSM). We found higher rates of spread for MSM (0.86 to 2.38 y−1, mean doubling time: 6 months) compared to HetM (0.24 to 0.86 y−1, mean doubling time: 16 months). We then developed a dynamic transmission model to reproduce the observed dynamics of N. gonorrhoeae transmission in populations of heterosexual men and women (HMW) and MSM. We parameterized the model using sexual behavior data and calibrated it to N. gonorrhoeae prevalence and incidence data. In the model, antibiotic-resistant N. gonorrhoeae spread with a median rate of 0.88 y−1 in HMW and 3.12 y−1 in MSM. These rates correspond to median doubling times of 9 (HMW) and 3 (MSM) months. Assuming no fitness costs, the model shows the difference in the host population’s treatment rate rather than the difference in the number of sexual partners explains the differential spread of resistance. As higher treatment rates result in faster spread of antibiotic resistance, treatment recommendations for N. gonorrhoeae should carefully balance prevention of infection and avoidance of resistance spread.

More and more infectious disease treatments fail because the causative pathogens are resistant to the drugs used for treatment. For the treatment of Neisseria gonorrhoeae, a sexually transmitted bacterium, drug resistance is a particularly big problem: there is only a single antibiotic left that is recommended for treatment. We aimed to understand how antibiotic-resistant N. gonorrhoeae spread in a sexually active host population and how the spread of resistance can be slowed. From antibiotic resistance surveillance data, we first estimated the rate at which antibiotic-resistant N. gonorrhoeae spread. Second, we reproduced the observed dynamics in a mathematical model describing the transmission between hosts. We found that antibiotic-resistant N. gonorrhoeae spread faster in host populations of men who have sex with men than in host populations of heterosexuals. We could attribute the faster spread of resistant pathogens to higher treatment rates. This finding implies that promoting screening to control antibiotic-resistant N. gonorrhoeae could in fact accelerate their spread.

Antibiotic-resistant Neisseria gonorrhoeae can evolve and spread rapidly [1]. Resistance is commonly observed against the antibiotic classes penicillin, tetracycline and fluoroquinolones [2–4]. Resistance also emerged against cefixime, an oral third generation cephalosporin, in recent years [2, 3]. Since 2010, cefixime is no longer recommended as first-line treatment [5] following guidelines from the World Health Organization (WHO) that an antibiotic should not be used when more than 5% of N. gonorrhoeae isolates are resistant [6]. Injectable ceftriaxone, in combination with oral azithromycin, is now the last antibiotic remaining as recommended first-line treatment [7]. Although other antibiotics are being tested for their safety and efficacy for N. gonorrhoeae treatment [8], no new classes of antibiotics are currently available [4] and management of antibiotics is urgently needed to preserve their efficacy. The current management strategy tries to reduce the overall burden of N. gonorrhoeae infection by expanded screening and treatment of hosts [9, 10], but the outcome of this strategy for resistance is uncertain. Understanding the drivers of resistance spread and anticipating future resistance trends will provide rationales for antibiotic management and help to improve antibiotic treatment strategies. Men who have sex with men (MSM) are host populations that have higher levels of antibiotic-resistant N. gonorrhoeae than heterosexual host populations [3]. In a study [5] based on the Gonococcal Resistance to Antimicrobials Surveillance Programme (GRASP) in England and Wales, cefixime-resistant N. gonorrhoeae were mainly found in MSM until 2011. The authors suggested that cefixime resistance was circulating in a distinct sexual network of highly active MSM and that bridging between MSM and heterosexuals was necessary for subsequent spread among heterosexual hosts. However, cefixime-resistant N. gonorrhoeae might have already been spreading undetected in the heterosexual host population. Mathematical models can help explain the differential observations of antibiotic-resistant N. gonorrhoeae in different host populations. In 1978, Yorke et al. [11] introduced the concept of core groups to model the transmission of N. gonorrhoeae. The concept of core groups posits that an infection can only be maintained in a host population if a highly sexually active group of hosts is responsible for a disproportionate amount of transmissions. More recent modeling studies have examined the transmission of antibiotic-resistant N. gonorrhoeae. Chan et al. [12] found that prevalence rebounds more quickly to a pre-treatment baseline when treatment is focused on the core group. Xiridou et al. [13] developed a N. gonorrhoeae transmission model to determine the impact of different treatment strategies on the prevalence of N. gonorrhoeae in Dutch MSM. They found that increased treatment rates could increase the spread of resistance, whereas re-treatment could slow it down. Hui et al. [14] used an individual-based N. gonorrhoeae transmission model in a heterosexual host population to investigate the effect of a molecular resistance test on the time until 5% resistance are reported. None of these studies has investigated or explained the differences in the spread of antibiotic-resistant N. gonorrhoeae in MSM and heterosexual host populations. In this study, we investigated the dynamics and determinants of antibiotic-resistant N. gonorrhoeae spread using surveillance data and mathematical modeling. We estimated the rates at which resistance spreads in heterosexual men (HetM) and MSM using surveillance data from the USA and from England and Wales. We then developed a mathematical model of N. gonorrhoeae transmission to reconstruct the observed dynamics of resistance spread. This allowed us to determine the major driver of resistance spread, and to explore the expected rates at which resistance spreads in MSM and heterosexual host populations. We fitted a logistic growth model to the proportion of antibiotic-resistant N. gonorrhoeae as observed in the two gonococcal surveillance programs (Fig 2). The proportion of cefixime-resistant N. gonorrhoeae in GRASP appears to increase for both HetM and MSM after 2006. Ciprofloxacin-resistant N. gonorrhoeae in HetM and MSM were spreading in all observed host populations after the year 2000. For a given antibiotic and surveillance program, the rates of resistance spread were consistently higher for MSM than for HetM (Table 4). The average rate of resistance spread was 0.53 y−1 for HetM and 1.46 y−1 for MSM, corresponding to doubling times of 1.3 y (HetM) and 0.5 y (MSM) during the initial exponential growth phase. Next, we studied the transmission of N. gonorrhoeae and the spread of resistance in the dynamic transmission model. We calibrated five model parameters to expected prevalence and incidence in MSM and HMW host populations. The posterior distributions of the parameters were based on 2,779 parameter sets for HMW and 65,699 parameter sets for MSM (Fig 3, Table 1). Distributions of the modeled prevalence and incidence of diagnosed infections after calibration are provided as Supporting Information (S1 and S2 Figs, S3 Table). The sexual mixing coefficient showed a tendency towards assortative mixing in both MSM and HMW (Fig 3a). The fraction of diagnosed and treated infections tended to be higher in MSM compared to HMW (Fig 3b), whereas the infectious duration was considerably shorter in MSM (median: 2.3 months, IQR: 1.7–3.0 months) than in HMW (median: 6.6 months, IQR: 5.5–7.9 months) (Fig 3c). The transmission probabilities per partnership were generally higher in HMW than in MSM (Fig 3d and 3e). After calibration, we used the dynamic transmission model to study the spread of antibiotic-resistant N. gonorrhoeae. The proportion of antibiotic-resistant N. gonorrhoeae increased faster in MSM than in HMW (Fig 4). In HMW, the median of all simulations reached 5% resistance in fewer than 4.5 y and 50% resistance in fewer than 7.8 y after appearance of the first antibiotic-resistant N. gonorrhoeae infection. In the MSM population, the median of all simulations reached a resistance level of 5% in fewer than 1.7 y and 50% in fewer than 2.6 y after resistance first appears in the population. The range spanned by all simulations was much wider in HMW than in MSM: 95% of all simulations reached the 5% threshold in fewer than 2.7–7.7 y (HMW), compared with 1.1–2.2 y (MSM). Antibiotic-sensitive and -resistant N. gonorrhoeae share the same resource for growth, i.e. the susceptible hosts. The rate at which one strain replaces the other strain in the host population is given by the difference in their net growth rates. We assume that the transmission probabilities and the infectious duration of the two strains are the same. Since the probability of resistance during treatment is very small (μ ≪ 1), the difference in net growth rates between the strains is approximated by the treatment rate τ and corresponds to the rate of spread of antibiotic-resistant N. gonorrhoeae. The observed distributions of treatment rates from the transmission model hardly overlap between HMW and MSM (Fig 5). The median treatment rates, i.e. the approximated median rates of resistance spread in the transmission model are 3.12 y−1 (MSM) and 0.88 y−1 (HMW). We tested whether changes in the probability of resistance during treatment, μ, and fitness costs in the antibiotic-resistant strain alter the model outcomes. Higher probabilities of resistance during treatment accelerate the establishment of antibiotic-resistant N. gonorrhoeae in the population and hence reduce the time until 5% resistance is reached (S3 Fig). Higher probabilities of resistance during treatment, however, do not affect rates of spread, unless the probability of resistance during treatment is unrealistically high (10%) (S4 Fig). Fitness costs in the antibiotic-resistant strain result in rates of resistance spread that are lower than the treatment rate τ (Fig. B in S1 Appendix). Fitness costs that reduce the transmission probability per partnership, βij, have a stronger effect than fitness costs that reduce the duration of infection. The effects of fitness costs are independent of the sexual partner change rate, πi, and βij if they affect the duration of infection, but can vary with πi and βij if they affect the transmission probability per partnership (Fig. C in S1 Appendix). While high fitness costs can prevent the spread of antibiotic-resistant strains (Fig. A in S1 Appendix), fitness costs between 0%–10% have only small effects on the rates of resistance spread (Fig. B in S1 Appendix). In this study, we quantified the rate at which antibiotic-resistant N. gonorrhoeae spread in heterosexual and MSM populations. We used data from two different surveillance programs and estimated that the proportion of ciprofloxacin- and cefixime-resistant N. gonorrhoeae doubles on average every 1.3 y in HetM and 0.5 y in MSM. The faster spread of antibiotic-resistant N. gonorrhoeae in MSM than in heterosexual hosts was corroborated using a dynamic transmission model, which was calibrated to observed prevalence and incidence rates. The model allowed us to identify the higher treatment rates in MSM, compared with heterosexual hosts, as the major driver for the faster spread of antibiotic-resistant N. gonorrhoeae. To our knowledge, this is the first study to have analyzed and interpreted N. gonorrhoeae antibiotic resistance surveillance data in a dynamic and quantitative manner. The transmission model was parameterized using sexual behavior data for HMW and MSM from Natsal-2 [22], a large probability sample survey of sexual behavior. Calibrating the model to observed prevalence and incidence rates allowed us to use largely uninformative priors for the model parameters. The calibration makes our model more robust to changes in parameters than using fixed parameter values, especially since for N. gonorrhoeae available parameter values are very uncertain [31]. It also allowed us to rely on few assumptions about the natural history of N. gonorrhoeae infection. The limitations to our study need to be taken into consideration when interpreting the findings. First, we used data from different sources, although all were collected in high income countries. The antibiotic resistance surveillance data are from programs in England and Wales and the USA. The mathematical transmission model was parameterized using British sexual behavior data [22] and calibrated to prevalence and incidence rates from the USA (HMW) [26, 27] and Australia (MSM) [28, 29]. For simplicity, we modeled the heterosexual and MSM host populations separately although there is some mixing between them. We assumed the sexual behavior of heterosexual men and women to be the same and pooled their behavioral data. Second, we assumed complete resistance against the antibiotic, i.e. 100% treatment failure. We further assumed that treatment of the sensitive strain is 100% efficacious. Both assumptions might explain why antibiotic-resistant N. gonorrhoeae spread at somewhat higher rates in the dynamic transmission model than estimated from data. Third, we restricted our model to resistance to one antibiotic with no alternative treatment or interventions. This is why we observe complete replacement of the antibiotic-sensitive strain in the model, a phenomenon that has not been observed in surveillance data. Fourth, resistance in our model is treated as a generic trait, but it likely depends on the underlying molecular mechanisms and possibly the genetic background of the N. gonorrhoeae strain. Different resistance mechanisms might explain some of the differences in the rates of resistance spread between the model and the different antibiotics from the surveillance data. Fifth, we did not include co- and superinfection with antibiotic-sensitive and -resistant N. gonorrhoeae strains. Since genetic typing provides evidence for mixed infections [32], it is worth speculating how they would affect the rate of spread from the transmission model. If antibiotic-sensitive and -resistant strains co-existed in a host and acted independently, we would not expect significant effects on the rate of spread. In contrast, if there was competition between the two strains within a host, the rate of spread would increase if the antibiotic-resistant strain outcompetes the -sensitive strain, and decrease otherwise. Sixth, we do not consider importation of resistance from another population. For example, importation of resistance from other countries might play a particularly important role during the early phase of resistance spread, when stochastic events can lead to extinction of the antibiotic-resistant strain. We expect that a high rate of importation of antibiotic resistance shortens the time to reach 5% resistance drastically, but that once the resistant strain is established in the population, importation hardly affects the rate of resistance spread. Finally, we assumed that the transmission probabilities per partnership and the durations of infection in the model represent average values for N. gonorrhoeae infections at different infection sites (urethral, pharyngeal, anal, cervical). The estimated posterior distributions of the parameters fit within the range of previously used values, and provide some insights into sexual mixing and the natural history of N. gonorrhoeae. The sexual mixing coefficient tends to be assortative for both HMW and MSM host populations in our model. Quantifying the degree of sexual mixing is difficult and largely depends on the study population, but our finding is consistent with other studies indicating assortative sexual mixing in the general population [30, 33]. The posterior estimates of the fraction of diagnosed and treated infections are consistent with the notion that a large proportion of N. gonorrhoeae infections are symptomatic, and that this proportion is expected to be higher in men than in women [34–36]. The average duration of infection was the only parameter with an informative prior, but we found marked differences between the duration of infection in HMW (6.6 months) and MSM (2.3 months). Per sex act transmission probabilities are generally considered to be lower from women to men than vice versa [37–39]. In our model, the median of the transmission probability per partnership was lower in MSM hosts than in HMW for both sexual activity groups. This could be explained by different numbers of sex acts per partnership in the two populations. The low transmission probability within the highly active MSM group (median: 30%) could reflect a single or a small number of sex acts per partnership. In contrast, the high transmission probability for HMW within the low sexual activity group (median: 87%) could be a result of a larger number of sex acts per partnership in those individuals. Furthermore, condom use is more frequent in MSM than in HMW [22], which could explain part of the observed differences in transmission probabilities. Our study found that the treatment rate is the driving force of resistance spread. Xiridou et al. [13] found that resistance could spread faster when the treatment rate was higher, but they did not identify the treatment as the major driver of resistance spread. Chan et al. [12] found that focusing treatment on the core group leads to a faster rebound to pre-treatment prevalence than equal treatment of the entire host population. Unfortunately, our findings cannot be compared with Chan et al. because they do not report the proportion of antibiotic-resistant N. gonorrhoeae. It was shown previously that treatment is the main selective force acting on resistance evolution due to the selective advantage to the resistant pathogen [40, 41]. We now expand this concept by showing that, assuming no fitness costs, treatment rates determine the rates of resistance spread even when the host populations has a heterogeneous contact structure. The intuitive argument that a faster spread of an infection, due to a higher number of sexual partners, will result in a faster spread of resistance does not hold. Instead, the proportion of resistant infections spreads equally in host populations with different number of partners as long as they receive treatment at the same rate and there are no fitness costs associated with the transmission probability per partnership. For N. gonorrhoeae, this insight challenges the current management strategy that aims to lower the overall burden of infection by expanding screening and treatment of hosts [9, 10]. As soon as antibiotic-resistant pathogens are frequent enough to evade stochastic extinction, expanded treatment will foster their spread and increase the burden of N. gonorrhoeae. Additionally, we show that fitness costs can decelerate or even prevent the spread of antibiotic-resistant N. gonorrhoeae strains. Fitness costs therefore might explain why highly resistant strains, such as the ceftriaxone-resistant N. gonorrhoeae strain H041, do not spread in the host population after their first detection [42]. Our findings also show that bridging between the HetM and the MSM host populations might not have been necessary for cefixime-resistance to spread in the HetM population after 2010 [5]. It is likely that cefixime-resistant N. gonorrhoeae had already been present in the HetM population but were spreading at a lower rate than in the MSM population. The results of our study will be useful for future N. gonorrhoeae research and for guiding treatment recommendations. The N. gonorrhoeae transmission model describes observed prevalence and incidence rates well and can reconstruct the spread of antibiotic-resistant N. gonorrhoeae. Estimating rates of resistance spread is useful for projecting future resistance levels and the expected time it will take until a certain threshold in the proportion of antibiotic-resistant N. gonorrhoeae is reached. Until now, treatment recommendations for N. gonorrhoeae are subject to change when 5% of N. gonorrhoeae isolates show resistance against a given antibiotic [6]. Our study shows the importance of the rate of spread: a level of 5% resistance results in a marginal increase to 8% in the following year if resistance spreads logistically at rate 0.53 y−1 (HetM mean estimate from Table 4), but reaches 18% in the next year if resistance spreads at rate 1.46 y−1 (MSM mean estimate from Table 4). Public health authorities could use surveillance data and adapt thresholds for treatment recommendation change to specific host populations using the method we describe. Our study challenges the currently prevailing notion that more screening and treatment will limit the spread of N. gonorrhoeae, as higher treatment rates will ultimately result in faster spread of antibiotic resistance. Future treatment recommendations for N. gonorrhoeae should carefully balance prevention of N. gonorrhoeae infection and avoidance of the spread of resistance.

10.1371/journal.pntd.0001519

Therapeutic DNA Vaccine Encoding Peptide P10 against Experimental Paracoccidioidomycosis

Paracoccidioidomycosis (PCM), caused by Paracoccidioides brasiliensis, is the most prevalent invasive fungal disease in South America. Systemic mycoses are the 10th most common cause of death among infectious diseases in Brazil and PCM is responsible for more than 50% of deaths due to fungal infections. PCM is typically treated with sulfonamides, amphotericin B or azoles, although complete eradication of the fungus may not occur and relapsing disease is frequently reported. A 15-mer peptide from the major diagnostic antigen gp43, named P10, can induce a strong T-CD4+ helper-1 immune response in mice. The TEPITOPE algorithm and experimental data have confirmed that most HLA-DR molecules can present P10, which suggests that P10 is a candidate antigen for a PCM vaccine. In the current work, the therapeutic efficacy of plasmid immunization with P10 and/or IL-12 inserts was tested in murine models of PCM. When given prior to or after infection with P. brasiliensis virulent Pb 18 isolate, plasmid-vaccination with P10 and/or IL-12 inserts successfully reduced the fungal burden in lungs of infected mice. In fact, intramuscular administration of a combination of plasmids expressing P10 and IL-12 given weekly for one month, followed by single injections every month for 3 months restored normal lung architecture and eradicated the fungus in mice that were infected one month prior to treatment. The data indicate that immunization with these plasmids is a powerful procedure for prevention and treatment of experimental PCM, with the perspective of being also effective in human patients.

Paracoccidioidomycosis (PCM) is the predominant systemic mycosis in Latin America causing half of the total deaths among systemic fungal infectious diseases in Brazil. Chemotherapy is the standard treatment, but the long time required, severe cases of immunosuppression and frequent relapses indicate that additional methods should be introduced such as immunotherapy combined with antifungal drugs. Previously, the protective activity of P10, a peptide derived from the major diagnostic antigen gp43, was demonstrated, alone or combined with chemotherapy. P10 elicited a vigorous IFN-γ mediated Th-1 immune response. Presently, the reduction of fungal load, and even sterilization, was attempted using a specific DNA vaccine encoding P10. Plasmid pcDNA3 expression vector with P10 insert was tested as a vaccine in intratracheally infected BALB/c and B10.A mice. Our results showed that vaccination with pP10 induced a significant reduction of the fungal burden in the lung. Co-vaccination of pP10 with a plasmid encoding mouse IL-12 proved to be even more effective in the elimination of the fungus with virtual sterilization in a long term infection and treatment assay system. The data suggest that immunization with these plasmids, without the need of an adjuvant, could be used in the prevention and treatment of PCM in human patients.

Paracoccidioides brasiliensis is a thermally dimorphic fungus that causes a systemic granulomatous disease known as paracoccidioidomycosis (PCM). PCM is widespread in Latin America, mainly affecting rural workers, and its incidence has increased in recently deforested areas associated with soil churning [1]. Acquisition of P. brasiliensis may arise from inhalation of aerosolized conidia. Recently we reviewed the death rates by systemic mycoses in Brazil [2]. PCM was the principal cause of death identified for 3,583 patients in the 1996–2006 decade and representing 51.2% of total deaths due to systemic mycoses. It ranks as the 10th most common cause of death among infectious diseases in Brazil [2]. There are distinct forms of PCM. The acute and sub-acute forms affect both genders with primary involvement of the reticuloendothelial/lymphatic system. The chronic form affects mainly adult males and predominantly causes pulmonary and/or mucocutaneous disease [3]. Antifungal chemotherapy is required for treatment, though treatment may not assure complete eradication of the fungus, with frequent relapses. Treatment with itraconazole usually takes 6–9 months in the low and 12–18 months in the moderately severe cases. Frequently, a combination of trimethoprim and sulfamethoxazole (TMP/SMZ) is used, held for 12 months in the low severity forms and 18–24 months in the moderately severe forms. Patients with severe PCM forms require endovenous treatment with anfothericin B or the TMP/SMZ association for long periods, monitored by clinical, radiological and serological tests [4]. The 43 kDa glycoprotein was characterized as the major diagnostic antigen of P. brasiliensis [5], [6], [7]. Immunization with gp43 elicited delayed hypersensitivity reactions in guinea pigs [8] and humans [9], implying the presence of T-CD4+ reacting epitopes. Based on the sequence of gp43 [6], which encodes a polypeptide of 416 amino acids with a single high mannose N-glycosylated chain [10], the T-cell epitope was mapped to a 15-mer peptide called P10 [11]. The hexapeptide HTLAIR comprises the essential core of P10 that induces proliferation of lymph node cells from mice sensitized to gp43 or infected with P. brasiliensis [11]. Type 1-T helper lymphocytes producing IL-2 and IFN-γ are induced by P10 [11], [12], [13]. Intratracheally infected mice previously immunized with P10 in the presence of complete Freund's adjuvant (CFA) had >200-fold reduction of lung P. brasiliensis colony-forming units (CFUs). In many cases the immunization rendered preserved lung architecture with few or no yeasts, whereas the infected, unimmunized mice displayed dense pulmonary inflammation characterized by epithelioid granulomas with numerous yeast cells [11], [12]. The immunoprotection by P10 depends on the IFN-γ-producing Th-1 response since mice deficient in IFN-γ, IFN-γ-R or IRF-1, but not IFN-α-R/IFN-β-R, were not protected by P10 immunization [12] The essential role of IFN-γ in organizing granulomas that contain P. brasiliensis yeasts has also been recognized by other investigators [13], [14], [15]. Several experimental avenues have been pursued to validate P10 as a vaccine candidate. These studies have included: a) the presentation of P10 by MHC molecules from different murine haplotypes [11]; b) its conservation in nature, confirmed by examining gp43 molecules from different isolates [16]; c) its immunogenicity and effective immunoprotection in formulations that do not require complete Freund's adjuvant [17]; d) its presentation by most human HLA-DR molecules as well as that of neighbor peptides to P10, based on the sequence of gp43 [18]; and e) the effectiveness of P10 as an adjuvant to chemotherapy in normal [19] and anergic [20] mice challenged intratracheally with virulent P. brasiliensis. The immunoprotective properties of P10 emulsified in Freund's adjuvant have been well documented in an established murine model of PCM [11]. Since CFA is not allowed in human vaccines and a tetramer of truncated P10 although immunogenic, involves laborious chemical methods [17], we have explored alternative approaches for P10 delivery. In the present work we have investigated the effectiveness of plasmid immunization with P10 and/or IL-12 inserts given prior to or after challenge with a virulent Pb18 isolate of P. brasiliensis using a murine pulmonary PCM disease model. Our results demonstrate that plasmid immunization with P10 with or without IL-12 inserts is highly therapeutic in mice intratracheally infected with this fungus. Most importantly, immunization was effective either prior to, or after infection suggesting that these plasmids are candidates for use in human PCM. Yeast cells of P. brasiliensis isolate 18 (Pb18) were grown in Sabouraud Dextrose Broth (BD, MD, USA) at 37°C for 7 days. Cells were washed and frozen in liquid nitrogen then disrupted by grinding on a mortar. Total RNA was isolated with Trizol according to manufacturer's instruction (Invitrogen, CA, USA). Complementary DNA was synthesized from 1 µg of total RNA in the presence of oligo(dT)18 (Fermentas, MD, USA) and Revertaid M-MuLV(Fermentas, MD, USA). The P10 nucleotide sequence was obtained using the sense PCR primer derived from the gp43 [6] 5′ nucleotide sequence: [5′-AAT AAG CTT CAA ACC CTG ATC GCC-3′], and the antisense primer derived from the 3′ end of the gp43 gene: [5′- AAT GAA TTC ATT GGC GTA ACG GAT TGC-3′]. A HINDIII site and an EcoRI site were added to the sense and antisense primers, respectively, for cloning into plasmid pcDNA3 (Invitrogen, CA, USA). PCR reactions (50 µl) were carried out following the protocol provided by Fermentas, using 100 ng of cDNA and 100 ng of each primer. The P10 PCR reaction started with one cycle at 94°C (2 min), followed by 40 cycles at 94°C (30 sec), 55°C (1 min) and 72°C (1 min), and a final 7-min extension at 72°C. PCR products were purified using Wisard SV gel and PCR Clean-UP system (Promega, Brazil) and each PCR product was digested with the appropriate restriction enzyme (Fermentas) and cloned into the pcDNA3 by directional insertion in the HINDIII/EcoRI sites. The resulting plasmid was called pP10. Plasmid pORF-mIL-12 was acquired from InvivoGen (CA, USA). The confirmation of the insert was done using the primers: sense [5′-CGG GTT TGC CGC CAG AAC ACA-3′] and antisense [5′-GGC CAC CAG CAT GCC CTT GT-3′]. The IL-12 PCR started with one cycle at 94°C (2 min), followed by 40 cycles at 94°C (1 min), 45°C (1 min) and 72°C (2 min), and a final 7-min extension at 72°C. To prepare plasmid DNA for immunization, Escherichia coli XL1Blue and DH5α cells were transformed by electroporation using Cellject Duo according to the manufacturer's directions (Hybaid, Middx, UK) with the DNA constructs or the vector plasmid alone and then cultured at 37°C in Luria broth supplemented with ampicillin (50 µg/ml). The positive clones were confirmed by automatic sequencing carried out following the protocol provided by Applied Biosystems (CA, USA) and analyzed by BioEdit and Blast. The parental vectors, pcDNA3 and pORF were used as negative controls. DNA for immunization was purified using the EndoFree Giga Kit (Qiagen, CA, USA) and was diluted in TE buffer to the final concentration of 1 µg/µl. For the expression of pORF-mIL-12 in mammalian cells, a transient-transfection assay was performed using Lipofectin (Invitrogen) and 1 or 2 µg plasmid transfected into HeLa cells (2×105 cells/well). The cells were grown in RPMI medium supplemented with 10% fetal calf serum (FCS) (Cultilab, SP, Brazil). After 24 h incubation, the cells were harvested, and total RNA was isolated with Trizol for reverse transcription (RT)-PCR. IL-12 PCR was used as described above. IL-12p70 was detected (80 ng/ml) by ELISA, in the supernatant of transfected HeLa cells. DNA immunization was performed by injecting groups of 5 six-week-old male BALB/c mice intramuscularly in both quadriceps with three doses of 100 µg of plasmid encoding P10 (pP10), 50 µg of either pP10 and pcDNA3 vector alone, or 50 µg of the pcDNA3 vector alone, each in 50 µl of TE buffer. A total of three immunizations were given at weekly intervals in alternating sites on the left and right hind legs. The mice were euthanized one week after the last immunization, their spleens were isolated and single-cell suspensions were prepared by gentle homogenization in RPMI medium supplemented with 1% FCS. Cells were suspended and treated with isotonic ammonium chloride to lyse erythrocytes. The splenocytes were washed by centrifugation, suspended in RPMI containing 10% FCS, and dispensed into wells on a microtitering plate (5×105 mononuclear cells per well). The cultures were stimulated with 20 µg/ml of synthetic P10. After 24 and 48-h incubation at 37°C with 5% CO2, supernatants were collected and IFN-γ was assayed by a sandwich enzyme-linked immunoassay (ELISA) (BD Pharmingen, CA, USA). Splenocytes from animals immunized with pP10 and stimulated with synthetic P10 produced 10 and 15 ng/ml IFN-γ after 24 and 48 h, respectively. When 50 µg of pcDNA3 was used for immunization, 9 and 11 ng/ml IFN-γ was released by splenocytes at the two examined times. This study was carried out as recommended by the Brazilian college of animal experimentation (COBEA). The protocol has been approved by the Ethical Committee on Animal Experimentation of University of São Paulo (Permit number: 039). BALB/c and B10.A mice were bred at the Institute of Biomedical Science of University of São Paulo, Department of Immunology animal facility under specific-pathogen-free conditions. Yeast cells of the virulent isolate Pb 18 of P. brasiliensis were maintained by weekly subculturing on Sabouraud Dextrose Agar and incubation at 37°C. Before experimental infection, 7–10 day-old cells were inoculated into Sabouraud Dextrose Broth and incubated at 37°C for 5–7days with rotary shaking. Fungal cells were washed three times in phosphate-buffered saline pH 7.2 (PBS) and counted in a haemocytometer. The viability of fungal cells in the inoculum was determined by staining with Janus B (Merck, Darmstadt, Germany) and was greater than 90%. BALB/c and B10.A mice (6- to 8-week-old males) were inoculated intratracheally (IT) with 50 µl suspension of 3×105 Pb18 yeast cells in sterile saline (0.85% NaCl). Mice were anesthetized i.p. with 200 µL of a solution containing 80 mg/kg ketamine and 10 mg/kg of xylazine (both from União Química Farmacêutica, Brazil). After approximately 10 min, their necks were extended to expose the trachea at the thyroid level and cell suspensions were injected with a 26-gauge needle. The incisions were sutured with 5-0 silk. Three different protocols were used (Fig. 1). Injections of 50 µg plasmid were given in the quadriceps muscle. First protocol: Groups of 10 BALB/c mice were injected with PBS (control), pcDNA3 (50 µg; control), pORF (50 ug; control), plasmid encoding P10 (pP10, 50 µg), plasmid encoding IL-12 (pIL-12, 50 µg) or with both pP10 and pIL-12 (50 ug each). A total of 4 injections were given weekly on alternating sites, on the left and right hind legs. One week after the last injection, mice were infected intratracheally then sacrificed 30 or 60 days later. Second protocol: BALB/c and B10.A mice (10 mice per group) were infected intratracheally. One month after infection, the mice received 4 weekly injections of either PBS (control), pcDNA3 (50 µg; control), pP10 (50 µg), pIL-12 (50 µg) or both pP10 and pPIL-12 (50 ug each). Mice were sacrificed 1 week after the last injection. Third protocol: B10.A mice (10 mice per group) were infected intratracheally. One month after infection, they were treated with 4 weekly injections followed by a monthly booster for 3 additional months (total of 7 injections). The injections were with either PBS (control), pcDNA3, pP10, pIL-12 or both pP10 and pPIL-12. The mice were sacrificed one month after the last injection, six months after infection. Mice were sacrificed and the lungs, liver and spleen were removed. Weighed tissue sections were homogenized and then washed 3 times with PBS and suspended in 1 ml PBS. Suspensions (100 µl) were inoculated on brain-heart infusion (BHI) agar medium supplemented with 4% FCS and 5% spent culture medium of P. brasiliensis (strain-192), streptomycin/penicillin 10 IU/ml (Cultilab) and cycloheximide 500 mg/ml (Sigma, MO, USA). Colonies were counted after 10 days of incubation at 37°C. Lung sections from sacrificed mice were fixed in 10% buffered formalin for 24 h and embedded in paraffin. Four-micra sections were stained with haematoxylin-eosin (HE) or silver nitrate (Gomori) and examined microscopically (Optiphot-2; Nikon, Tokyo, Japan). Sections of excised lungs were homogenized in 2 ml of PBS in the presence of protease inhibitors: benzamidine HCl (4 mM), EDTA disodium salt (1 mM), N-ethylmaleimide (1 mM) and Pepstatin (1.5 mM) (Sigma, St Louis, MO). The supernatants were assayed for IL-4, IL-10, IL-12, and IFN-γ using ELISA kits (BD OpTeia, San Diego, CA). The detection limits of the assays were as follows: 7.8 pg/ml for IL-4, 31.3 pg/ml for IFN-γ and IL-10, 62.5 pg/ml for IL-12, as previously determined by the manufacturer. Statistical analyses were performed using GraphPad Prism5 software. The results are expressed as means and standard deviations (SD). The nonparametric Kruskall-Wallis honestly significant difference test was used. p values are shown in the Figure legends. To explore the effects of the plasmid with the P10 insert (pP10) with or without the murine IL-12 gene insert (pIL-12), BALB/c mice were immunized and then infected intratracheally with virulent P. brasiliensis Pb 18. Animals were sacrificed after 30 or 60 days, and the fungal burden in the lungs, spleens and livers was determined. The number of lung CFU per gram of tissue was significantly reduced in animals immunized with pP10 and/or pIL-12 compared to controls at both time intervals (Fig. 2). Notably, we observed that the empty plasmids (pcDNA3 and pORF) also induced a significant reduction in CFU relative to mice that received PBS alone, which is presumably a result of dendritic cell activation through Toll-like receptor 9 binding of plasmid unmethylated CpG motifs. Immunostimulation by DNA from P. brasiliensis also attributed to CpG motifs showed protective effects in susceptible mice [21], [22]. Nevertheless, the fungal load measured in CFUs in mice receiving pP10 and/or pIL-12 was significantly lower than that in mice treated with control pcDNA3 and pORF. Livers and spleens from all animals had no detectable fungal cells. The therapeutic protocol attempts to reproduce the clinical reality of patients presenting to medical attention after developing symptomatic PCM. We studied two mouse strains with different susceptibilities to PCM, BALB/c (susceptible) and B10.A (highly susceptible) [23]. The data showed that immunization with pP10 and/or pIL-12 was therapeutic in mice infected with P. brasiliensis for 1 month prior to receiving plasmid immunizations (Fig. 3). CFU reductions were significant in infected mice receiving pP10 and/or pIL-12 compared to mice injected with PBS or pcDNA3. In contrast to the first protocol, injection of pcDNA3 after installing PCM was not sufficient to reduce the fungal burden. The most significant reduction in the lung CFUs from B10.A mice was achieved when pP10 and pIL-12 were combined. The CFUs from the livers and spleens were barely detectable in all groups. This protocol allowed us to analyze the efficacy of therapeutic plasmid treatment during long-term infection (six months) of the highly susceptible mouse strain, B10.A. Treatment of mice with PCM using pP10 and/or pIL-12 significantly reduced lung CFUs (Fig. 4). However, the impact of pIL-12 alone was not as dramatic as either pP10 alone or pP10 with pIL-12. Notably, treatment with the combination of pP10 and pIL-12 virtually eradicated the infection in all organs examined. The lungs of control animals in each experimental protocol group showed intense inflammation and large numbers of yeast cells, whereas mice receiving pP10 with or without pIL-12 had significantly reduced inflammation, and lower or undetectable fungal cells. Analysis of the lungs of animals from the third protocol (6 months infection) that received control plasmids revealed dense infiltration of inflammatory cells, mainly of macrophages, lymphocytes and epithelioid cells, and numerous fungal cells (Fig. 5A). Around the foci of epithelioid granulomas, giant cells were observed. In contrast, there were large areas of normal lung architecture in pP10-immunized mice and a global reduction in the number of granulomas with few yeast cells (Fig. 5B). Treatment with pIL-12 resulted in histological findings that were more similar to controls than to pP10-immunized mice (Fig. 5C). Importantly, the lungs of mice treated with the combination of pP10 and pIL-12 were mostly histologically normal and no yeast cells were identified (Fig. 5D). Previous studies with BALB/c mice have established that P10 elicits a protective Th-1 immune response [11]. BALB/c and B10.A mice have different genetic backgrounds that strongly influence their response to infection by P. brasiliensis. Their different susceptibility to fungal infection depends in part on their capacity to produce pro-inflammatory cytokines, which are often reduced in B10.A relative to BALB/c. IL-4, IL-10, IL-12 and IFN-γ were measured in the lungs of infected B10.A mice and BALB/c. In mice subjected to the second protocol, BALB/c mice responded to pcDNA3 and pP10 gene immunization with significant increase in IFN-γ in the lung homogenate compared to untreated or pcDNA3 treated animals (data not shown). In contrast, B10.A mice produced significantly less IFN-γ after immunization with pP10 and pcDNA3 in comparison to treatment with pcDNA3 gene alone, which suggests that the increase in IFN-γ in both of these groups relative to untreated mice could be due to dendritic cell activation by plasmid CpGs. The cytokine production in the group of animals submitted to the third protocol, in which B10.A mice treated with pP10 with or without pIL-12 had undetectable yeasts in the lung tissue, is shown in Table 1. After 6 months post-infection, cytokine analyses in these mice showed a persistent IFN-γ production regulated by an IL-10-rich immune response, which is compatible with a protective therapeutic effect in B10.A mice. A vaccine against P. brasiliensis using plasmid DNA was first tested in 2000 [24], [25]. BALB/c mice were immunized with a mammalian expression vector carrying the full gene of the gp43 under the control of CMV promoter with Freund's adjuvant resulting in the induction of both B and T cell-mediated immune responses characterized by a mixed Th-1/Th-2 long-lasting cellular immune response, chiefly modulated by IFN-γ. This immunization method was protective when performed in mice prior to challenge with virulent P. brasiliensis. When tested for immunoprotection, P10 in Freund's adjuvant was also active in the murine model of PCM, eliciting an IFN-γ-dependent Th-1 immune response [11], [20]. The combined treatment of P10-vaccine in Freund's adjuvant and chemotherapy, either an azole, amphotericin B, or sulfamethoxazole, stimulated a protective Th-1 response, rich in IL-12 and IFN-γ, that was therapeutically beneficial if initiated 2 or 30 days after intratracheal infection [19]. The combined treatment was also effective in anergic animals challenged with the fungus [20]. Presently we used a DNA vaccine encoding P10 with or without a plasmid encoding IL-12 that is administered without adjuvant. We found that the pP10 vaccine, either given prior to or 1 month after intratracheal infection, induced a significant reduction in the fungal burden in the lungs of mice. Co-vaccination with murine pIL-12 significantly enhanced vaccine effectiveness, particularly in a long- term infection model in B10.A mice. The combined DNA vaccine (Protocol 3) achieved virtual sterilization after 6 months with histologically normal lungs and undetectable fungal burden. Full protection was mediated by IFN-γ production and the pro-inflammatory effect of pP10 and pIL-12 was regulated by IL-10 in these susceptible mice. The mechanism of fungal killing by gene immunization is not solely mediated by cytokines since the empty plasmid pcDNA3 is a strong stimulator of the immune system. However, a significant protection is only achieved with pP10 or pP10+pIL-12 administration. P10 is not protective in IFN-γ-KO mice [12], indicating that this cytokine is essential for fungal killing through macrophage activation. T-CD4+ lymphocytes recognizing P10 and other cells induced by fungal infection are the main producers of IFN-γ. A role for a simultaneous induction of protective antibodies against fungal antigens [26] is also recognized. IL-12 administration has been previously studied in experimental PCM [27]. Our current results show that IL-12 protected mice against disseminated infection. In the long term infection protocol, pIL-12 alone was only partially effective in the protection of infected mice, but the cytokine facilitated the elimination of P. brasiliensis when combined with pP10. This is a very encouraging result and strongly suggests that a pP10-based vaccine associated with pIL-12 could be used as a powerful adjuvant to chemotherapy. Despite the effectiveness of chemotherapy, fatalities from invasive or systemic fungal diseases are not uncommon. Vaccines against fungal diseases are gaining increasing attention, owing to their capacity to effectively modulate the immune response (reviewed in [28]. The frequent occurrence of clinical relapses and sequellae, such as pulmonary fibrosis, following antifungal chemotherapy suggest that immunoprotective vaccines could also reduce the incidence of these complications [29]. In addition to our work with P10, there have been other notable attempts to develop vaccine strategies for the treatment of PCM. They included a cDNA encoding the antigenic protein rPb27 [30], the recombinant heat shock protein 60 emulsified in adjuvant [31], radioattenuated P. brasiliensis yeast cells [32] and Mycobacterium leprae DNAhsp65 plasmid in infected BALB/c mice [33]. Braga et al. [34] immunized BALB/c mice either with recombinant purified flagellins (FliC) genetically fused with P10 or with the synthetic P10 peptide mixed with purified FliC. A prevailing Th1-type immune response was obtained that reduced P. brasiliensis growth and lung damage in infected mice. From a practical standpoint, the broad use of antifungal vaccines is not realistic when considering the perspective of a large number of infected people relative to the number of individuals who develop the disease. Mycoses caused by dimorphic fungi, such as PCM, coccidioidomycosis, histoplasmosis and blastomycosis, have low incidence as a deep-seated disease. Certain fungal diseases such as cryptococcosis, aspergillosis and candidiasis typically occur in immunocompromised hosts. Hence, targeted prophylactic vaccination may be a more practical approach to control disease. In the case of PCM, immunization of those at highest risk, such as farmers in highly endemic regions would be reasonable. However, we have also demonstrated that the pP10/pIL-12 combination is highly efficacious after PCM has developed. Therefore, immunization could be most useful in combination with standard therapy in PCM patients in order to enhance treatment efficacy, reduce treatment duration and, perhaps, prevent relapses.

10.1371/journal.pgen.1007039

The creation and selection of mutations resistant to a gene drive over multiple generations in the malaria mosquito

Gene drives have enormous potential for the control of insect populations of medical and agricultural relevance. By preferentially biasing their own inheritance, gene drives can rapidly introduce genetic traits even if these confer a negative fitness effect on the population. We have recently developed gene drives based on CRISPR nuclease constructs that are designed to disrupt key genes essential for female fertility in the malaria mosquito. The construct copies itself and the associated genetic disruption from one homologous chromosome to another during gamete formation, a process called homing that ensures the majority of offspring inherit the drive. Such drives have the potential to cause long-lasting, sustainable population suppression, though they are also expected to impose a large selection pressure for resistance in the mosquito. One of these population suppression gene drives showed rapid invasion of a caged population over 4 generations, establishing proof of principle for this technology. In order to assess the potential for the emergence of resistance to the gene drive in this population we allowed it to run for 25 generations and monitored the frequency of the gene drive over time. Following the initial increase of the gene drive we observed a gradual decrease in its frequency that was accompanied by the spread of small, nuclease-induced mutations at the target gene that are resistant to further cleavage and restore its functionality. Such mutations showed rates of increase consistent with positive selection in the face of the gene drive. Our findings represent the first documented example of selection for resistance to a synthetic gene drive and lead to important design recommendations and considerations in order to mitigate for resistance in future gene drive applications.

Gene drives are selfish genetic elements that are able to bias their own inheritance among offspring. Starting from very low frequencies they can rapidly invade a population in just a few generations, even when imposing a fitness cost. Gene drives based on the precise DNA cutting enzyme CRISPR have been shown recently to be highly efficient at copying themselves from one chromosome to the other during the process of gamete formation in mosquitoes, resulting in transmission to 99% of offspring instead of the 50% expected for a single gene copy. One proposed use for CRISPR-based gene drives is in the control of mosquitoes by designing the gene drive to target mosquito genes involved in fertility, thereby reducing their overall reproductive output and leading to population suppression. Like any intervention designed to suppress a population these gene drives are expected to select for mutations in the mosquito that are resistant to the drive and restore fertility to mosquitoes. We have analyzed the origin and selection of resistant alleles in caged populations of mosquitoes initiated with a gene drive construct targeting a female fertility gene. We find the selected alleles are in-frame insertions and deletions that are resistant to cleavage and restore female fertility. Our findings allow us to improve predictions on gene drive behaviour and to make concrete recommendations on how to improve future gene drive designs by decreasing the likelihood that they generate resistance.

Naturally occurring gene drives—selfish genetic elements that are able to bias their own inheritance and rapidly invade a population, even starting from very low frequencies—have inspired proposals to harness their power to spread into a population of insect disease vectors traits that manipulate their biology in ways that could suppress or eliminate disease transmission [1–4]. In particular for malaria, transmitted exclusively by mosquitoes of the Anopheles genus, historical gains in reducing the disease burden have been largely achieved by the correct implementation of vector control measures (residual insecticides and bed nets) [5]. Though these measures have been instrumental in substantially reducing malaria transmission, they are insufficient by themselves to eradicate the disease in the near future at the current level of investment [6]. Gene drive technology could help in developing a self-sustaining, species-specific and affordable vector control measure much needed to achieve disease eradication in the future. Gene drives based on the activity of DNA nucleases able to recognise specific target sequences were first proposed over a decade ago and have received much attention recently due to the advent of new, easily programmable nucleases such as CRISPR-Cas9 that have allowed us and others to build functioning gene drives that show rates of inheritance from a heterozygous parent close to 100%, compared to the expected Mendelian inheritance of 50% [1, 7–9]. The principle behind the technology is to re-program a nuclease to cleave a specific site of interest in the genome and to insert the nuclease within this recognition site. The gene drive is designed to be active in the germline, so that in diploid organisms heterozygous for the gene drive the nuclease causes a double stranded break (DSB) at the target site on the homologous chromosome not containing the gene drive. The DSB can be repaired either by simple end-joining (EJ) of the broken strands or via homology-directed repair (HDR) where the DSB is resected and the intact chromosome used as a template to synthesise the intervening sequence. In the case of a gene drive, repair via HDR thus leads to a copying of the drive from one chromosome to another and the conversion of a heterozygote into a homozygote. Hence the force of gene drive is determined by a combination of the rate of cleavage of the nuclease in the germline, and the propensity for the cell machinery to repair the broken chromosome by HDR. We and others have shown that in germline cells the rates of HDR following a nuclease-induced DSB can be almost two orders of magnitude greater than EJ, a fact which explains the extraordinarily high rates of gene drive inheritance observed [7, 8, 10, 11]. On the other hand EJ repair can lead to the creation of small insertions or deletions at the target site that, although occurring initially at low frequency, might be expected to be selected for in the target organism if they prevent the gene drive nuclease acting and there is a negative fitness cost associated with the gene drive [1, 7, 11–13]. This possibility has been recognised since the first proposal of this type of gene drive [1], with much theory being dedicated to it recently [13, 14] and recent empirical evidence of its occurrence in Drosophila[10]. To lower the likelihood of resistance arising there are several potential mitigation strategies including, but not limited to, the targeting of conserved sequences that are less tolerant of mutations and the targeting of multiple sequences, akin to combination therapy [1, 12]. We previously developed a gene drive designed to spread into a mosquito population and at the same time reduce its reproductive potential by disrupting a gene essential for female fertility, thus imposing a strong fitness load on the population [11]. To investigate the long term dynamics of the emergence of resistance to a gene drive imposing such a load we continued to monitor the frequency of this gene drive over generations and analysed the target locus for evidence of mutagenic activity that could lead to the development of resistant alleles that block gene drive activity and restore gene functionality. Our findings show that a range of different resistant alleles can be generated and some of these are subsequently selected for and show dynamics consistent with our modelling predictions. These results provide a quantitative framework for understanding the dynamics of resistance in a multi-generational setting and allow us to make recommendations for the improvement of future gene drive constructs that relate to choice of target site and regulation of nuclease expression in order to retard the emergence of resistance. A proof-of-principle CRISPR-based gene drive designed for population suppression was previously developed in our laboratory (Fig 1A). This gene drive disrupted a haplosufficient gene (AGAP007280, the putative mosquito ortholog of nudel [15]) required in the soma and essential for female fertility [11]. The gene drive also contained an RFP marker gene for the visual detection of individuals inheriting the drive. In our experiments individuals heterozygous for the gene drive transmitted the drive, regardless of their sex, to more than 99% of their offspring. We observed in these mosquitoes a marked reduction in fertility (~90%) in females heterozygous for the drive, due to ectopic expression of the nuclease under control of the germline vasa2 promoter that resulted in conversion to the null phenotype in somatic cells. In spite of this fitness disadvantage experimental data showed that the gene drive could increase rapidly in frequency in a caged population due to the exceptionally high rates of inheritance bias. From a starting population (G0) in two duplicate cages of 600 individuals with a 1:1 ratio of transgenic heterozygotes and wild type individuals, the gene drive progressively increased in frequency to 72–77% by G4. This rate of increase was slightly higher than predicted by a deterministic model but within the limits of stochastic variation expected [11]. Due to a combination of the partial dominance of the sterility phenotype in heterozygous females and the previously documented generation of target site mutations conferring resistance to the gene drive [11], this first gene drive was not expected to maintain high levels of invasion. Nonetheless it represented a useful experimental model to investigate the long term dynamics of the de novo generation of target site mutations and their selection at the expense of a gene drive imposing a large reproductive load. We therefore maintained this cage experiment for 25 generations and used the presence of the RFP marker in the gene drive construct as a proxy to estimate the frequency of individuals containing it. The frequency of gene drive progressively increased in both cages, peaking at around generation 6, and thereafter we observed a gradual and continuing decrease such that by G25 the frequency of individuals with the gene drive was less than 20%. To investigate whether the gradual decline in the gene drive frequency observed in the cage experiment was due to the selection of pre-existing variant target sites in the population or the generation and selection of nuclease-resistant indels, we used deep sequencing of a PCR amplicon comprising sequences flanking the target site on pooled samples of mosquitoes from early (G2) and late (G12) generational time points (Fig 1B). The expected amplified region from the original wild type sequence was 320bp long, with the putative cleavage point within the target site residing after nucleotide 208 (Fig 2A). Ultra-deep sequencing of PCR reactions were performed on pooled DNA under non-saturating conditions so that the number of reads corresponding to a particular allele at the target site is proportional to its representation in the pool. We developed a computational method to analyse the sequences edited by the CRISPR-based gene drive close to the nuclease target site. In the colony of mosquitoes that we used there are a number of pre-existing single nucleotide polymorphisms (SNPs) within the amplicon that do not overlap the nuclease target site and are present at varying frequencies (S1 Fig). Our method identified small insertions and deletions introduced by the repair system and used the presence of surrounding SNPs to characterize the haplotypes on which they arose. Because the PCR only amplifies the non-drive allele, the frequencies reported below refer to their frequency within this class, rather than within the population as a whole. Mapping amplicon sequences reconstructed from the sequenced against the Anopheles gambiae reference genome (PEST strain, AgamP4, Vectorbase) we observed a large repertoire of deletions already in the G2 generation, with a wide range of sizes and centred around the predicted nuclease cleavage site after nucleotide 208 (Fig 2A—one cage trial shown as a representative example), and a lower proportion of small insertions, consistent with the known mutational activity of the nuclease. By contrast, ten generations later we observed a much reduced diversity of indels. We then considered all alleles that reached a frequency of at least 1% in any sample, classified these as to whether the indel caused a frameshift in the coding sequence of the target gene or was in-frame, and analysed their frequency over time (Fig 2B). The predominant target allele in the G2 was still the reference (non-mutated) allele at 63% and 48% in cages 1 and 2, respectively (Fig 2B and S1 Table), while the second major class (at least 15% in each replicate) was represented by a wide range of non-reference alleles, each present at low frequency (<1%), consistent with the stochastic generation of a broad range of indels. Thus at a time when the gene drive was still increasing in frequency there was a significant accumulation of mutations at the target site that would likely render it refractory to the homing mechanism of copying. Of note, three separate indels causing in-frame deletions of 3- or 6bp (202-TGAGGA, 203-GAGGAG, 203-GAG; where 203 refers to the starting site of the indel in the reference amplicon and “-”means deletion) were present among a large number of indels at low but appreciable frequencies in the G2. Such short in-frame deletions may result in only minimal disruption to the final encoded protein while at the same time proving resistant to the gene drive. Indeed these three deletions, plus a 6bp in frame insertion (207+AAAGTC), had increased significantly in frequency to make up the 4 most abundant non-drive alleles in the G12, almost to the exclusion of the reference allele (present at 6% and 0.4%; S1 Table). At the same time, a wide range of frameshift indels that were present in G2 had fallen in frequency in G12 to either below the 1% threshold or were not detected at all (Fig 2B and S1 Table). The most parsimonious explanation for these results is that a wide range of frameshift and in-frame indels was created by the gene drive, yet only short in-frame indels were selected for because they restore functionality to the target gene while protecting the sequence from gene drive activity. These ‘restorative’ mutations are likely to be most strongly selected when the frequency of the gene drive is high in the population—when the majority of individuals are homozygous for the driver, the relative gain in viable offspring from an individual with a gene drive balanced by a resistant restorative mutation is that much higher. Small in-frame deletions can arise by either classical non-homologous end-joining (NHEJ), or an alternative form of end-joining (microhomology-mediated end-joining, MMEJ) that relies on alignment of small regions of microhomology, as little as 2 base pairs, on either side of the DSB, resulting in loss of the intervening sequence [16]. Consistent with this latter possibility, at least three of the most frequent alleles in the G12 generation can be explained by MMEJ via 3bp repeats (Fig 2C). To investigate whether the most common indels at G12 had single or multiple origins, we used the naturally occurring SNPs in the sequences flanking the recognition sequence. In cage 1 the deletion 203-GAGGAG was present on 10 separate haplotypes in G12, with each haplotype being present at ratios broadly similar to their ratios in the starting population (S1 Fig), suggesting that the same deletion was generated at least 10 times independently and that there was no detectable selective advantage to any particular haplotype surrounding the deletion. In cage 2 the predominant allele was 202-TGAGGA (68% of all non-reference alleles), an end-joining deletion that shows no apparent features of a MMEJ event, and was found on 5 separate haplotypes. The progressive increase in frequency of specific mutations at the target site, concomitant with a decrease of gene drive activity, strongly suggested that they conferred resistance to cleavage while still ensuring a normal functional activity of nudel. To confirm this hypothesis we crossed individual RFP+ females from G20 with wild type males and assessed both their fertility and the transmission rate of the drive. We also sequenced the target site of each parental female to characterize allelic variants at the target locus. This analysis failed to detect wild type sequence at the target site (Fig 3A) among 70 individuals tested; instead every individual showed an indel, indicating that each female tested was heterozygous for the gene drive and balanced by a mutated target site. In cage 1 the 203-GAGGAG 6bp deletion was the predominant allele (23/31 individuals) while in cage 2 another 6bp deletion (202-TGAGGA) was predominant (37/39 individuals). The relative frequency of each allele was consistent with the results obtained using pooled amplicon sequencing performed on the G12 individuals. Of those heterozygous females that could be confirmed as having mated, the vast majority (56/58) generated viable progeny (average clutch size 119 +/- 35.9 eggs, average hatching rate 78.5% +/- 19.9%; Fig 3B and S2 Table) at rates significantly higher than those previously observed in females heterozygous for the gene drive and a wild type allele (90.7% overall reduction in fecundity) [11], suggesting that the mutations detected at the target site substantially restored the functionality of nudel. There was no obvious difference in fertility associated with the 5 indels sampled in this assay (Fig 3B). The two major indels across the two cages result in relatively conservative changes to the overall amino acid sequence of the final gene product—two glutamate residues missing (203-GAGGAG) or two glutamates missing and a conversion of a serine to arginine (202-TGAGGA). Finally, the offspring of these individuals showed a non-biased inheritance of the gene drive (50.82% mean +/-5.96% standard error; total RFP+ offspring 50.85%), consistent with normal Mendelian segregation (Fig 3B). Thus, as well as restoring nudel function, the mutated sequences were also resistant to the gene drive. Conceivably a breakdown in the nuclease component (e.g. mutation in the Cas9 coding sequence, or the gRNA sequence) could be an additional explanation for the Mendelian transmission of the gene drive element and restored fertility in heterozygous females that we observed. To assess this possibility we took the male offspring (‘sons’) of the above crosses that inherited the construct and crossed them in turn to wild type females. We assumed that if the gene drive construct was still functional it should show a biased inheritance when the resistant target site allele had been replaced with a wild type one. Indeed, in these sons we saw a significant increase in the transmission of the gene drive to their progeny, but the observed rate (mean 60.13% +/- 13.9% S.E.; total RFP+ progeny 59.6%) was much lower than that previously observed (~99% inheritance) [11]. A similar phenomenon of reduced homing has been observed in the offspring of another mosquito species [8], and more recently in Drosophila [10], when the drive construct was inherited from the mother and when the same vasa germline promoter was used to transcribe the Cas9 nuclease. The reduced gene drive activity in the immediate offspring of heterozygous mothers was attributed to the persistence of maternally-deposited Cas9 in fertilized embryos, leading to double stranded DNA breaks being repaired preferentially by end-joining mechanisms in the early zygote possibly before paternal and maternal homologous chromosomes are aligned. Consistent with this explanation, males in the subsequent generation (‘grandsons’) that had received only a paternal copy of the gene drive had exceptionally high homing rates, with 97.5% of progeny inheriting the gene drive (Fig 3C). The drop in homing seen in sons receiving a maternal dose of Cas9 (59.6% transmission vs. 97.5% in grandsons) allows us to estimate an ‘embryonic end-joining’ rate of 79.6% of wild type alleles being converted to cleavage-resistant alleles. This rate of embryonic end-joining is much higher than that observed in the germline at or just prior to meiosis (~1% [11]) and is predicted to reduce the rate of spread of the gene drive, due to a reduced frequency of cleavable alleles [17], and increase the rate at which restorative resistant alleles can arise and be selected. Cleavage due to maternally deposited Cas9 could potentially be followed by HDR instead of end-joining, effectively leading to ‘embryonic homing’, where the cleaved allele is converted to the non-cleaved allele. In the case of a resistant allele this could lead to an individual heterozygous for the allele being converted into a homozygote in the early zygote, thereby further accelerating the spread of the resistant allele in the population. One signature of embryonic homing of a resistant allele would be novel hybrid haplotypes due to partial conversion of the haplotype surrounding the wild type allele where the DSB was generated to the haplotype surrounding the resistant allele. Looking in detail at the most abundant resistant allele in each cage, we failed to observe such a signature, and all resistant haplotypes were already pre-existing in the population (S1 Fig), suggesting that if this phenomenon is occurring then the resection following cleavage and resultant conversion encompasses a section longer than the ~300bp covered in our sequenced amplicon. The key qualitative results from the cage experiments—that a gene drive can increase in frequency in a susceptible population even if it reduces individual fitness, and that the spread of a gene drive can in turn lead to the spread of mutants that are resistant to cleavage and restore individual fitness—are fully consistent with expectations from population genetic models [13, 18, 19]. To investigate how well such models can account for the quantitative details of the cage experiments, we extended the model of Deredec et al. [18] to incorporate our observations of embryonic cleavage by maternally derived nuclease (which is independent of inheritance of the gene drive), and have two classes of resistant allele (in-frame functional and frame-shift non-functional; see S1 Text and S1 File). Due to the sex-specific fitness effects of our construct, the model also has a separate treatment of females and males. Using this model and the baseline parameter values from the single-generation crosses, we generated the expected allele frequency dynamics over the 25 generations of the experiment (Fig 4). Again, the qualitative fit to the observed dynamics is good, but there are quantitative differences. For example, the model predicts that at G12 the original wild type allele will be 9.3% of all non-drive alleles, while our observed rates were 6% and 0.4% in cages 1 and 2, respectively (Fig 2B, S1 Table). The model also recapitulates the observation that while non-functional resistant alleles initially outnumber the functional ones, because they are produced more frequently, by the end of the experiment it is the functional ones that predominate. Importantly, for the gene drive itself, the model captures the essential aspects of the observed dynamics, showing an initial increase in frequency followed by an eventual loss, though in earlier generations our observed frequencies exceeded the predicted frequencies (Fig 4B). To investigate what might explain this discrepancy, we examined the effect of varying each of the different parameter values individually in the model, and found that small variations in the fertility of females heterozygous for the gene drive had the largest effect in increasing the match between observations and expectations. Keeping the experimental estimates of all other parameters unchanged, the least squares best fit occurred at a dominance coefficient of 0.70 (Fig 4B and S1 File), compared to our previous direct estimate of 0.9, with lower confidence limit of 0.86 [11]. We also used our model to investigate the potential impact of HDR after embryonic cleavage caused by maternal deposition of Cas9, and found this parameter has little effect on the expected rate of resistance emergence when the rates of meiotic homing are as high as we observe. We have analyzed the dynamics of a gene drive deliberately designed to impose a fitness load on a population, and characterized the resistant or compensatory mutations which it generated and selected for. As with any control approach aimed at suppressing an organism, ‘push back’ from the target organism is to be expected. One of the advantages of the modular gene drives investigated here is that contingency in planning for and overcoming resistance can be foreseen and built into the system in a number of ways. First, the use of multiple gene drives targeting separate sequences has long been considered an essential pre-requisite for any gene drive intended as a functional vector control tool [19] and the ease with which the guide RNA expression constructs can be multiplexed lends CRISPR-based gene drives this flexibility [12, 20]. Second, it will be useful to target sites at which sequence changes are likely to destroy function. The nuclease target site in the gene AGAP007280 described in this report was not chosen according to any prioritisation based on high levels of sequence conservation that would imply functional constraint—a feature expected to mean that resistant mutations are less likely to restore function of the gene. Clearly the choice of the target should be guided by the extensive genomic data that is now available on sequence conservation both across different Anopheles species [21] and within An. gambiae [22]. This data revealed a posteriori that for the target site in AGAP007280 used here there is pre-existing variation in this species at at least 8 of the 20 nucleotides covered by the gRNA. Going forward low tolerance of sequence variation at the target site should be a key criterion for designing a gene drive. Third, our results show that one of the key drivers in the generation of resistant alleles is the nuclease activity itself, followed by end-joining, and a significant proportion of these alleles are created as a result of maternally deposited nuclease in the early zygote where end-joining repair predominates over homology-directed repair. We suggest this maternal effect may be suppressed either through the use of more tightly regulated promoters to restrict nuclease expression to the early germline or through the addition of destabilising modifications to the nuclease, either of which are expected to reduce perdurance in the embryo. Fourth, an additional consideration in the choice of target site may take into account the propensity for a particular double strand break to be repaired more readily into a resistant, restorative (R2) allele, for example due to microhomology either side of the cleavage site that more readily re-creates an in-frame allele than a frameshift allele. Where MMEJ is the predominant end-joining repair pathway, this feature could be incorporated into target site choice to ensure the most likely end-joining repair event is an out of frame allele and therefore not likely to be selected. Our approach of pooled sequencing of a targeted region allowed us to reliably detect even low frequency signatures of gene drive activity and reveal the complex dynamics of different genotypes emerging over time. Certainly for the future improvement of gene drives it will be important to have a faster method to triage for the most robust gene drives least prone to resistance without a multi-generational cage experiment, a laborious and time consuming process that should be reserved for more extensive evaluation of the best candidates. A simple way to do this would be to apply the method of amplicon sequencing described here in a screen where all generated mutant alleles are balanced against a known null allele to see if they restore function to the target gene. The potential for rapid emergence and spread of resistance highlights not only one of the technical challenges associated with developing a gene drive, but also how intentionally releasing a resistant allele in to a population could be a simple and effective means of reversing the effects of a gene drive, if it fully restores function [23]. These experiments were essentially as described before in Hammond et al. [11] Briefly, in the starting generation (G0) L1 mosquito larvae heterozygous for the CRISPRh allele at AGAP007280 were mixed within 12 hours of eclosion with an equal number of age-matched wild-type larvae in rearing trays at a density of 200 per tray (in approx. 1L rearing water). The mixed population was used to seed two starting cages with 600 adult mosquitoes each. For subsequent generations, each cage was fed after 5 days of mating, and an egg bowl placed in the cage 48h post bloodmeal to allow overnight oviposition. After allowing full eclosion a random sample of offspring were scored under fluorescence microscopy for the presence or absence of the RFP-linked CRISPRh allele, then reared together in the same trays and 600 were used to populate the next generation. After a generation had been allowed the opportunity to oviposit, a minimum of 240 adults were removed and stored frozen for subsequent DNA analysis. For the sequence analysis, a minimum of 240 adult mosquitoes taken at generations G2 and G12 of the cage trial experiments were pooled and extracted en masse using the Wizard Genomic DNA purification kit (Promega). A 332 bp locus containing the target site was amplified from 40 ng of genomic material from each pooled sample using the KAPA HiFi HotStart Ready Mix PCR kit (Kapa Biosystems), in 50 μl reactions. Specially designed primers that carried the Illumina Nextera Transposase Adapters (underlined), 7280-Illumina-F (TCGTCGGCAGCGTCAGATGTGTATAAGAGACAGGGAGAAGGTAAATGCGCCAC) and 7280-Illumina-R (GTCTCGTGGGCTCGGAGATGTGTATAAGAGACAGGCGCTTCTACACTCGCTTCT) were used to tag the amplicon for subsequent library preparation and sequencing. The annealing temperature and time were adjusted to 68°C for 20 seconds to minimize off-target amplification. In order to maintain the proportion of the reads corresponding to particular alleles at the target site, the PCR reactions were performed under non-saturating conditions and thus they were allowed to run for 20 cycles before 25 μl were removed and stored at -20°C. The remnant 25 μl were run for an additional 20 cycles and used to verify the amplification on an agarose gel. The non-saturated samples were used to prepare libraries according to the Illumina 16S Metagenomic Sequencing Library Preparation protocol (Part # 15044223 Rev.A). Amplicons were then purified with AMPure XP beads (Beckman Coulter) followed by a second PCR amplification step with dual indices and Illumina sequencing adapters using the Nextera XT Index Kit. After PCR clean-up via AMPure XP beads and validation performed with Agilent Bioanalyzer 2100, the normalized libraries were pooled and loaded at a concentration of 10 pM on Illumina Nano flowcell v2 and sequenced using the Illumina MiSeq instrument with a 2x250bp paired-end run. Sequencing data of the amplified genomic region were analysed using available tools and developed scripts in R v3.3.1. Raw reads were cleaned up for low quality and trimmed for the presence of adapters using Trimmomatic v.0.36 [24]. Paired-end reads were merged together in order to reconstruct the whole amplicon sequence using PEAR v0.9.10 [25]. Resulting assembled identical fragments were then clustered using fastx_collapser module from the FASTX v0.0.13 suite (http://hannonlab.cshl.edu/fastx_toolkit/) and aligned to the reference amplicon with vsearch tool v2.0.3 [26]which implements a global alignment based on the full dynamic programming Needleman-Wunsch algorithm. We considered for downstream analysis only sequences represented by at least 100 reads in each dataset. The blast6 output files from the alignment phase were parsed by ad hoc written R scripts to identify sequence variants containing insertions and/or deletions in the target site. The quantification of each allelic variant was measured as relative alternative allele frequency by summing up the reads representing that particular variant in the dataset. Finally, for each identified variant, we examined the single nucleotide variants (SNVs) along the full amplicon and selected the ones with a minimum alternative allele frequency of 2.5% for the purposes of haplotype calling. As part of a sequencing effort one year prior to the start of this experiment 12 males and 12 females from our A. gambiae G3 laboratory colony were subjected to individual genome resequencing. Mosquitoes were chosen randomly as pupae of differing ages from separate trays of a large cohort of the colony population (census population size >2000) in order to minimise biased sampling from a reduced number of founders. After emerging as adults whole mosquitoes were individually homogenised and genomic DNA was extracted with Promega Wizard Genomic DNA extraction kit. Paired-end reads (2 x 100bp) obtained from the Illumina HiSeq 2000 sequencing were aligned to A. gambiae PEST reference genome assembly (AgamP4, VectorBase) using BWA-MEM (Li and Durbin 2009, v0.7.15). Alignments were sorted using Samtools (v1.5) and raw SNPs and indels were called using HaplotypeCaller tool from Genome Analysis Toolkit (GATK, v3.7) for each of the 24 samples in the same 320bp region around the nuclease target site in AGAP007280 gene that was used for the pooled amplicon sequencing. Raw SNPs were then merged using GATK GenotypeGVCFs tool. No indels were observed in this step for the selected region. We used SHAPEIT2 (Delaneau et al. 2013, v2) for the final haplotype estimation from previously obtained unphased genotypes of 24 individuals. Raw sequencing reads from the 24 individuals have been submited to NCBI Sequence Read Archive (SRA) under project accession number PRJNA397539. We use a discrete generation deterministic model to explore how the dynamics of gene-frequencies depend on underlying parameters. We suppose there are four possible alleles: Wildtype (W), driver allele (H), and two mutant alleles that are resistant to homing, R1 which is fully functional and R2 which is recessive but non-functional (i.e., H/R2 and R2/R2 type females are sterile). We assume alleles segregate at meiosis according to Mendelian inheritance except in W/H males and females, where segregation may be distorted by cleavage followed by either homing or non-homologous repair. Our model also allows for the possibility that eggs from females with at least one H allele will contain the driver nuclease (regardless of the egg’s own genotype), in which case cleavage and repair may occur in the embryo. The mathematical details of the model are given in the S1 Text, and model outputs from user-defined parameters values can be seen using a computable document format (Wolfram CDF Player) available as a file(S1 File). Baseline parameter values for the model are provided in the legend accompanying Fig 4. Individual females containing at least one copy of the RFP-linked CRISPRh gene drive were selected as virgins from the G20 generation and allowed to mate with 5 wild type male mosquitoes, essentially as in Hammond et al. [11]. The fecundity of females and transmission of the gene drive was measured by counting larval offspring positive for the RFP marker. To check mating status of females, spermathecae were dissected and examined for the presence of sperm. Unmated females were censored from the fertility assay. Sons of each gene drive mother from the G20 generation were kept together and allowed to mate in groups of approximately 5 males with an equal number of wild type females and assessed for rates of transmission of the gene drive. The male offspring of these sons (grandsons) inheriting the drive from their fathers were in turn assessed in the same way, keeping lineages separate.

10.1371/journal.pcbi.1002489

Predictive Features of Persistent Activity Emergence in Regular Spiking and Intrinsic Bursting Model Neurons

Proper functioning of working memory involves the expression of stimulus-selective persistent activity in pyramidal neurons of the prefrontal cortex (PFC), which refers to neural activity that persists for seconds beyond the end of the stimulus. The mechanisms which PFC pyramidal neurons use to discriminate between preferred vs. neutral inputs at the cellular level are largely unknown. Moreover, the presence of pyramidal cell subtypes with different firing patterns, such as regular spiking and intrinsic bursting, raises the question as to what their distinct role might be in persistent firing in the PFC. Here, we use a compartmental modeling approach to search for discriminatory features in the properties of incoming stimuli to a PFC pyramidal neuron and/or its response that signal which of these stimuli will result in persistent activity emergence. Furthermore, we use our modeling approach to study cell-type specific differences in persistent activity properties, via implementing a regular spiking (RS) and an intrinsic bursting (IB) model neuron. We identify synaptic location within the basal dendrites as a feature of stimulus selectivity. Specifically, persistent activity-inducing stimuli consist of activated synapses that are located more distally from the soma compared to non-inducing stimuli, in both model cells. In addition, the action potential (AP) latency and the first few inter-spike-intervals of the neuronal response can be used to reliably detect inducing vs. non-inducing inputs, suggesting a potential mechanism by which downstream neurons can rapidly decode the upcoming emergence of persistent activity. While the two model neurons did not differ in the coding features of persistent activity emergence, the properties of persistent activity, such as the firing pattern and the duration of temporally-restricted persistent activity were distinct. Collectively, our results pinpoint to specific features of the neuronal response to a given stimulus that code for its ability to induce persistent activity and predict differential roles of RS and IB neurons in persistent activity expression.

Memory, referred to as the ability to retain, store and recall information, represents one of the most fundamental cognitive functions in daily life. A significant feature of memory processes is selectivity to particular events or items that are important to our survival and relevant to specific situations. For long-term memory, the selectivity to a specific stimulus is seen both at the behavioral as well as the cellular level. For working memory, a type of short-term memory involved in decision making and attention processes, stimulus selectivity has been observed in vivo using spatial working memory tasks. In addition, persistent activity, which is the cellular correlate of working memory, is also selective to specific stimuli for each neuron, suggesting that each neuron has a ‘memory field’. Our study proposes that both the location of incoming inputs onto the neuronal dendritic tree and specific temporal features of the neuronal response can be used to predict the emergence of persistent activity in two neuron models with different firing patterns, revealing possible mechanisms for generating and propagating stimulus-selectivity in working memory processes. The study also reveals that neurons with different firing patterns may have different roles in persistent activity expression.

Working memory reflects the temporary storage of information that is necessary for immediate decisions/actions. Delay-period activity, which corresponds to neural activity that persists after the end of the initiating stimulus, represents the cellular correlate of working memory [1], [2]. This activity, referred from now on as persistent activity, is stimulus-selective: a specific pyramidal neuron will only exhibit persistent activity if a stimulus is presented in specific locations of the visual field, in the spatial working memory tasks for example, which represents the neuron's memory field. [3], [4]. ‘A large body of work has been devoted to understanding the biophysical mechanisms underlying induction and maintenance of persistent activity, which have emphasized the importance of a delicate balance between excitatory and inhibitory recurrent network connections [5], [6], [7], [8], [9], as well as the contribution of intrinsic cellular conductances [10], [11], [12], [13]. However, very little is known regarding the cellular mechanisms that enable stimulus selectivity in the PFC. How does a neuron ‘recognize’ the relevant stimulus and therefore, enters a persistent activity state? Previous studies have suggested that formation of these memory fields entails proper inhibitory transmission [14], as well as fine interactions between pyramidal neurons and interneurons [15], similar to the mechanisms underlying the formation of orientation columns [16]. However, additional cell-specific features, such as the latency to the first action potential or the sequence of inter-spike intervals (ISIs), could also be involved in the formation of memory fields, as shown in the visual cortex [17]. In the prefrontal cortex (PFC), the brain area heavily involved in mediating working memory functions and expression of persistent activity, layer V pyramidal neurons come in at least two flavors with respect to their firing patterns: intrinsic bursting (IB) neurons, characterized by an initial burst of action potentials (APs) followed by single APs or regular spiking (RS) neurons, characterized by a sequence of single APs [18], [19]. These neurons can also be categorized as adapting, whose firing frequency in response to a constant current step decreases during the stimulation and non-adapting [20], [21]. These different pyramidal neuron subtypes based either on their morphology or firing pattern, can form distinct sub-networks [22], [23], [24] that project to different subcortical areas, such as the pons or the striatum, suggesting that they might serve distinctive functional roles. This is further supported by recent data showing that cortico-pontine pyramidal neurons, compared to cortico-cortical neurons, have increased levels of the hyperpolarization activated cation current (H-current), contributing to increased temporal summation and increased amplitude of the slow afterdepolarization (dADP), which in turn facilitates the probability of persistent activity induction [25]. The present study uses detailed compartmental models of IB and RS neuron sub-types to identify (a) features of incoming signals that determine persistent activity induction (stimulus selectivity) and (b) characteristics of the neuronal response to these signals that may be used by downstream neurons to decode information about the probability of persistent activity emergence (encoding of preferred stimuli). Our results predict that stimulus-selectivity is tightly linked to the spatial location of activated synapses. Moreover, while the properties of persistent activity differ between the two subtype models, in both neurons the latency to the first action potential and the initial inter-spike-intervals of the stimulus-induced response contain predictive information regarding the emergence of persistent activity, providing a mechanism for encoding and propagating the occurrence of preferred signals. Following the construction of a morphologically and biophysically-detailed layer V PFC pyramidal neuron model, biophysically relevant variations in the sodium and R-type calcium currents (see Methods) led to the emergence of two distinct neuronal sub-types: a Regular Spiking (RS) and an Intrinsic Bursting (IB) model neuron. Specifically, a combination of doubling the R-type calcium and the persistent sodium conductances changed the firing pattern of the model neuron from an RS to an IB one. The experimentally documented range of these conductances [26], [27] indicates that such differences are often seen in layer V PFC pyramidal neurons. Neuronal responses were first validated extensively against known experimental data in order to verify that both model neurons exhibit: a) physiological values of input resistance (81 MΩ for both model neurons, experimental average 79.60±6.6 MΩ [28]), b) physiological responses to step pulse current injections (Fig. 1B, C), c) proper back-propagating action potentials (BPAPs) in the apical as well as the basal dendrites (Supplemental Fig. S1) and d) physiological synaptic responses in the basal dendrites (Supplemental Fig. S2, also see Methods). Two biophysical mechanisms have thus far been implicated in the generation of persistent activity: the NMDA [7], [9], [29], [30] and the CAN conductance [7], [10], [11], [29]. The NMDA current was validated with respect to the AMPA current based on experimental data from connected layer V PFC pyramidal neurons showing that the NMDA-to-AMPA ratio is 1.2, and that NMDA currents have relatively slow kinetics of inactivation (Fig. 2A, B1, B2) [31]. Furthermore, it has been shown that basal dendrites of layer V pyramidal neurons exhibit NMDA spikes at their basal dendrites (Fig. 2C) [32], [33]. These NMDA spikes are generated in an all-or-none manner, and once generated, stronger stimuli affect mostly the duration of the NMDA spike, while a slight increase in the amplitude may also be seen [32]. We tested whether NMDA spikes could be evoked at the basal dendrites of the neuron models. A dendritic branch located about 100 µm from the soma was stimulated with an increasing number of excitatory synapses. When using at least 40 excitatory synapses to induce the necessary depolarization, dendritic NMDA spikes could be evoked in both model neurons. Further increase in the number of synapses resulted in an increase of the NMDA spike duration along with a slight increase in the spike amplitude, in accordance with the experimental data (Fig. 2D1, D2). Layer V PFC pyramidal neurons have been shown to exhibit a delayed afterdepolarization (dADP) following stimulation of Gq-coupled receptors [34], [35]. This dADP is induced following a burst of action potentials and has small amplitude (average ∼3 mV) and very slow kinetics (decay τ = 3 sec) [35], rendering it a possible mechanism for induction and maintenance of persistent activity [11]. The dADP has been shown to be primarily generated by the CAN current [36] and possibly results from the activation of TRPC4/5 channels which are found in layer V PFC pyramidal neurons [35]. We simulated the dADP by including an additional ionic mechanism, which is mainly dependent on two variables: a) the half point of calcium-induced activation and b) the rate of inactivation. These two variables were adjusted so that the dADP was induced following a burst of 5 spikes, but was much smaller following just 2 spikes (Fig. 3A, B), in accordance with experimental findings [11]. The amplitude of the dADP could be modified by changing the conductance of the CAN mechanism (Fig. 3C) and was kept within the physiological range (2–8 mV), based on existing experimental data ([11] and Fig. 3D). Cortico-cortical connections that are thought to underlie the emergence and maintenance of persistent activity [7], [37] form synapses onto the basal dendrites of pyramidal neurons [38]. Therefore, the basal dendrites of the model neurons were stimulated with a total of 200 excitatory synapses (containing both AMPA and NMDA receptors), evenly distributed within a few basal dendrites (see Methods), 10 times at 20 Hz (synchronously), while the soma was stimulated with 5 inhibitory synapses (both GABAA and GABAB) at 50 Hz (also synchronously) [39]. This stimulation protocol was repeated 50 times and the location (set of dendritic branches), but not the stimulation time, of activated synapses varied between trials (see Methods). Persistent activity was induced in a probabilistic manner, in a percentage of these trials. Synaptic stimulation alone (in the absence of the CAN mechanism) did not lead to persistent activity in any single cell model, even when the number of stimulated synapses was gradually increased up to 400. However, since neuromodulators, such as dopamine and serotonin, are known to increase NMDA currents in layer V PFC pyramidal neurons [40], [41], [42] and the NMDA conductance is required in large-scale networks for stabilizing persistent activity [7], we next tested whether increasing the NMDA current by 25% could induce persistent activity in the single neuron models. Increasing the NMDA-to-AMPA ratio (abbreviated “N*”, with * equal to the ratio) from 1.2 to 1.5 did not induce persistent activity in any model neuron (data not shown) although it resulted in decreased inter-spike-intervals (ISIs) of the neuronal response during the stimulus (Supplemental Table S1). The latter concurs with experimental data showing modulation of neuronal excitability by NMDA in vitro [43]. Activation of the dADP mechanism on the other hand, resulted in induction of persistent activity, that is, neuronal activity that lasted more than 3 seconds following the end of the stimulus (Fig. 4A2, B2) in both neuron models. Increasing the magnitude of the CAN conductance (i.e., increasing the amplitude of the resulting dADP, tested with five somatic step pulses, within the physiological range) increased the probability of inducing persistent activity. We characterized the magnitude of the CAN current that would induce persistent activity with at least 50% probability in the 50 experimental trials in which the spatial arrangement of the synapses on basal dendrites was varied (i.e., at least 25/50 trials exhibited persistent activity). The dADP required for induction of at least 50% persistent activity for the RS and IB neuron models was 3.2 and 3.9 mV, respectively (Fig. 4C, white bars) and dropped by 1.3 mV in both models when the NMDA-to-AMPA ratio increased to 1.5 (Fig. 4C, black bars). Note, however, that the slightly larger dADP in the IB model cell corresponds to a smaller CAN conductance compared to the RS model cell (Fig. 3C). This can be explained by the enhanced R-type calcium and persistent sodium currents in the IB model cell which may contribute to the long-lasting depolarization produced by the CAN mechanism, thus partially substituting the CAN conductance. Taken together, these findings show that induction of persistent activity requires a larger dADP (although a smaller CAN conductance) in the IB than the RS model cell. In the previous analysis, we classified persistent activity as the neuronal activity that continues past the end of the initiating stimulus and lasts at least 3 seconds. However, the neuron models could exhibit self-terminated persistent activity (500–2000 ms) (Fig. 4D, white bars), even in the experimental trials classified as not having persistent activity (‘no persistent’ trials). We notice that the RS neuron model exhibits significantly less temporally-restricted persistent activity compared to the IB neuron model (Fig. 4D, white bars, p<0.001). Increasing the NMDA-to-AMPA ratio however facilitates short-lasting persistent activity to a much larger extent in the RS than the IB model (Fig. 4D, black bars, p<0.001). A possible explanation could lie in the fact that the IB model cell has larger R-type calcium and persistent sodium currents, which together with the CAN mechanism contribute to the prolonged depolarization needed for persistent activity. Thus, an additional slow increase in Ca++ influx due to enhanced NMDARs would have a greater impact on the RS model, where the primary conductance responsible for the dADP is the CAN conductance, than the IB model, where several mechanisms -with different kinetics- already contribute to this depolarization. Furthermore, this short-lasting persistent activity could be significant in an in vivo situation where network mechanisms could maintain it for longer periods of time. These findings show that, while persistent activity in the single neuron models is primarily dependent on the CAN current, altering the NMDA-to-AMPA ratio modulates the duration of persistent activity that lasts less than 3 seconds. Having characterized the conditions leading to persistent activity emergence in both model cells, our next goal was to search for features of the input and/or the models' response that would be associated with stimulus-selectivity. The presence of ‘memory fields’ has been shown in individual PFC neurons with respect to delay-period activity [44]. That is, a specific neuron exhibits robust delay-period activity (i.e., an increase in firing rate during the delay compared to the stimulus period) only for a specific set of locations in the visual field [3]. The way a PFC pyramidal neuron, however, identifies its memory field remains an open question. It is possible that different incoming stimuli, such as stimuli located in different parts of the visual field, activate synapses in different dendritic branches on PFC pyramidal neurons and this spatial specificity of inputs is in turn used to discriminate between preferred (i.e., those leading to persistent activity) and non-preferred stimuli. In our models, we used 50 simulation trials, in which the set of dendritic branches containing synaptic mechanisms varies with each trial (see Methods). This variability in the location of incoming contacts could be assumed to represent different incoming stimuli [45], hence, we conjecture that the spatial location of activated synapses may play a role in persistent activity induction. To test this hypothesis, we measured the distance from the soma and the center of each dendritic branch that contained stimulated synapses, and averaged the values of each of these features for all dendritic branches in ‘persistent’ versus ‘no persistent’ trials. We found that in both neuron models, synaptic mechanisms were on average located further away from the soma for the ‘persistent’ trials, compared to the ‘no persistent’ trails and this difference was statistically significant (p<0.001) (Fig. 5B, RS model and Supplemental Fig. S3, IB model). The distributions of all activated dendritic segments in ‘persistent’ and ‘no persistent’ trials (Fig. 5C, RS model and Supplemental Fig. S3, IB model) show that this difference stems from a rightward shift as well as a change in the shape of the ‘persistent’ trial distribution due to the activation of dendritic segments located further away from the soma. A possible mechanistic explanation as to why inputs that are further away from the soma lead to persistent firing may be linked to the generation of NMDA spikes. As shown in Fig. 5D, the magnitude of NMDA spikes is much larger when synapses are stimulated in distal compared to proximal locations within the basal dendrites of both model neurons. It is thus possible that distal inputs lead to persistent activity emergence via the facilitation of NMDA spikes which in turn promote the supralinear integration of synaptic inputs [46] and provide much larger and longer lasting somatic depolarizations. These findings are supported by recent experimental data showing that inputs to proximal basal dendrites of cortical pyramidal neurons sum linearly and require precise temporal coincidence for effective summation, whereas distal inputs are combined supralinearly over broader time windows in an NMDAR-dependent manner [47]. Finally, these findings suggest that the relative distance of incoming signals from the cell body may code for the neuron's memory field and therefore, their ability to induce persistent activity. Since the spatial location of incoming contacts is significantly different between ‘persistent’ and ‘no persistent’ trials, it is likely that these differences are reflected in the neuronal response to these stimuli and, if so, this information can be used by downstream neurons to decode the upcoming emergence of persistent activity before it occurs [48]. To test this hypothesis, we first examined whether features of the neuronal response to the stimulus, such as the average firing frequency or the AP latency differed between preferred and non preferred inputs. We found that the average ISIs of the neuronal response during the stimulus was not different between ‘persistent’ and ‘no persistent’ trials in either the RS or the IB neuron model (see Supplemental Table S2). However, the first AP latency of the models' response was clearly different in the ‘persistent’ trials when compared to the ‘no persistent’ trials (see Fig. 6A–D and Supplemental Table S3). Specifically, ‘persistent’ trials in the RS model had AP latencies that were significantly larger than the ‘no persistent’ trails for both NMDA-to-AMPA ratios tested (p<0.001 (N1.2) and p<0.001 (N1.5), non-overlapping boxes in Fig. 6C, D). For the IB model neuron, differences in the AP latencies were highly significant only when the NMDA-to-AMPA ratio was increased to N1.5 (p = 0.0065 (N = 1.2), overlapping boxes in Fig. 6C and p<0.001 (N = 1.5), non-overlapping boxes in Fig. 6D). In both models, persistent activity emergence was associated with a slightly slower onset of the neuronal response, which could be explained by the more distal location of activated synapses, compared to the ‘no persistent’ trials (Fig. 5B–C and Supplemental ). Although the differences in the AP latencies between ‘persistent’ and ‘no persistent’ trials were small (200–300 µs), recent studies have shown that even submillisecond differences in AP emergence or width could represent meaningful coding parameters for neurons [49], [50], [51], suggesting that the magnitude of the AP latency maybe used to code for the occurrence of a preferred stimulus. To test whether differences in the AP latency can discriminate between ‘persistent’ and ‘no persistent’ trials in a more systematic manner, we assessed the ability of the AP latency values to predict the emergence of persistent activity using Linear Discriminant Analysis (LDA). For this, we used a training set (consisting of the AP latencies for 20 ‘persistent’ and 10 ‘no persistent’ trials) in order to determine the optimal cut-off that separates the two distributions based solely on the value of the AP latency. The method was validated using leave-five-out cross validation (LFOCV) and subsequently tested on a previously unseen set of another 30 trials (see Methods), to assess how well can the AP latency of the response to a new input determine whether this input will induce persistent activity or not. Each observation (i.e. trial) of the test set was passed through the 6 ‘trained’ LDA models produced by the LFOCV and a class label was assigned by each model (0 for ‘no persistent’ and 1 for ‘persistent’). All model outputs were then averaged and if the average was 0.5 or higher, then that specific observation was classified as a ‘persistent’ trial, otherwise it was classified as a ‘no persistent’ trial. Using the percentage of correctly predicted ‘persistent’ (sensitivity) and ‘no persistent’ (specificity) trials to assess the method's performance accuracy, we found that discrimination was more successful in the RS than the IB model neuron. Specifically, for the RS model and an NMDA-to-AMPA ratio of 1.2, the sensitivity of the method was very high (100% or 1), while the specificity was a bit lower (0.7), resulting in total accuracy of 0.85 (Fig. 6E–F, black squares). This means that out of the 30 trials tested, all 20 ‘persistent’ trials and 7/10 of the ‘no persistent’ trials are correctly identified by their respective AP values. For the IB model, the sensitivity value was 0.8, the specificity was only 0.4, and the total accuracy was 0.6, considerably decreased compared to the RS model (Fig. 6E–F, ‘x’ marks). However, for a larger NMDA-to-AMPA ratio (N 1.5), the performance accuracy was high for both models, with the RS cell reaching 100% and the IB cell reaching 90% (Fig. 6G). Taken together, these findings suggest that the AP latency may be used as a discriminatory feature for signaling whether a given stimulus will or will not lead to persistent firing, that the accuracy of this prediction is higher in the RS than the IB model neuron and is strongly dependent on the NMDA contribution. We next investigated whether some other characteristic of the stimulus-induced response can better predict persistent activity emergence even for a lower NMDA-to-AMPA ratio. Towards this goal, we used the ISIs during the stimulus-induced response as input to a linear perceptron (see Methods and Fig. 7A for a graphical illustration) and tested whether ‘persistent’ trials could be discriminated from ‘no persistent’ trials based on these features. The perceptron was trained and validated with 30 trials (as in LDA), using the leave-one-out cross validation (LOOCV) method and subsequently tested on a previously unseen set of another 30 trials (see Methods and LDA analysis above). Sensitivity and specificity measures were again used to assess the method's performance accuracy. The sensitivity of the perceptron for both the RS and IB neuron models was 100% when the first 2, first 5, or all ISIs of the stimulus-induced response were used as input (Fig. 7B). Similarly, specificity of the perceptron for both models was above 80%, with the IB slightly better than the RS model (0.9 or greater, Fig. 7C) for all ISI sequences tested. In particular, the specificity for the RS model was 80%, 90% and 90% when the first 2, 5 or all ISIs were used as input features (Fig. 7C, black squares) whereas the specificity of the IB model was 90%, 100% and 90%, respectively (Fig. 7C, ‘x’ marks). The perceptron's performance was also assessed on shuffled datasets (in which ‘persistent’ and ‘no persistent’ trials were randomly labeled) and the performance was severely degraded: the sensitivity dropped to 0% and the specificity to 60% for both RS and IB models. These results show that the initial ISIs of the stimulus-induced response contain highly accurate predictive information regarding the emergence of persistent activity in both the IB and the RS model neurons. Overall, our findings suggest that temporal characteristics of the stimulus-induced response, such as the first spike latency and the first 2 ISIs, contain significant predictive information about the emergence of persistent activity beyond the end of preferred stimuli while average characteristics such as the firing frequency don't capture such information, in accordance with data from other brain regions [48], [52], [53]. These findings are particularly important as they pinpoint specific features of the neuronal response, at the single neuron level, which are common across two major sub-types of pyramidal neurons and which encode stimulus preference with respect to persistent activity emergence. If experimentally validated, these findings suggest a potential mechanism by which stimulus-selectivity that initiates in primary cortices may be decoded by downstream PFC pyramidal neurons within less than 100 miliseconds from the stimulus presentation, and this rapid decoding may have serious implications for the expression of goal-directed behaviors that have been documented in the PFC [54]. Since both model neurons seem to use similar codes for stimulus-selective persistent activity induction, we wondered whether their different firing patterns influenced persistent activity at a different level. We thus contrasted the properties of persistent activity, such as its induction threshold, firing frequency and firing pattern and their dependence on CAN and NMDA in the two model cells. In this study, we used a modeling approach to investigate the mechanisms that underlie stimulus-specific induction of persistent activity in two major subtypes of layer V PFC pyramidal neurons: regular spiking and intrinsic bursting cells. We found that persistent activity can emerge in both single neuron models when the basal dendrites are stimulated with realistic synaptic inputs, provided that the CAN mechanism is activated. In addition, we showed that RS and IB model neurons have distinct persistent activity properties, such as different firing patterns during persistent activity as well as differences in its modulation by the NMDA current. More importantly, our findings suggest that the spatial arrangement of activated synapses may determine whether a given signal will lead to persistent activity induction, thus pinpointing a mechanism for stimulus selectivity. Specifically, we found that in both model cells, preferred stimuli consisted of inputs impinging on more distal parts of the basal dendritic tree compared to non-preferred stimuli. Finally, we show that the temporal features of the stimulus-induced response near its onset code for persistent activity induction in both model cells. Specifically, in the RS neuron, both the action potential latency and the first few ISIs are sufficient for discriminating between preferred and non-preferred stimuli while in the IB model neuron only the first few inter-spike-intervals play the same role. These findings suggest the potential decodability of preferred inputs by downstream PFC neurons upon stimulus presentation and long before persistent activity induction. The ability of PFC pyramidal neurons to display neuronal activity that persists after the end of stimulation was first recorded in vivo in monkeys [57]. This persistent activity has been considered as a network property and particularly a property of recurrent networks due to reverberating excitation [5]. Thus, single neurons have to be part of a recurrent network in order to exhibit persistent activity firing. This hypothesis was further corroborated by the fact that persistent activity could not be induced in neurons recorded in PFC slices where many of the recurrent connections could be severed. Following a modification of the artificial cerebrospinal fluid used, persistent activity lasting for about 1–2 seconds could be recorded from single PFC pyramidal neurons in the slice preparation. This persistent activity, or rather the ‘UP’ state, which occurs both spontaneously and following a stimulus [8], [9], [30], is mediated by AMPA and NMDA [9] and is modulated by GABAB currents [58] and dopamine [59], [60]. Single neurons have been shown to exhibit persistent activity following activation of metabotropic receptors, such as the muscarinic acetylcholine receptor (mAchR) and the metabotropic glutamate receptors (mGluR) [10], [11], [12], due to an underlying depolarizing envelope (i.e., dADP) activated by these receptors [34], [35], [61], [62]. The average dADP in PFC pyramidal neurons following a short 20 Hz stimulus ranges between 2 and 8 mV, not large enough to induce persistent activity by itself, as it has been suggested for enthorhinal cortical neurons [63]. Our computational study shows that this small depolarization when coupled with synaptic activation can induce persistent activity in single neurons. The common characteristic of both dADP and NMDA mechanisms is their slow inactivation kinetics, previously suggested to be required for the persistent activity to maintain ‘physiological’ firing rates [31]. Our study examined the role of these mechanisms in persistent activity. We showed that while an increased NMDA modulates the neuronal firing rate during the stimulus, increasing the CAN current specifically increases the firing frequency of persistent activity. Furthermore, while increasing the NMDA current increases the variability of firing during persistent activity, increasing the CAN current decreases this variability. Based on analysis from in vivo delay-period activity, stimulus-selective -and thus more informative (or significant for mediating behavior)- persistent activity has high firing frequency rates (increased compared to cue-response) as well as increased variability [55]. Our results suggest a dual role for both NMDA and CAN current mechanisms: CAN current acts to enhance persistent activity firing but makes it more regular, while NMDA acts to decrease persistent activity firing but increases its irregularity. Thus, a delicate balance between these two mechanisms in vivo is likely to be critical for proper persistent activity firing. Persistent activity in PFC is stimulus-selective, that is, a neuron will only exhibit persistent firing to specific stimuli, for example stimuli that appear on a specific location of the visual field [44]. The selection of stimuli that a neuron responds to is called a ‘memory field’, in analogy to the receptive fields in the visual cortex [64], or the place fields in the hippocampus [65]. Inhibitory mechanisms play a significant role in shaping the memory fields in PFC, since blockade of GABAA receptors disrupts the emergence of stimulus-selective persistent activity [14]. In our study, we made the assumption that different environmental stimuli could be mapped as different spatial arrangements of synaptic inputs on the basal dendrites. Dendritic activation has been shown to map direction-selective responses in the fly [45] as well as place cells in the hippocampus [66], hence, it is possible that different spatial locations in the neuron's receptive field correspond to the activation of spatially distinct synaptic patterns. The fact that persistent activity emergence in the model neurons is associated with activation of synapses that are located further away from the soma suggests that perhaps in vivo circuits are refined so that stimuli within a memory field project to more distal basal dendrites compared to stimuli outside the neuron's memory field. Since persistent activity is characterized by slow kinetics, it is likely that inputs to distal dendrites, which are generally characterized by slower integration, are more suitable for carrying signals related to persistent activity induction. Finally, stimulation of distal dendrites can generate larger NMDA spikes (Fig. 5 and [67]) which will in turn prolong the window for temporal summation of incoming signals thus resulting in larger and longer-lasting somatic depolarization. Therefore, distal inputs may facilitate persistent activity emergence via the enhancement of NMDA spikes [47]. In addition to a strong link between the spatial arrangement of preferred stimuli and the emergence of persistent activity, our data showed that features of the neuronal response during the stimulus such as the AP latency [48] and the first few inter-spike-intervals, can code for the emergence of persistent activity. Specifically, we found that the AP latency in ‘persistent trials’ is on average significantly longer compared to ‘no persistent’ trials in both the RS and IB model neurons. This may be due to the fact that ‘persistent trials’ corresponded to synaptic arrangements in which activated synapses were located significantly further away from the soma than in ‘no persistent’ trials. Although the differences in the AP latencies between ‘persistent’ and ‘no persistent’ trials were submillisecond, several studies suggest that they could still be decoded by downstream neurons [49], [50], [51]. This finding adds to the coding capabilities of the AP latency which has also been found to code for differences in spatiotemporal characteristics of the input in CA1 model neurons [48] as well as the location of sound in secondary auditory neurons [52]. While AP latency was not as powerful predictor at lower NMDA-to-AMPA ratio, particularly in the IB model cell, stimulus selectivity was encoded in the first few inter-spike-intervals of the stimulus-induced response. For both the RS and IB model neurons the emergence (or not) of persistent activity could be predicted with high accuracy when utilizing the first few (2 or 5) ISIs. These findings are particularly important as they suggest the potential decodability of preferred stimuli by neuronal circuits downstream the L5 PFC pyramids, as early as a few hundreds of milliseconds following the stimulus presentation and long before the emergence of persistent activity. In support of this conjecture, recent data suggest that PFC neurons can categorize input signals as early as the stimulus presentation time [68], [69]. This information could in turn be used by downstream striatal [70] and pontine neurons [71], [72] to prepare for the execution of a specific movement and may provide a neuronal basis for goal-directed behavior. Overall, our findings regarding the coding of information in ISIs are in agreement with studies from other brain regions where ISI sequences were shown to contain more information about receptive fields in the visual cortex than the average firing frequency [17], and could be used to filter and modulate receptive fields in retinal ganglion cells [73]. Recently, in vivo patch-clamp techniques uncovered the importance of intrinsic cellular features in active place cell in the hippocampus [74]. Thus, it is now possible to use patch-clamp recordings in PFC pyramidal neurons during virtual working memory tasks to test the prediction that cellular features such as the AP latency or the ISIs can be used to code for the occurrence of preferred stimuli and the emergence of persistent activity. While both IB and RS pyramidal neurons have been documented in the prefrontal cortex [19], [20], their functional role remains unclear. According to recent studies, there could be a link between neuronal sub-types and their preferred target areas. For example, both RS and IB cortical neurons project to the pons (cortico-pontine) or the striatum but no IB neurons project to the contralateral cortex (cortico-cortical) [23]. Similarly, IB neurons in the distal parts of the subiculum project primarily to the medial enthorhinal cortex but not the amygdala [75]. This segregation is likely to be associated with some form of functional specialization of RS and IB neurons. Furthermore, corticopontine neurons, which consist of both RS and IB neurons, in PFC seem more likely to express persistent activity in response to acetylcholine modulation compared to cortico-cortical neurons, in which no IB neurons are found [25]. Our findings are in line with this hypothesis as they support a differential role of RS and IB pyramidal neurons in persistent activity emergence. Moreover, since different neuronal properties have been suggested to provide a recurrent network with different persistent activity characteristics [76], our data suggest that RS and IB neurons may form distinct subnetworks when connected in a recurrent network. Future modeling and experimental work is needed to further investigate this hypothesis. Dopamine, acting through D1/5 receptors, increases the NMDA current [42] while it decreases the dADP [11]. Our computational study, as well as a previous one [77], suggested that increasing the NMDA component of synaptic stimulation decreases the CAN current required for induction of persistent activity. Thus, D1 signaling seems to modulate both of these mechanisms in order to maintain stability of neuronal excitability. In the case where only NMDA currents are increased while the dADP amplitude in response to metabotropic receptors remains the same, persistent activity will be elicited even in response to non-relevant stimuli. On the other hand, if dADP alone was reduced without any change in the NMDA currents, then no stimulus would be able to induce persistent activity. Our modeling work showed that when increasing the NMDA current in a similar amount that DA does, the ISI variability increases in both model neurons but it remains elevated for the entire 3 second recoding period of persistent activity only in the IB model neuron (Fig. 9F). Furthermore, increasing the NMDA current also changes the persistent activity properties of the RS model neuron to resemble those of the IB model neuron, with a bursting firing pattern during persistent activity and the emergence of time-limited persistent activity (Fig. 9B and 4D). Our results are in agreement with the well established idea that an increase in DA is necessary for proper expression of persistent activity since the properties of persistent activity in our model neurons are closer to the ones observed in vivo, when the NMDA current contribution is increased. Furthermore, increasing the NMDA current contribution also improves the ‘persistent’ vs. ‘no persistent’ trials discrimination particularly in the IB model neuron, enhancing the coding capabilities of this neuronal subtype, since both the AP latency and/or the first few ISIs can be used to predict the emergence of persistent activity. DA also modulates other biophysical mechanisms in pyramidal neurons of the prefrontal cortex, such as the L-type calcium channels [78], [79], sodium currents [27], [80] and potassium currents [81], [82]. Modulation of all these mechanisms is likely to affect properties of persistent activity; however, such an analysis is beyond the scope of this work. Our detailed model reproduces closely the electrophysiological activity of PFC pyramidal neurons. Nonetheless, sources of inaccuracy may have been introduced since the experimental data used to constrain the model are products of in vitro preparations. In that sense however, model limitations do not significantly differ from those of the in vitro preparations whose findings are readily replicated by the model. Simplifications that have been adopted in this work include: (i) a strictly phenomenogical model of the CAN current, (ii) absence of background synaptic activity which is known to occur in vivo (although we do include membrane noise), (iii) stimulation delivered only to the basal dendrites of the model neurons (which are known to receive the majority of inputs from other cortical areas), while more spatially distributed stimulation in combination with the experimentally observed accumulation of extracellular potassium [83] could reduce the threshold for persistent activity induction and require more complex spatial coding features, (iv) same biophysical mechanisms in both model cells, yet different conductance values for the R-type calcium and sodium currents, while in nature, there is probably some variability, and (v) no modeling of plasticity or neuromodulator effects. In spite these simplifications, our model findings are important as they have identified neuronal features that could code for the emergence of persistent activity in both neuronal subytpes found in the cortex, as well as differential properties of persistent activity between RS and IB neurons. In summary, our modeling results allow the formulation of several predictions, which when tested experimentally could further the current knowledge on persistent activity, its underlying mechanisms and its contribution to working memory. First, we predict that the location of activated synapses is critical for the emergence of persistent firing: stimuli that lead to persistent activity consist of inputs that arrive in distal parts of the basal tree, far away from the soma. This prediction could be easily tested experimentally in slice preparations that generate Up and Down states. Specifically, synaptic stimulation of proximal versus distal basal dendrites should have a different impact on the probability of generating Up states and/or modulating the firing frequency or duration of the Up state. Second, we have identified the AP latency and first few ISIs of the stimulus-induced response as features that could discriminate between stimuli that result in persistent activity or not. These findings could be tested experimentally in slice preparations and/or in vivo when a stimulus is used to induce persistent activity. Finally, our models predict that the dADP threshold for persistent activity induction is lower in RS than IB neurons, suggesting that RS neurons should comprise the majority of layer 5 PFC pyramidal neurons that exhibit persistent firing. This can also be tested experimentally in slice preparations that generate Up and Down states, where the effect of synaptic stimulation on the UP state is contrasted between RS and IB neurons. Overall, our modeling work identifies key features of the neuronal response that could predict the emergence of persistent activity and pinpoints a differential role of RS and IB model neurons in persistent activity properties. A detailed compartmental model of a layer V PFC pyramidal neuron comprising of a large variety of membrane mechanisms was implemented in the NEURON simulation environment [84] and was applied on a reconstructed layer V PFC pyramidal neuron available at the neuromorph database (neuron C3_5 from the Smith lab, http://neuromorpho.org/neuroMorpho/index.jsp, shown in Fig. 1A). This neuron was taken from adult Long-Evans rats, at 64–78 days of age [85]. When converted to NEURON, C3_5 had a total of 45 compartments (1 somatic, 1 axonic, 18 basal, and 25 apical dendritic compartments). We assumed a uniform membrane resistance of Rm = 30 kΩ.cm2; a uniform intracellular resistivity Ra = 100 Ω.cm; and a specific membrane capacitance of 1.2 µF.cm−2 in the soma and 2.0 µF.cm−2 in the dendrites. The resting membrane potential was set at −66 mV. Active mechanisms included two types of Hodgkin–Huxley-type Na+ currents (transient: INaf; persistent INap;), three voltage-dependent K+ currents (IKdr; IA; ID), a fast Ca++ and voltage-dependent K+ current, IfAHP; a slow Ca++-dependent K+ current, IsAHP; a hyperpolarization-activated non-specific cation current (Ih); a low-voltage activated calcium current IcaT; three types of Ca++- and voltage-dependent calcium currents (IcaN; IcaR; IcaL); and the calcium-activated non-selective cation (CAN) current. Channel equations for all the different voltage-gated calcium currents (IcaN; IcaR; IcaL, ICaT), IsAHP, IA and Ih are described in [86], while the channel equations for IKdr, ID, IfAHP, INap and INaf are described in [56]. A combination of variations of the R-type Ca++ current and the persistent Na+ current mechanism could generate two different firing patterns: a) a regular spiking (RS) and b) an intrinsic bursting (IB) firing pattern. Specifically, doubling the conductance values for both of the two aforementioned currents switched the firing behaviour of the model neuron from an RS one to an IB one (see Supplementary Text S1). The possibility that such a variability exists in PFC pyramidal neurons is evident from data showing that the current density of persistent sodium current ranges from 2 to 6 pA/pF in PFC pyramidal neurons [27], and that the R-type calcium current contribution could range from 5–25% of total calcium current in layer V pyramidal neurons [26]. The equation for the Ca++-activated non-selective cation (CAN) current mechanism was based on [87]. The ‘cac’ and ‘beta’ parameters were adapted so that the current reached its maximum value by the first second following the end of the inducing stimulus and decayed with a time constant of 3 seconds, as observed experimentally (Fig. 3). The actual equations used for the CAN mechanism are the following: , and , based on [35], where about 70% of the dADP is Na+ current. The state m was calculated by the following set of equations(1)(2)(3)(4)where beta = 0.00001(1/ms) and cac = 0.0004 (mM). The beta and cac values were adjusted so that the dADP was induced following more than 4 spikes and had decay kinetics in the order of a few (∼3) sec. Furthermore, mechanisms modelling four synaptic currents, AMPA, NMDA, GABAA and GABAB, were used in the model neuron (channel equations also described in [86] and Supplementary Text S1). The ‘default’ NMDA-to-AMPA ratio was set to 1.2, based on [31]. The ratio was measured based on the somatic current recorded at −70 mV and +60 mV, following stimulation of 10 synapses on the basal dendrites. We also performed these experiments with the same number of synapses used in the ‘persistent activity’ simulations, and did not find any difference in the ratio of +60 mV and −70 mV currents. In order to achieve this NMDA-to-AMPA ratio, the gNMDA was equal to 3× the gAMPA. In order to increase the ratio to 1.5, the gNMDA was equal to 4× the gAMPA. Both the RS and IB models were validated with respect to passive and active membrane properties as well as apical and basal dendritic responses (see Supplemental Fig. S1). Dendritic and somatic voltage traces in response to glutamate release were validated based on experimental data [32] (see Supplemental Fig. S2). Dendrites were stimulated with a total of 200 excitatory synapses (containing both AMPA and NMDA receptors), equally distributed in 10 different dendritic branches (20 synapses on each branch) which were activated 10 times at 20 Hz. The 10 dendritic branches were selected randomly from the pool of all basal dendrites. Synapses were distributed at random locations within each branch, according to a uniform distribution. The soma of the neuron model was stimulated with 5 inhibitory synapses [39] at 50 Hz. Both excitatory and inhibitory synapses were activated synchronously, without any temporal variability between different trials. Since we were interested in studying the suprathreshold response of neurons to a specific stimulus, we used the above number of synapses for which the neuron model responded with at least 5 APs in the 10 event stimulus. In addition, for best simulation of membrane potential fluctuations as observed in vitro due to the stochastic ion channel noise [88], [89], an artificial current with Poisson characteristics was injected in the soma of both RS and IB neuron models. This simulation of channel noise is simple compared to ones recently reported [89], yet sufficient for the complexity of the model used and the purpose of the current study. We define persistent activity as the prolongation of neuronal activity following the end of the stimulus for at least 3 seconds. All simulations were recorded for 5 seconds, and if neuronal activity persisted past the 3 seconds following the stimulus, it did not stop before the 5 sec recording. Simulations included 50 repetition trials for each condition (i.e. specific set of NMDA and CAN conductances), where the spatial distribution of activated synaptic mechanisms at the different basal dendritic branches changed in each trial. For the data analysis, only conditions in which persistent activity emerged in at least 50% of the runs were used, unless otherwise noted. Estimation of inter-spike-intervals (ISIs) of the simulated neuronal responses, as well as generation of the ISI histograms was performed with custom-made macros using IgorPro software (Wavemetrics, Inc) and Matlab (Mathworks, Inc). Prediction of persistent activity emergence based on the ISI values was done using a custom-made Artificial Neural Network written in Java. The network used was a simple perceptron, which was initially trained with 30 randomly selected trials, validated (using leave-one-out cross validation) and subsequently tested on another 30 trials (both training and test sets comprised of 20 persistent and 10 no persistent trials). Prediction of persistent activity emergence based on the AP latency values was done using Linear Discriminant analysis (LDA) in Matlab (Mathworks, Inc.), with code downloaded from the file exchange site (http://www.mathworks.com/matlabcentral/fileexchange/29673-lda-linear-discriminant-analysis). The method was initially trained with 30 randomly selected trials, validated (using leave-five-out cross validation), and then tested on another set of unseen 30 trials, similarly to the Perceptron analysis. The following conditions were used for the prediction analysis: RS neuron model −N1.2/−N1.5, 40 persistent trials, 20 no persistent trials; IB neuron model −N1.2/−N1.5, 40 persistent trials, 20 no persistent trials (as shown below in the text: N = NMDA-to-AMPA ratio, dADP = delayed afterdepolarization). In each trial, synaptic mechanisms were placed according to a uniform distribution within 10 dendritic branches which were selected at random. For each of the selected dendritic branches, the path distance from the center of the branch to the soma was calculated and used to estimate the average dendritic distance for any given trial. All experimental results shown here are parts of the neuronal database recorded by Kyriaki Sidiropoulou while she was at Dr. Francis White laboratory and have been reported in previous publications [11], [28]. The model is available for download from ModelDB (Link: https://senselab.med.yale.edu/modeldb/ShowModel.asp?model=144089, accession number: 144089).

10.1371/journal.ppat.1004284

Paenibacillus larvae Chitin-Degrading Protein PlCBP49 Is a Key Virulence Factor in American Foulbrood of Honey Bees

Paenibacillus larvae, the etiological agent of the globally occurring epizootic American Foulbrood (AFB) of honey bees, causes intestinal infections in honey bee larvae which develop into systemic infections inevitably leading to larval death. Massive brood mortality might eventually lead to collapse of the entire colony. Molecular mechanisms of host-microbe interactions in this system and of differences in virulence between P. larvae genotypes are poorly understood. Recently, it was demonstrated that the degradation of the peritrophic matrix lining the midgut epithelium is a key step in pathogenesis of P. larvae infections. Here, we present the isolation and identification of PlCBP49, a modular, chitin-degrading protein of P. larvae and demonstrate that this enzyme is crucial for the degradation of the larval peritrophic matrix during infection. PlCBP49 contains a module belonging to the auxiliary activity 10 (AA10, formerly CBM33) family of lytic polysaccharide monooxygenases (LPMOs) which are able to degrade recalcitrant polysaccharides. Using chitin-affinity purified PlCBP49, we provide evidence that PlCBP49 degrades chitin via a metal ion-dependent, oxidative mechanism, as already described for members of the AA10 family. Using P. larvae mutants lacking PlCBP49 expression, we analyzed in vivo biological functions of PlCBP49. In the absence of PlCBP49 expression, peritrophic matrix degradation was markedly reduced and P. larvae virulence was nearly abolished. This indicated that PlCBP49 is a key virulence factor for the species P. larvae. The identification of the functional role of PlCBP49 in AFB pathogenesis broadens our understanding of this important family of chitin-binding and -degrading proteins, especially in those bacteria that can also act as entomopathogens.

American Foulbrood and its etiological agent, Paenibacillus larvae, pose a serious threat to global honey bee health. So far, molecular mechanisms of host-microbe interactions are poorly understood in this system and no key virulence factor for the entire species has been identified. Here, we demonstrate that P. larvae expresses a chitin-binding and -degrading protein PlCBP49 harboring one module that belongs to the auxiliary activity 10 (AA10) family of lytic polysaccharide monooxygenases (LPMOs). We provide evidence that PlCBP49 degrades chitin via a metal ion-dependent, oxidative mechanism, as already described for other members of the AA10 enzyme family. Using P. larvae mutants lacking PlCBP49 expression, we demonstrate that PlCBP49 is crucial for the degradation of the chitin-rich peritrophic matrix, a key step in pathogenesis of P. larvae infections. In the absence of PlCBP49 expression the peritrophic matrix remained nearly intact and about 95% of the infected larvae survived infection. This clearly indicated that PlCBP49 is a key virulence factor of P. larvae. These results constitute important progress in our understanding of both P. larvae pathogenesis and the biological role of LPMOs in entomopathogens. Furthermore, knowing PlCBP49 and its role in pathogenesis opens new possibilities to develop curative measures for this disease.

Vertebrates and invertebrates alike need to protect their intestinal epithelia against various chemical, physical and biological challenges while the transport of nutrients and water must remain uninterrupted to aid in digestion. For this purpose, mucosal secretions line the digestive tract of vertebrates, while in most invertebrates the midgut epithelium is lined by the peritrophic membrane or peritrophic matrix (PM) [1], [2], an organized layer made of secreted chitin and (glyco)proteins. Chitin, an insoluble linear β-(1,4)-linked polymer of N-acetyl-D-glucosamine (GlcNAc), is considered the major structural component of the PM where it forms chitin fibrils. These fibrils are held together by chitin-binding proteins, while the interstitial spaces are filled by glycans (for a recent review see [3]). The resulting lattice acts as a molecular sieve with a large range of pore sizes. The functions attributed to the PM include (i) compartmentalization of digestive processes, (ii) protection from ingested xenobiotics, and (iii) acting as mechanical barrier against abrasive foodstuffs and pathogens (for a recent review see [2]). The latter function makes the PM a first-line defense against ingested pathogens and, hence, an important part of the invertebrates' complex system to combat infections. Accordingly, insect-pathogenic bacteria infectious per os must breach the PM before they can attack and invade or cross the epithelium. To this end, hydrolytic enzymes such as proteases and, most importantly, chitin-degrading proteins enabling PM degradation are produced and secreted by these pathogens. Most of the chitin-degrading proteins are chitinases belonging to the family of glycosyl hydrolases (GH). To date, 133 different GH families classified on the basis of sequence similarities and forming 14 clans (GH-A – GH-N) of related families based on similarities in protein folds can be found in the Carbohydrate Active Enzymes (CAZy) database [4]. Another family of bacterial chitin-degrading proteins comprises proteins with a carbohydrate-binding module (CBM) and belong to the auxiliary activities 10 (AA10) family (formerly chitin binding module 33 (CBM33) family) of lytic polysaccharide monooxygenases (LPMOs) as defined in the Carbohydrate Active Enzymes (CAZy) database [4]. These proteins were originally thought to lack any catalytic activity of their own [4], although they had been shown to be involved in chitin degradation [5]. However, it was recently demonstrated that CBP21, an AA10 (formerly CBM33) family member expressed by the Gram-negative soil bacterium Serratia marcescens [6], [7], as well as EfCBM33A, expressed by the Gram-positive, opportunistic pathogen Enterococcus faecalis, are capable of degrading crystalline chitin via a novel, copper-dependent, oxidative enzymatic mechanism [8]–[10]. LPMOs comprise only two families, (i) the AA10 (formerly CBM33) family with bacterial, viral, and some eukaryotic members and (ii) the AA9 (formerly glycoside hydrolase 61 (GH61)) family with only fungal members so far. Both families are monooxygenases and target recalcitrant polysaccharides. Bacterial chitin-degrading proteins are produced mainly to meet nutritional needs of the bacteria because chitin-degradation products, once transported into the bacterial cell, can be used as carbon sources [11], [12]. Because, in addition, many insect pathogens need to overcome chitin-containing structures (e.g., cuticula or peritrophic membranes) to enter a host and establish an infection, degradation of the PM by bacterial pathogens might be a process serving two purposes: nutrition and invasion. Paenibacillus larvae is a bacterial pathogen of honey bee larvae which causes a notifiable disease called American Foulbrood (AFB) [13]. AFB is a highly contagious disease and is fatal for the entire colony when detected in too late a stage of disease. However, since most authorities consider burning of diseased colonies and contaminated hive material the only workable control measure, diseased colonies are inevitably lost in most cases. Hence, AFB causes considerable economic losses in apiculture worldwide. The etiological agent, P. larvae, is a Gram-positive, rod-shaped bacterium forming tenacious spores under adverse environmental conditions like lack of nutrients. These spores are the infectious form of P. larvae and they initiate a fatal infectious process in bee larvae once ingested with contaminated larval food. Honey bee larvae become less susceptible to infection with increasing age and as soon as two days after egg hatching they can be considered “resistant” (for recent reviews: [14]–[16]). This phenomenon has been attributed to the growing PM already in the early days of AFB research [17], [18]. Recently, we demonstrated that the honey bee larval gut is lined by a chitin-containing PM which is degraded during P. larvae infection [19], confirming earlier results showing that the PM is the first barrier the bacteria have to overcome when trying to breach the epithelium and to enter the haemocoel [20]. Proteases and chitin-degrading proteins are most likely involved in this process. For P. larvae it is long since known that an impressive number of proteases is expressed and secreted [21], [22]. These proteases, although poorly characterized, have already been implicated as virulence factors as they might aid in degrading the PM, breaching the epithelium and converting larval into bacterial biomass [23]–[25]. In contrast, little is known so far about the nature and expression of P. larvae chitin-degrading proteins and their role in the pathogenesis of AFB. Recently, the genomes of representatives of two P. larvae genotypes, ERIC I and ERIC II, were sequenced, annotated and used for comparative genome analysis [26]. Surprisingly, no complete, uninterrupted and, hence, putatively functional gene coding for a classical chitinase could be detected in any of the genomes [26] posing the intriguing question: how is the described chitin-degradation by P. larvae during infection [19] achieved? We are now answering this question by describing the identification and functional characterization of P. larvae PlCBP49, a novel member of the AA10 (formerly CBM33) family of chitin-binding and –degrading LPMOs. PlCBP49 was confirmed in the genome and secretome of P. larvae. Chitin-affinity purified PlCBP49 was used to demonstrate chitinolytic activity both on soluble and insoluble substrates as well as to confirm that chitin degradation was metal ion-dependent and involved an oxidative step. Furthermore, we studied the biological role of PlCBP49 in P. larvae infected larvae and were able to demonstrate that PlCBP49 is involved in PM degradation during infection and is a key virulence factor of P. larvae. Recently, degradation of the larval midgut PM by P. larvae was demonstrated to be a key step in the pathogenesis of P. larvae infections [19]. In order to identify proteins which might be involved in this process, we isolated and analyzed chitin-binding proteins from P. larvae culture supernatants using chitin-coated beads. SDS-PAGE analysis of the chitin-binding fractions revealed two chitin-binding proteins (CBP) migrating around 60 kD (CBP60) and 49 kD (PlCBP49I) secreted by ATCC9545 (P. larvae ERIC I), while only one band migrating around 49 kD (PlCBP49II) was visible in the supernatant of DSM25430 (P. larvae ERIC II) (Fig. 1A). To determine the chitinolytic activity of the isolated proteins, the chitin-binding fractions were subjected to zymography performed with denaturing gels containing ethylene glycol chitin (EGC) as the substrate (Fig. 1B). The chitin binding fractions of both isolates produced a clear zone in the range of 49 kD suggesting that in both strains a protein migrating around 49 kD was able to degrade the chitin analogue EGC. We hypothesized that these bands corresponded to PlCBP49I and PlCBP49II identified in SDS-PAGE analysis of the chitin-binding fractions of ATCC9545 and DSM25430, respectively. In addition, chitinolytic activity could also be detected around 42 kD in the chitin-binding fraction of ATCC9545 and a similar activity was weakly but not reproducibly detectable in the chitin-binding fraction of DSM25430. These activities could not be related to any proteins detectable in the Coomassie-stained gels of the chitin-binding fractions hampering their further characterization. This phenomenon was not surprising because zymography is a highly sensitive method of detecting enzymatically active proteins even when present in such low amounts that they cannot be visualized with common staining procedures. In the chitin-binding fraction of ATCC9545, no chitinolytic activity could be observed around 60 kD indicating that CBP60, although binding to chitin, was not able to degrade EGC. To further substantiate that PlCBP49I and PlCBP49II are not only chitin-binding but also chitin-degrading proteins, we determined their protein sequences via mass spectrometry analysis after separation of the bead-bound proteins by SDS-PAGE. Comparison of the obtained peptide sequence data suggested that PlCBP49I was identical to PlCBP49II (Fig. 2A). Protein BLAST analysis of the peptide sequences indicated that PlCBP49I and PlCBP49II were not classical chitinases but rather homologs of CBP21, a chitin-binding and –degrading protein of the chitinolytic bacterium S. marcescens belonging to the AA10 family of LPMOs [5], [8], [10] (Fig. 2A). To analyze whether ATCC9545 and DSM25430 harbor a common gene putatively coding for PlCBP49 we determined the sequences of the corresponding genes cbp49I and cbp49II in ATCC9545 and DSM25430, respectively, by comparing the obtained peptide fragment sequences with the genomic sequences of P. larvae BRL 230010 [27] via TBLASTN analysis [28]. Nucleotide sequence analysis confirmed the presence of the same functional open reading frame (ORF) in both strains (Genbank accession numbers JX185746 for ATCC9545, JX185745 for DSM25430) indicating that both strains harbor the gene cbp49 coding for the protein PlCBP49 with 443 amino acids. Screening a collection of P. larvae field isolates for the presence of cbp49 confirmed the presence of this gene in all strains analyzed so far (data not shown). Protein BLAST alignment of the translated genomic sequence of PlCBP49 followed by domain analysis of the deduced amino acid sequence using the Conserved Domain Database (CDD; NCBI) revealed a modular protein with an N-terminal domain homologous to AA10 (formerly CBM33) family members of LPMOs followed by two fibronectin type III-like domains, which are often found in bacterial glycosyl hydrolases, and a second small chitin-binding domain (CBM 5/12) at the C-terminus, which is also found in many different glycosyl hydrolase enzymes and presumed to have a carbohydrate binding function (Fig. 2B). Similar domain architecture has recently been described for several other members of the AA10 family of LPMOs [29]–[31]. Amino acid alignment of the AA10 domain of P. larvae PlCBP49 protein with other AA10 family members (Fig. 2C) identified an N-terminal signal sequence (framed) and two histidine residues (His35, His122; arrows) which are highly conserved in all proteins of the AA10 family and which are involved in coordinating copper [8], [10]. Some other conserved residues implicated in binding to chitin [5] were also present such as Tyr62, Glu68, Thr119, Ala120, Asp188, and Thr189 (asterisks) [8], [10] with Tyr62 in P. larvae (red asterisk) corresponding to Trp56 in CBP21 and Tyr54 in EfCBM33A [8], [10]. These results indicated that P. larvae PlCBP49 contains a domain characteristic for the AA10 family of LPMOs and suggested that PlCBP49 has LPMO activity. To further substantiate that P. larvae PlCBP49 is a member of the AA10 family of LPMOs we analyzed the chitin-degrading activity of PlCBP49 in greater detail by zymography in the presence of the metal chelator ethylenediaminetetraacetic acid (EDTA) and the di-oxygen mimic cyanide, which were both shown to be potential inhibitors of AA10 family members [8], [9]. P. larvae PlCBP49 activity was strongly inhibited in the presence of 20 mM EDTA and 2 mM potassium cyanide (KCN) (Fig. 3) supporting the involvement of divalent cations and the crucial role of the oxidative step in chitin degradation through PlCBP49. In contrast, presence of 20 mM caffeine, a competitive inhibitor of chitinases [32], did not inhibit chitin degradation (Fig. 3) suggesting that PlCBP49 indeed is not a classical chitinase. To better analyze the chitin-degrading activity of PlCBP49, to functionally characterize this protein in the P. larvae genotypes ERIC I and ERIC II [13], [16], and to assess its role in the pathogenesis of AFB, P. larvae mutants deficient in the expression of PlCBP49 were constructed from ATCC9545 and DSM25430 as parent strains. The chitin-binding fractions of both mutant strains were analyzed by SDS-PAGE (Fig. 4A) and zymography (Fig. 4B) to demonstrate loss of PlCBP49 expression and of chitinolytic activity concomitant with cbp49 gene disruption. In both mutants, ATCC9545 Δcbp and DSM25430 Δcbp, the protein bands corresponding to PlCBP49 (Fig. 4A) as well as the chitinolytic activity migrating around 49 kDa (Fig. 4B) were missing in the corresponding chitin-binding fractions. In contrast, the chitinolytic activity visible around 42 kDa was unaffected (Fig. 4B) indicating that gene disruption of cbp49 had no downstream effects on the chitinolytic machinery detectable via zymography. These results clearly confirmed that the identified cbp49 gene encodes the identified protein PlCBP49 which in turn is responsible for the observed activity towards soluble chitin. We recently showed that P. larvae is able to metabolize insoluble, colloidal chitin [19]. However, the data related to the chitin-degrading activity of PlCBP49 obtained so far were based on using EGC as a soluble substrate in zymograms. To further verify that degradation of insoluble recalcitrant polysaccharides is mediated by PlCBP49, as it is described for members of the AA10 family of LPMOs, we tested whether or not PlCBP49 might be able to degrade chitin-containing structures like an insect PM. To this aim, we used an Ussing chamber (Fig. 5A) to perform permeability assays with PMs which were isolated from S. frugiperda last instar larvae and subjected to the chitin bound fractions of the knock-out P. larvae strains (ATCC9545 Δcbp, DSM25430 Δcbp) and of the corresponding parent wild-type strains (ATCC9545, DSM25430). This comparative approach allows differences in PM permeabilization between mutant and wild-type bacteria to be linked with differences in PlCBP49 expression. Permeability of the PMs was measured as methylene blue (MB) efflux and was significantly higher after incubation with the chitin-binding fractions of the wild-type bacteria than after incubation with the chitin-binding fractions of the corresponding mutants (Fig. 5B). For P. larvae ATCC9545 MB efflux significantly (student's t-test, p-value = 0.0125) decreased from 0.0429±0.006 µg/ml/mm2/h in the presence of PlCBP49 expression (ATCC9545, mean values ± SEM) to 0.021±0.0016 µg/ml/mm2/h in the absence of PlCBP49 expression (ATCC9545 Δcbp, mean values ± SEM). Similar results were obtained for P. larvae DSM25430 (0.02418±0.002 µg/ml/mm2/h, mean values ± SEM) compared to P. larvae DSM25430 Δcbp (0.0094±0.0013 µg/ml/mm2/h, mean values ± SEM) which were also significantly different (student's t-test, p-value = 0.004). Remarkably, exposure of PMs to chitin-bound fractions of DSM25430 Δcbp resulted in PM permeability not significantly different from the negative control (student's t-test, p-value = 0.983) indicating that in the absence of PlCBP49 expression no PM degrading activity was active in these fractions. In contrast, chitin-bound fractions of ATCC9545 Δcbp resulted in PM permeability that was significantly higher than the negative control (student's t-test, p-value = 0.006) although also significantly reduced when compared to the effect achieved with ATCC9545 wild-type, meaning in the presence of PlCBP49 expression. These results indicated that the chitinolytic activity of PlCBP49 can act on chitin in its native crystalline form suggesting a role for PlCBP49 also in PM degradation observed during P. larvae infection of honey bee larvae [19]. To test the proposed involvement of PlCBP49 in PM degradation in P. larvae infected honey bee larvae, we isolated (Fig. 6A) and stained PMs (Fig. 6B–D) from non-infected control larvae as well as from larvae infected with either wild-type P. larvae (ATCC9545) or mutant P. larvae (ATCC9545 Δcbp) in order to assess PM integrity. While control larvae at the age of six days after egg hatching contained intact PMs (Figs. 6A, B), the PM in ATCC9545-infected larvae of the same age appeared almost totally degraded with only small patches of stainable structures (Fig. 6C). In contrast, larvae of the same age infected with ATCC9545 Δcbp and, therefore, lacking the chitin-degrading activity of PlCBP49, did still contain an almost intact and stainable PM (Fig. 6D). This result indicated that PM degradation only occurs in the presence of PlCBP49 expression and, therefore, confirmed that PlCBP49 plays a key role in the degradation of the larval PM during infection. PM degradation in P. larvae infected larvae has been shown to be a key step in AFB pathogenesis [19]. Therefore, we hypothesized that PlCBP49 involved in PM degradation in infected larvae should be a key virulence factor of this pathogen. To test this, we performed laboratory infection assays by feeding first instar honey bee larvae larval diet containing spores of the knock-out P. larvae strains (ATCC9545 Δcbp, DSM25430 Δcbp) and of the corresponding parent wild-type strains (ATCC9545, DSM25430). Comparing mortality in the groups infected with the wild-type and with the mutant bacteria revealed that the lack of PlCBP49-expression led to a significant reduction in mortality (Fig. 6E). While 500 cfu/ml of ATCC9545 resulted in a total mortality of 75.5%±5.0% (mean values ± SEM), the same spore concentration killed only 4.4%±1.1% (mean values ± SEM) in the group infected with the mutant strain ATCC9545 Δcbp (student's t-test, p-value = 0.0001). The results were similar for DSM25430 and DSM25430 Δcbp: 100 cfu/ml killed 77.7%±8.7% (mean values ± SEM) in the group infected with the wild-type DSM25430 and only 3.3%±1.9% (mean values ± SEM) died in the group infected with DSM25430 Δcbp (student's t-test, p-value = 0.0011). In other words, lack of PlCBP49 activity resulted in about 94% reduction in mortality for ATCC9545 and nearly 96% reduction in mortality for DSM25430. Therefore, the lack of PlCBP49 expression nearly abolished the virulence of both P. larvae strains. These results clearly indicated that PlCBP49 is a key virulence factor for both P. larvae genotypes. Attempts to understand the pathogenesis of P. larvae infections of honey bee larvae have been undertaken since the early days of AFB research. It was suggested that the PM acts as a protective shield against P. larvae infections because it might hamper vegetative bacteria in their attempts to cross the gut wall [17], [33]–[35]. Recently, we demonstrated that the PM indeed is the first barrier vegetative P. larvae bacteria have to overcome before attacking and breaching the epithelium [20]. In addition, we recently showed that P. larvae is able to metabolize colloidal chitin [19] and that the larval midgut PM contains chitin which is degraded during P. larvae infection, leading to loss of PM integrity [19]. We now describe the identification and functional analysis of P. larvae PlCBP49, a new member of the AA10 (formerly CBM33) family of chitinolytic LPMOs which were recently shown to be able to degrade crystalline chitin via a novel, oxidative mechanism [8]. We demonstrate that PlCBP49 is involved in PM degradation in infected honey bee larvae and that it is a key virulence factor for both P. larvae genotypes, as in the absence of PlCBP49 expression P. larvae virulence was almost lost (for a definition of the term virulence please see [36]). When invading the infected larva, P. larvae faces two physical barriers: first the peritrophic membrane and second the epithelium [20]. Although the destruction of the epithelial cell layer by P. larvae is still poorly understood, the identification and characterization of PlCBP49 now allows the biochemical mechanism by which P. larvae penetrates and traverses the PM to be unravelled. Cbp49 encodes a chitin-binding and –degrading protein belonging to the AA10 (formerly CBM33) family of chitin-degrading LPMOs. The best analyzed member of this family is CBP21 expressed by S. marcescens, a Gram-negative soil bacterium described as an insect pathogen but which is also an opportunistic pathogen of mammals [6], [37], [38]. CBP21 was originally described as lacking any catalytic activity of its own despite its essential role in chitin-degradation [5]. However, it was recently shown to possess chitin-degrading activity [8] and it was demonstrated that CBP21 degrades crystalline chitin through a novel copper-dependent oxidative mechanisms [8], [10]. So far, the only function attributed to CBP21 is that of a food-scavenging enzyme allowing S. marcescens to use chitin as carbon source. However, S. marcescens is pathogenic to many invertebrates including insects where it causes intestinal infections [37]. Based on our results we now propose that the biological role of S. marcescens CBP21 is not only to enable these bacteria to use chitin as carbon source but also to allow them to degrade the protective PM as part of the pathogenic process when S. marcescens acts as insect pathogen. This assumption is further substantiated by recent findings showing that Drosophila melanogaster deficient in forming a proper PM due to a loss-of-function mutation in the gene dcy coding for Drosocrystallin, an integral component of D. melanogaster PM, are more susceptible to oral infection by S. marcescens than wild-type flies [39]. Furthermore, for EfCBM33A it was recently speculated that it is involved in host-microbe interactions based on the observed gene regulation [9]. PlCBP49 is 443 amino acids in length and in addition to the AA10 module also contained two fibronectin type III-like domains and another chitin-binding module, CBM5/12. The latter domain has often been found in proteins expressed by bacteria thriving in the digestive tract of invertebrates [40]. For instance, CBP50 expressed by B. thuringiensis has a similar domain architecture [30] and it was demonstrated that the FN III-like domains and the CBM5/12 domain are involved in substrate binding. Future experiments should address the question of the function of these domains for PlCBP49 activity. According to our results, PlCBP49 not only degraded insoluble chitin structures but also acted on EGC, a soluble analog of chitin. Further biochemical and molecular studies are needed to unravel the function of these additional modules in the context of LPMO activity and to analyze whether these modules are involved in cleavage of soluble chitin, a capability not yet described for other members of the AA10 family of LPMOs. Normally, Bacilli and Paenibacilli express a wide range of classical chitinases. However, recent evidence suggests that the situation is totally different for P. larvae. When the entire genomes of two P. larvae strains representing the genotypes ERIC I (DSM25719) and ERIC II (DSM25430) were sequenced, manually curated, annotated, and searched for putative virulence factors, in both genomes only a pseudogene containing a chitinase A (GH18 family) N-terminal domain was identified. While the P. larvae genomes harbored more than 100 protease genes, genes coding for classical chitinases were missing [26]. These most recent in silico results further support our finding that PlCBP49 is a key chitin-degrading protein of P. larvae. By using a chitin-binding activity based approach, we identified PlCBP49 as chitin-degrading protein and we demonstrated that PlCBP49 is able to degrade both soluble and insoluble chitin and is essential for PM degradation in vitro and in vivo. However, while in the Ussing chamber experiment gene disruption of cbp49 in DSM25430 resulted in loss of S. frugiperda PM degradation by DSM25430 Δcbp, the chitin bound fraction of ATCC9545 Δcbp still seemed to contain some activity affecting PM integrity. We speculate that this activity might be due to CBP60, present only in ATCC9545 and ATCC9545 Δcbp chitin bound fractions but absent in the corresponding DSM25430 fractions (Fig. 1). Although CBP60 is a chitin-binding protein, it did not show any chitin-degrading activity. However, CBP60 might have protease activity targeting non-chitin components of the PM like insect intestinal mucin (IIM) or other proteins thereby affecting PM integrity. More than hundred putatively functional protease genes were present in the P. larvae genomes [26] leaving abundant CBP60-candidates which might contribute to PM degradation in ATCC9545. Obviously, much further work is needed to unravel and understand the entire PM degrading system of ATCC9545, to identify CBP60, and to analyze whether it plays a role in PM degradation. Identification and characterization of PlCBP49 presented here is a first and important step towards this end. Unfortunately, it was impossible to construct complementation mutants to provide the final control for the assays performed with mutant P. larvae. However, our data (i) on PM degradation during P. larvae infection [19], (ii) on the lack of classical chitinase genes in the genome of P. larvae [26], and (iii) the combination of in vitro and in vivo data presented here provide a convincing body of evidence for our interpretation of the role and relevance of PlCBP49 in AFB pathogenesis. Weakening or even destroying the larval PM allows more ready access of bacteria or bacterial toxins to gut epithelial cells [3], [39], [41]. We recently described the identification of several putative toxin genes in the genome of P. larvae ERIC I [42] and the detailed analysis of two of these toxins, Plx1 and Plx2, which belong to the family of AB-toxins [43]–[45]. Binding of P. larvae AB-toxins to their cognate cell surface receptors of the midgut epithelial cells might be accomplished more easily or even only when the epithelium is no longer protected by a PM like already shown in other insect systems [3]. We, therefore, suggest that degradation of the PM is a prerequisite for P. larvae AB-toxins Plx1 and Plx2 to act on the epithelial cells. We recently identified SplA, a P. larvae ERIC II-specific surface-layer (S-layer) protein which presumably mediates bacterial adhesion to midgut epithelial cells [46]. Direct adhesion of P. larvae ERIC II to host cells via SplA might only be possible if the cells are no longer protected by a PM. We, therefore, suggest that in the case of P. larvae ERIC II, degradation of the PM is a prerequisite for the bacteria to directly approach the epithelial layer. Based on these scenarios we propose the following model for pathogenesis of P. larvae infections: during the initial, non-invasive phase of infection, degradation of the PM serves nutritional needs and enables P. larvae to use chitin as an additional carbon source. However, at some time point the PM is totally degraded and bacteria (e.g. via SplA for P. larvae ERIC II) and/or bacterial toxins (Plx1 and Plx2 for P. larvae ERIC I) can directly approach and act on the host cells leading to bacterial breaching of the epithelium and initiation of the invasive phase. According to this model, blocking PM degradation blocks the transition from the non-invasive phase to the invasive phase explaining the impressive loss of virulence connected with the inactivation of PlCBP49 activity. This makes PlCBP49 a key virulence factor of P. larvae. Paenibacillus larvae strains ATCC9545 and DSM25430 representing the two P. larvae genotypes ERIC I and ERIC II [13], respectively, were used in this study. Strain ATCC9545 is the type reference strain and was obtained from the American Type Culture Collection (ATCC, USA) through U. Rdest (Biocenter Würzburg, Germany). Strain DSM25430 (Deutsche Sammlung von Mikroorganismen) corresponds to the field isolate 04-309 originating from an outbreak of American Foulbrood in Germany [47]. Both strains have been characterized in several previous studies [13], [42], [46], [48]–[50]. Non-manipulated P. larvae wild-type bacteria were cultivated either in MYPGP liquid broth or on Columbia sheep blood agar plates at 37°C as previously described [47], [51]. Manipulated knockout clones were cultivated on MYPGP-agar plates [52] supplemented with 5 µg/ml chloramphenicol and incubated at 37°C for 2–3 d as previously described [46]. Escherichia coli DH5α cells (Invitrogen) transformed with plasmids pTT_wsfA243 [53] or pTT_cbp573 (see below) were cultivated in selective Luria Bertani (LB) media (agar and broth) supplemented with 30 µg/ml chloramphenicol. Plasmid DNA was prepared following the manufacturer's protocols (QIAprep Spin Miniprep kit, Qiagen). Concentration and purity of the plasmid preparations were analyzed by photometric analysis (Nanodrop) and agarose gelelectrophoresis. Preparation of spore suspensions for exposure bioassays and determination of spore concentrations by cultivating serial dilutions on Columbia sheep blood agar plates were performed as described previously [13], [48], [49]. For small-scale affinity purification of P. larvae proteins exhibiting a chitin-binding domain (CBD), P. larvae was grown under the conditions described above and the culture supernatant was collected in the stationary phase by centrifugation. Aliquots (50 µl) of magnetic beads coated with chitin (chitin magnetic beads, New England Biolabs) were washed twice with CBD Column Binding Buffer (500 mM NaCl, 20 mM Tris-HCl, 1 mM EDTA, 0.05% Triton X-100; pH 8.0) at 25°C. An aliquot of filtered (2 µm-filter, Roth) P. larvae culture supernatant (500 µl) was incubated with an aliquot of pre-washed beads for 1 h at 4°C with mild agitation. Beads were magnetically collected and subsequently washed three times with CBD buffer, resuspended in non-reducing sodium dodecyl sulfate–polyacrylamide gel electrophoresis (SDS-PAGE) sample buffer [54], heated for 5 min at 96°C to release the chitin-binding proteins which were then separated by SDS–PAGE and analyzed after staining with Coomassie Brilliant Blue. Chitin-degrading activity was analyzed by zymography in 10% SDS-PA gels containing 0.1% ethylene glycol chitin (EGC) according to Trudel and Asselin [55]. Briefly, chitin-bound fractions were mixed with non-reducing SDS-PAGE sample buffer [54], heated for 5 min at 96°C to release the chitin-bound proteins and subjected to SDS-PAGE. Subsequently, gels were incubated with mild agitation at 37°C for 2 h in sodium acetate buffer (NaAc buffer; pH 5) containing 1%Triton X-100. Gels were stained for 5 min with 0.01% calcofluor in 0.5 M Tris-HCl (pH 9) and washed four times for 15 min with deionised water. Gel analysis was performed under UV-light. For analyzing the reaction mechanism of PlCBP49, zymography was also performed with NaAc buffer supplemented with 2 mM potassium cyanide (KCN), 20 mM ethylenediaminetetraacetic acid (EDTA), or 20 mM caffeine to determine the conditions inhibiting chitin-degrading activity of PlCBP49. In order to determine the protein sequence of PlCBP49I and PlCBP49II, chitin-binding fractions of ATCC9545 and DSM25430 were separated via SDS-PAGE. Coomassie stained bands of both, PlCBP49I and PlCBP49II were excised from the gel and analyzed by mass spectrometry (Alphalyse, Denmark) as already described [50]. The provided protein sequence analysis service included reduction and alkylation of cysteine residues, digestion of the proteins with trypsin, extraction and micro-purification of the obtained peptides using C18 ziptips followed by peptide mapping via MALDI-ToF and partial peptide sequencing via MALDI-ToF/ToF. Proteins were identified on the basis of peptide masses and sequence information by using the in-house databases of Alphalyse (Denmark) and the NCBI database. The obtained peptide sequences of PlCBP49I and PlCBP49II were compared to each other using the protein alignment tool of Vector NTI (Invitrogen). To identify the genetic information belonging to the sequenced peptides in P. larvae, obtained peptide fragment sequences were compared by TBLASTN analysis [28] with the sequence of P. larvae BRL 230010 [27]. The best hit was with the sequence ZP_09067740.1 annotated as chitin-binding protein. The corresponding ORF was highly homologous to a hypothetical protein of Bacillus cereus (EJQ09528.1; E-value 4e-161) and chitin-binding domain 3 protein of Bacillus thuringiensis serovar berliner (ZP_04102491.1; E-value 4e-161). In order to determine the genomic sequence of PlCBP49 in P. larvae ATCC9545 and P. larvae DSM25430, we selected a primer pair amplifying the complete predicted ORF of PlCBP49 (cbp_F and cbp_R, Table 1) and a second primer pair located up and downstream of the predicted ORF (cbp_up and cbp_down, Table 1). The obtained amplicons for both genotypes were sequenced (Eurofins, Germany) and sequences were aligned using the DNA alignment tool of Vector NTI (Invitrogen). PCR-analysis with primer pair cbp_F and cbp_R (Table 1) of several field isolates of P. larvae ERIC I and ERIC II confirmed the presence of the PlCBP49-gene in all strains analyzed so far (data not shown). The PlCBP49-gene in the genomes of P. larvae ATCC9545 and P. larvae DSM25430 was disrupted via a recently described strategy [46], [56] using vector pTT_wsfA243 containing the bacterial mobile group II intron LI.LtrB sequence [46], [53] for constructing a targetron vector for targeted intron insertion at position 573/574 from the start codon of cbp49 in P. larvae determined as optimal insertion site by a computer algorithm provided by the manufacturer (http://www.sigma-genosys.com/targetron). Retargeting of the LI.LtrB targetron of vector pTT_wsfA243 prior to transformation into P. larvae was performed following the manufacturer's protocol (Sigma) and essentially as already described for disruption of the P. larvae S-layer gene splA [46]. In brief, primers IBS_cbp_573, EBS1d_cbp_573, and EBS2_cbp_573 (Table 1) designed by a computer algorithm provided by the manufacturer (see above) were used for modification of the LI.LtrB targetron to generate a specific cbp49 targetron. Replacement of the wsfA243 targetron in pTT_wsfA243 by the cbp49 targetron gave rise to the new vector pTT_cbp573 which was subsequently transformed into E. coli DH5α cells for plasmid replication and preparation. For creation of P. larvae knockout mutants, electrocompetent P. larvae ATCC9545 and P. larvae DSM25430 cells were prepared as described [57] and 1 µg of plasmid pTT_cbp573 was transformed by electroporation as recently established [58]. Positive clones were selected on MYPGP-agar containing 5 µg/ml chloramphenicol. Successful insertion of the cbp49-specific retargeted intron (915 bp) into the P. larvae target gene cbp49 was demonstrated by PCR-analysis of the corresponding genomic regions in both knockout-strains designated P. larvae ATCC9545 Δcbp and P. larvae DSM25430 Δcbp using primers cbp_F and cbp_R flanking the cbp49 intron insertion position 573/574 of the ORF (Table 1) followed by sequencing the obtained PCR products. The mutant amplicons carrying the insertion migrated at 2259 bp whereas the wild-type amplicons had the expected size of 1344 bp (Fig. 7A). Further analyses of the mutant strains in comparison to their respective parent wild type strains were performed as already described [46] and did not reveal any differences in germination, sporulation, and growth in liquid broth (Fig. 7B,C) which otherwise might have influenced functional analyses. Clones P. larvae ATCC9545 Δcbp and P. larvae DSM25430 Δcbp were further analyzed for the absence of PlCBP49 in the chitin-binding fractions of the secretomes via SDS-PAGE and zymography. Non-infected control larvae as well as larvae infected with wild-type strain ATCC9545 or with PlCBP49 deficient strain ATCC9545 Δcbp were reared in 24-well plates as described below. Larvae at 6 d of age were immobilized for 5 min on ice. Larvae were placed under the binocular and opened longitudinally. Fat body was carefully separated from larval midgut, discarded and midgut was washed with 1XPBS (phosphate buffered saline). Clean midgut was opened lengthways and peritrophic membrane was carefully pulled out. 1 ml of 1XPBS was dispensed per well in a 24-well plate and each PM corresponding to different larvae was disposed in a different well. Afterwards, clean PMs were extended on a microscope slide, dried out and stained with a methylene blue-basic fuchsine staining technique essentially as previously described [19]. Changes in PM permeability were measured using an Ussing chamber (CHM8; World Precision Instruments, Stevenage, United Kingdom) and methylene blue (MB) as described before [59]. Spodoptera frugiperda was employed as a model insect since the required PM area necessary for performing permeability experiments could not be obtained from honey bee larvae. S. frugiperda PMs were isolated from actively feeding last-instar larvae. Each larva was anesthetized shortly on ice; midgut was extracted and longwise opened. PM was carefully pulled out, disposed on a cotton film and opened. The film was cut again to adjust PM containing area and gently washed with 1XPBS to remove food content. Cotton film was then assembled in an Ussing chamber. A suitably sized piece of the PM was used to cover the hole (12.6 mm2) in the Ussing chamber separating the two compartments. The ectoperitrophic side of the PM faced the compartment filled up with 400 µl of 0.2 mg/ml methylene blue solution, and the endoperitrophic side faced the second compartment with 400 µl of 1XPBS. Optical density at 661 nm (OD661) was measured at the beginning of the experiment and after 30 min to ensure the integrity of the PM. To test the activity of PlCBP49 against the PM, 100 µl of the buffer solution in the endoperitrophic compartment was replaced with an equivalent volume of chitin binding fraction of each P. larvae strain (ATCC9545, ATCC9545 Δcbp, DSM25430 and DSM25430 Δcbp). After additional 2 h of incubation the solutions in both compartments were recovered, and the concentration of MB was calculated based on the OD661 nm measures. The MB flux was expressed as µg of dye that passed through the 12.6-mm2 portion of mounted PM in 1 h. The flux measured at 30 min was subtracted from the final flux, to normalize the initial permeability of the PMs. Negative controls were performed with mock incubated chitin beads to rule out any influence of the beads or the buffers used on the PM. At least three independent replicates were performed for each data collection. Data represent mean values ± SEM. Data were statistically analyzed by student's t-test. In order to analyze the functional role of P. larvae PlCBP49 during pathogenesis of P. larvae infections, honeybee larvae reared in vitro were experimentally infected with the P. larvae knockout strains ATCC9545 Δcbp and DSM25430 Δcbp as well as with the corresponding P. larvae wild-type strains ATCC9545 and DSM25430. These exposure bioassays were performed essentially as previously described [20], [46], [48]. Briefly, spore suspensions with a defined concentration of colony forming units (cfu) were prepared from each of the four strains to be tested. First-instar larvae were collected from different Apis mellifera colonies of the institute's apiary. Larvae were reared in 24-well plates and larval diet (66% royal jelly (v/v), 33% glucose (w/v) and 33% fructose (w/v)) was fed ad libitum. During the first 24 h, larval diet of the infection groups was contaminated with spores to achieve infection. Subsequently, larvae were fed with normal larval diet and fresh larval diet was provided every day. Mock infected control larvae fed with normal larval diet during the whole experiment were used as internal quality control of the exposure bioassays. Only assays with less than 15% mortality in the control groups were considered valid. Spore concentrations were identical for the corresponding strains (500 cfu/ml for both ATCC9545 and ATCC9545 Δcbp; 100 cfu/ml for both DSM25430 and DSM25430 Δcbp). The final spore concentrations corresponded to the LC80 of the wild-type strains and resulted in 75.5%±5.0% and 77.8%±8.7% mortality in the larvae infected with P. larvae ATCC9545 and P. larvae DSM25430, respectively, corroborating previous findings [13], [48]. Taking the LC80 ensured that observation of both decrease and increase in mortality would be possible. Larval health status and mortality were monitored daily over 15 d. Larvae were only considered to have died from AFB if they contained a high number of vegetative P. larvae bacteria after overnight-cultivation of larval remains on CSA plates. P. larvae infection was never detected in control animals or in surviving pupae at day 15 post-infection of any of the infection groups. However, vegetative P. larvae entrapped in non-degraded PM could be demonstrated in engorged larvae of the P. larvae knockout mutant infected groups shortly before defecation. Cbp49 knockout stability (presence of the targetron insertion) in P. larvae cultivated from larval remains of the knock-out-infected groups was confirmed by PCR with the gene specific primer pair cbp_F and cbp_R (Table 1), flanking the intron insertion site 573/574. PCR amplicons were analyzed by gel electrophoresis on a 1% agarose gel, stained with ethidium bromide and visualised by UV light. Total mortality of the P. larvae wild-type strains and the P. larvae knockout mutants was calculated for each replicate as the proportion of larvae that died from AFB compared to the total number of exposed larvae. For each strain, three biological replicates were performed with thirty larvae each. Data represent mean values ± SEM. Data were statistically analyzed by student's t-test.

10.1371/journal.ppat.1000963

Cognitive Dysfunction Is Sustained after Rescue Therapy in Experimental Cerebral Malaria, and Is Reduced by Additive Antioxidant Therapy

Neurological impairments are frequently detected in children surviving cerebral malaria (CM), the most severe neurological complication of infection with Plasmodium falciparum. The pathophysiology and therapy of long lasting cognitive deficits in malaria patients after treatment of the parasitic disease is a critical area of investigation. In the present study we used several models of experimental malaria with differential features to investigate persistent cognitive damage after rescue treatment. Infection of C57BL/6 and Swiss (SW) mice with Plasmodium berghei ANKA (PbA) or a lethal strain of Plasmodium yoelii XL (PyXL), respectively, resulted in documented CM and sustained persistent cognitive damage detected by a battery of behavioral tests after cure of the acute parasitic disease with chloroquine therapy. Strikingly, cognitive impairment was still present 30 days after the initial infection. In contrast, BALB/c mice infected with PbA, C57BL6 infected with Plasmodium chabaudi chabaudi and SW infected with non lethal Plasmodium yoelii NXL (PyNXL) did not develop signs of CM, were cured of the acute parasitic infection by chloroquine, and showed no persistent cognitive impairment. Reactive oxygen species have been reported to mediate neurological injury in CM. Increased production of malondialdehyde (MDA) and conjugated dienes was detected in the brains of PbA-infected C57BL/6 mice with CM, indicating high oxidative stress. Treatment of PbA-infected C57BL/6 mice with additive antioxidants together with chloroquine at the first signs of CM prevented the development of persistent cognitive damage. These studies provide new insights into the natural history of cognitive dysfunction after rescue therapy for CM that may have clinical relevance, and may also be relevant to cerebral sequelae of sepsis and other disorders.

Cerebral malaria (CM) is a deadly consequence of Plasmodium falciparum infection. Severe neurologic deficits are frequent during CM. Although most resolve within 6 months, several retrospective studies have described high frequencies of long-lasting cognitive impairment after an episode of CM. We developed behavioral tests to identify cognitive impairment due to experimental CM. During infection with Plasmodium berghei ANKA (PbA), mice susceptible to CM (C57BL/6) developed long-lasting cognitive impairment in contextual and aversive memory. The same profile was seen in Swiss Webster mice infected with Plasmodium yoelii XL, a lethal strain that also induces neurological dysfunctions in susceptible mice strains, confirming that the cognitive dysfunction is closely associated to the development of CM. Reactive oxygen species are described as mediators of neurological and cognitive impairment associated to sepsis and Alzheimer's disease. Here we found enhanced production of malondialdeyde and conjugated dienes in brains of PbA-infected C57BL/6 mice, indicating oxidative stress. Antioxidant therapy with N-acetylcisteine and desferroxamine, as an additive to chloroquine, prevented the cognitive impairment, confirming the importance of oxidative stress in CM-associated cognitive sequellae. Administration of additive antioxidants may be a successful therapeutic strategy to control long-lasting consequences of CM and in other severe systemic inflammatory syndromes with neurological involvement.

Malaria, together with tuberculosis and human immunodeficiency virus/acquired immunodeficiency syndrome (HIV/AIDS), is one of three most important infectious diseases worldwide, with devastating morbidity and mortality and deleterious economic consequences [1]. More than 400 million people suffer from malaria, which causes over two million deaths annually, mainly among African children [2]. Cerebral malaria (CM) is the most severe neurological complication of infection with Plasmodium falciparum and is the main cause of acute non-traumatic encephalopathy in tropical countries. Mortality is high. In addition, physical and neurologic deficits are frequently seen at the time of hospital discharge in children surviving CM, although most resolve within 6 months after discharge [3]. Nevertheless, several retrospective studies suggest that cognitive deficits in children with CM are more frequent, and persist far longer than physical and neurologic deficits [4], [5], [6], [7], [8]. Boivin et al. [4] reported that 21% of children >5 years old with CM have cognitive deficits 6 months after discharge, and that increased seizure frequency and prolonged coma duration are associated with persistent cognitive deficits. Desruisseaux and coworkers [9] reported cognitive dysfunction in the acute phase of experimental infection with Plasmodium berghei ANKA in mice. A test of work memory performed at the 7th day of infection demonstrated significant impairment in visual memory in C57BL/6 mice associated to significant histological alterations as well as hemorrhage and inflammation [9]. Although the pathogenesis of CM has been extensively investigated, many aspects of the cellular and molecular pathogenesis remain incompletely defined [10]. This is in part due to the complexity of the host-pathogen interaction, which includes intricate biologic and inflammatory responses, variations in immune status and genetic background of the host, and factors unique to the malarial parasites [1]. This complexity has been revealed by clinical and experimental observations that have recently included informative mouse models [11], [12], [13]. Biochemical features also influence the natural history and complications of CM [14]. For example, there is evidence that oxidative stress mediates some of the tissue damage caused by experimental malarial infection and in cultured human cells [15], [16]. Until recently physicians have focused on survival of patients with CM and not on long-term outcomes and sequellae and, as a result, the incidence and impact of chronic neurocognitive dysfunction have been underestimated and underreported [17]. Similarly, there has been little inquiry into these issues in experimental CM. Here, we establish and characterize permissive and resistant murine models that clearly demonstrate sustained cognitive dysfunction due to CM. In addition, we demonstrate that a component of the cognitive dysfunction is related to oxidative stress and that this can be favorably modified by an interventional strategy that includes antioxidants in addition to specific rescue chemotherapy aimed at the malarial pathogen. Because oxidative stress is a pathogenetic mechanism in other syndromes of neurocognitive injury and neurodegeneration, the findings may also be relevant to other systemic inflammatory syndromes with cerebral involvement. In order to establish the clinical course of neurobehavioural complications of CM, mice from diverse genetic backgrounds were infected with different strains of Plasmodium. Mortality, parasitemia and behavior alterations (detected by the SHIRPA protocol – see below) were recorded. First, we compared C57BL/6 and BALB/c mice infected with PbA (Figure 1A, C). Ninety five percent of C57BL/6 mice died between 7 and 10 days (with an average of 7.7 days of survival, Figure 1C) after infection, with cerebral manifestations including convulsion, paralysis, and coma. Mean parasitaemia was 23%. Seventy percent of BALB/c mice died within 15 days after infection with severe anemia and overwhelming parasitemia (∼80%), but no signs of cerebral malaria. C57BL/6 mice had 40% mortality within 15 days after infection with Pch (Figure 1C). These animals showed high parasitemia (average of 46%) on day 7 and profound anemia, but no signs of CM were observed at anytime during the experiment. In this model, parasitemia at day 10 post infection was an average of 11% in surviving mice. In an additional model of infection, used to examine the effect of genetic background and parasite variables, SW mice infected with PyXL had 100% lethality (Figure 1D) within 8 days, surviving an average of 7.25 days and displaying clear signs of CM at day 6 associated with substantial parasitemia (approximately 32%). In contrast, when SW mice were infected with the non-lethal PyNXL, 100% of the animals survived for at least 15 days post infection with parasitemia over 20% and no signs of CM. These results are in agreement with others in the literature [12], [13], [18] and confirm that C57BL/6 and SW mice are susceptible to CM when infected with PbA or the lethal strain of Py, respectively. The summary results of primary screening by SHIRPA on days 3 and 6 post-infection are shown in Table 1. On day 3 post-infection, no alterations were observed in any of the groups tested (C57BL/6 versus BALB/c infected with PbA; C57BL/6 infected with PbA versus Pch; or SW mice infected with PyNXL versus PyXL) (Table 1). On day 6, however, C57BL/6 mice infected with PbA and SW infected with PyXL displayed significant alterations of reflex and sensory function, motor behavior, and autonomic function. On day 7, the animals demonstrated additional alterations of muscle tone and strength (not shown). BALB/c mice showed minor alterations of motor behavior and autonomic function (only in two tests in this protocol, while susceptible C57BL6 infected with PbA and SW mice infected with PyXL displayed positive findings in three and four assessments, respectively). Previously, Lackner et al. [19] reported that alterations of autonomic function and muscle tone and strength are specific and early signs of CM. Using these criteria, the SHIRPA protocol was prospectively applied on day 6 to identify CM. Positive results diagnosing CM were then taken as an indication to start chloroquine treatment, and to conduct further assessment of cognitive function in CM-positive animals. Interestingly, when we started treatment with chloroquine at 25 mg/kg on day 6, signs of neurological involvement were rapidly responsive and were abolished by day 7 post infection (data not shown). To investigate the occurrence of late cognitive impairment, PbA-infected C57BL/6 mice that had early signs of CM as detected by the SHIRPA protocol were treated from day 6 to 12 with chloroquine and submitted to the open field-task analysis at day 15 post infection. Chloroquine treatment was very effective in controlling parasitemia, since infected red blood cell counts were reduced to 0.66±0.6% at day 16 and parasites were not recovered at day 30 (1.1±0.56%) post infection. There were no differences in the numbers of crossings and rearings observed when groups of PbA-infected C57Bl/6 and BALB/c mice subjected to the same rescue treatment with chloroquine were studied in the training session (Figure 2). In the test session, non-infected C57BL/6 mice treated with chloroquine or saline demonstrated a significant decrease in the numbers of crossings and rearings, indicating intact cognitive skills. In contrast, there was no reduction in crossings or rearings in PbA-infected C57BL/6 mice rescued with chloroquine (Figure 2A, B; right bars), indicating diminished cognitive capacity [20]. Importantly, the decrease in cognitive ability was persistent for at least 30 days indicating a long lasting dysfunction (Figure 2C, D). In parallel, PbA-infected BALB/c mice that did not have CM based on SHIRPA analysis (Table 1), but were, nonetheless, treated with chloroquine showed a significant reduction in both crossings and rearings (Figure 2E, F, p<0.05, Student's T Test) that was not different from what was observed in non-infected controls. Thus, despite being infected with PbA, as confirmed by parasitological examinations, BALB/c mice do not develop CM and its sequelae, i.e., late cognitive impairment. Importantly, when C57BL6 mice were infected with Pch, a Plasmodium strain that does not induce CM [13] (Table 1), the pattern was similar to that of uninfected animals and CM-resistant BALB/c mice (Figure 3A, B). Therefore, even though C57BL/6 mice are susceptible to CM, when animals of this genetic background are infected with a Plasmodium strain that does not cause central nervous system involvement they do not develop signs of CM or consequent cognitive impairment based on our tools of detection. Conversely, cognitive impairment identified by our analytic instruments was not restricted to the C57BL/6 background since it was also observed in SW mice. SW mice infected with lethal strain PyXL [21] developed early signs of cerebral dysfunction that was not detected after infection with a non-lethal PyNXL strain (Table 1). SW mice infected with PyNXL showed a significant reduction in the numbers of test events when training and testing sessions were compared and the pattern was not different from non-infected control animals (Figure 3C, D). Nevertheless, when SW mice infected with PyXL were subjected to testing there was no reduction in test events in training and testing sessions (Figure 3C, D, right bars). A similar pattern was observed in PbA-infected C57BL/6 animals. Finally, we also performed experiments on PbA infected C57BL/6 animals that were depleted of CD8+ lymphocytes by treatment with anti-CD8 monoclonal antibody. CD8+ cells were previously shown to have an important role in CM [22]. In agreement with previous reports, single dose treatment with anti-CD8 temporarily reverse or stabilize the progression of CM [22], [23]. However, parasitemia and, consequently, anemia, are persistent in anti-CD8 treated mice and probably contribute to late deaths observed in these animals [22], [23]. In our hands, the first death in the anti-CD8 treated group was observed on day 13, but the majority of deaths occurred later on days 16–18. Importantly, the results from an open-field test can be altered if the mice are seriously ill, since the motor activity and general behavior are usually affected under this condition, interfering with the performance of the animals during the test. Therefore, to ensure that the results of the cognitive tests were not reflecting compromised behavior due to an ongoing severe systemic illness we decided to perform the experiments on animals that were treated both with chloroquine and anti-CD8. In fact, combined treatment with chloroquine and anti-CD8 monoclonal antibody prevented the occurrence of cognitive damage in these animals (reduction in crossings/rearings between training and testing sessions in untreated animals 34.0/32.5% versus reduction in crossings/rearings between training and testing sessions in anti-CD8 treated animals 13.0/0.0%). Together, these results indicate a clear correlation between the occurrence of CM and the development of late cognitive impairment. To determine if CM differentially influences memory skills, we submitted mice to different cognitive tasks including step-down latency and inhibitory avoidance, continuous multiple-trials step-down inhibitory avoidance and object recognition task. For this purpose, we elected to use PbA infection in C57BL/6 and BALB/c strain as positive and negative comparative models, respectively. The step-down latency and inhibitory avoidance in the test session at day 15 post infection (Figure 4) was not different from training and test in PbA-infected C57BL6 mice treated with chloroquine (mean of latency of 9 and 9.5 s, training and test sessions respectively; Z = −1.075; p = 0.282, Wilcoxon's Test), suggesting impairment in aversive memory. On the contrary, non-infected mice treated with chloroquine or saline showed an increase in step-down latency, indicating intact aversive memory, when comparing their behavior in training and test sessions. A similar pattern was seen with Pb-infected BALB/c mice, where comparisons between infected and non-infected mice were not statistically different (Figure 4). When we applied the continuous multiple trials step-down inhibitory avoidance task analysis (Figure 5), we observed a significant increase in the number of training trials required to reach the acquisition criterion (50 sec on the platform) with PbA-infected C57Bl/6 mice treated with chloroquine as compared to the non-infected controls (f(5–54) = 8; p = 0.0001, Wilcoxon's Test). The results of this task suggest that PbA-infected C57Bl/6 mice required approximately two times more stimulation to reach the acquisition criterion compared to non-infected animals receiving the same treatment, indicating learning impairment after recovery from CM [24]. As expected, PbA-infected BALB/c mice did not show any differences in the number of training trials required to reach the acquisition criterion when compared to non-infected controls. In the retention test, there was no difference between groups at all the time points tested. Therefore, learning ability, but not long term aversive memory retention skills, is impaired in PbA-infected C57BL/6 mice. PbA-infected C57Bl/6 mice treated with chloroquine showed an impairment of novel object recognition memory, i.e., they did not spend a significantly higher percentage of time exploring the novel object during short (Z = −1.782; p = 0.075, Kruskal-Wallis's Test) or long-term (Z = −1.753; p = 0.080, Kruskal-Wallis's Test) retention test sessions in comparison to the training trial (Figure 6). In contrast, this pattern was not reproduced in PbA-infected BALB/c mice (Figure 6). This result indicates that, as in other memory tasks, CM is associated with late deficits in cognition and memory skills that are not shared by infected animals that did not have clinical or neurobehavioural evidence for CM. Oxidative stress is thought to be an important mechanism in the pathogenesis of neurodegenerative diseases and in sepsis-associated encephalopathy [25], [26]. To examine this issue in experimental CM, we measured lipid peroxidation by the production of MDA, and the formation of diene conjugated species. On day 3 post infection, no significant differences in lipid peroxidation were detected in brains of C57BL/6 mice infected with PbA compared to those inoculated with control RBC (Figure 7A, C). On day 6 post infection, however, the amount of both MDA and diene conjugates (Figure 7B, D, p<0.05, Student's T Test) were increased in brain tissue from PbA-infected mice when compared to the RBC group. Conversely, C57BL/6 mice infected with Pch and BALB/c mice infected with PbA, which do not develop CM, did not show increased production of MDA (Figure 7E, F) or diene conjugates (data not shown). These data identify oxidative stress in the brains of C57BL/6 mice infected with PbA but not in non-infected controls or mice infected with Pch, a Plasmodium strain that does not cause CM, suggesting that oxidative injury is a component of neurological impairment and, potentially, cognitive dysfunction in murine CM. Taoufiq and coworkers [27] proposed that the protection of the endothelium by antioxidant delivery may constitute a relevant strategy in CM. Therefore, we asked if antioxidants used as an additive together with antimalarials therapy would reduce subsequent cognitive impairment in mice that developed early clinical signs of CM. We treated PbA-infected C57BL/6 that showed signs of CM, detected by the SHIRPA protocol, with chloroquine plus a combination of desferoxamine and N-acetylcysteine treatment starting when antimalarial treatment was initiated on day 6 post-infection and continuing for 7 days. As described previously, treatment with chloroquine alone dramatically reduced mortality and parasitemia, but did not prevent cognitive damage (Figure 2). On the other hand, treatment with desferoxamine or N-acetylcysteine alone or in combination had no effect on the parasitemia curve (data not shown). We found that the treatment with a combination of chloroquine, desferoxamine and N-acetylcysteine ameliorated cognitive impairment in infected mice. Importantly, combination of chloroquine, desferoxamine and N-acetylcysteine was equally effective in controlling parasitemia as the treatment with chloroquine alone (0.66%±0.65 in chloroquine treated animals vs 0.71%±0.49 in animals with combination treatment, ns). Figure 8 shows that there was a significant reduction in numbers of crossing and rearing events when analysis in test and training sessions of mice treated with anti-parasitic and an antioxidant drugs (p<0.05) was compared to analysis of animals given chloroquine alone. The combined administration of desferoxamine and N-acetylcysteine is a necessary condition, since when chloroquine was given with either desferoxamine or N-acetylcysteine we did not see protection against the cognitive damage (Figure 8A, B). Combination therapy was also able to abolish microvascular congestion and plugging detected by histological examinations of the cortex, hippocampus and cerebellum of treated mice (Figure 8, panels G, J and M) at day 7 post-infection, histologic features that were present in untreated mice with clinical signs of CM (Figure 8, panels F,I and L). Administration of desferoxamine plus N-acetylcysteine without chloroquine did not protect animals from early death with high parasitemia and therefore could not be tested as a treatment for cognitive impairment. The protection of cognitive function by chloroquine together with desferoxamine and N-acetylcysteine was seen both in C57BL/6 mice infected with PbA (Figure 8A, B) and SW mice infected with PyXL and (Figure 8C, D), indicating that the additive therapy with antioxidants is able to prevent cognitive impairment due to CM in relevant models of the disease and diverse genetic backgrounds. Because artesunate has become the standard therapy to treat P. falciparum malaria in humans [28], we also performed an experiment in which the animals were treated with a combination of artesunate (100 mg/kg, b.w., p.o.) plus desferoxamine and N-acetylcysteine following the same protocol described above. As seen with chloroquine, combination therapy with artesunate was able to prevent the cognitive damage observed in untreated C57BL/6 mice infected with PbA (reduction in crossings/rearings between training and testing sessions in untreated animals 14.0/13.3% versus reduction in crossings/rearings between training and testing sessions in artesunate together with deferoxamine and N-acetylcysteine treated animals 32.8/23.8%). More than 500,000 children develop CM in sub-Saharan Africa each year, of whom 110,000 die [29]. Additionally, survivors may not fully recover from CM since long-term cognitive impairment is observed in 12–14% of those individuals [6]. In a study conducted by Dugbartey and coworkers [7], children with a history of CM performed significantly poorer than those without previous CM in bimanual tactile discrimination, accuracy of visual scanning, visual memory, perceptual abstraction and rule learning skill, right ear auditory information processing, and dominant-hand motor speed. The social and economic burden of persistent cognitive dysfunction is not yet fully clear. Nevertheless, these residual deficits may affect future cognitive development in children, and this establishes the potential for devastating impact in adulthood. CM may thus be the chief cause of cognitive impairment in children in Sub-Saharan Africa and an important cause of cognitive impairment in adults in this region. Additional insights regarding the pathogenesis of cognitive deficits in CM and strategies for effective therapy to prevent this devastating complication are urgently required. The natural history of cognitive dysfunction in experimental CM and its response to rescue therapy with antimalarial are unknown. Here we addressed these issues and provide new insights that may have clinical relevance. In the present work we demonstrated cognitive damage in animals rescued from CM by treatment with the antimalarial drugs chloroquine and artesunate in the early phase of the disease. In addition, we found that antioxidant agents that have previously been used in clinical regimens reduce cognitive dysfunction when given as additive to antimalarial therapy. Experimental CM is characterized by brain edema, parenchymal lesions, blood brain barrier breakdown, and reduced cerebral blood flow. These pathophysiological responses are associated with impaired brain metabolism reflecting cellular injury and bioenergetic disturbances [30]. Magnetic resonance imaging studies suggest lesions in the corpus callosum and striatum [30]. The corpus callosum is one of the most prominent fiber systems of the mammalian brain. Patients with callosal damage cannot read text presented in the left visual field, and animals in which the callosum is divided, and sensory input restricted to one hemisphere, fail to show interhemispheric transfer of learning [31]. Taken together, these date suggest that damage in specific regions of the brain due to CM could generate cognitive damage as well as lack of memory or learning, similar to what was observed in neurocognitive impairment following CM in African children [4]. Additional studies are required to elucidate the mechanisms of central nervous system injury in children with CM as a necessary precursor to the development of interventions to prevent consequent long-term cognitive impairment [32]. We developed surrogate models that mimic clinical CM and its cognitive sequelae after parasitic cure by chloroquine, establishing invaluable tools to study mechanisms and consequences of cerebral involvement in malaria. We found that distinct cognitive abilities are affected in this condition, and that the use of antioxidant therapy concomitant with anti-malarial drugs was an effective therapy to prevent late cognitive damage to the host. In experimental malaria, infection can vary in severity depending on the species and strain of Plasmodium, the dose of parasites and the mouse genetic background. We chose our innoculum based on previous work on experimental CM in the literature [12], [19], [33], [34], [35], but we recognize the possibility that different results could have been obtained if we had used a mild infection model. In non-lethal infections, such as those caused by Pch and Py 17XNL, resolution generally results in immunity to a second challenge with the same strain, but not to a heterologous parasite. Some parasite strains are lethal only to a particular strain of mice (for instance Pch to 129sv, A/J and DBA/2 mice) and some are uniformly lethal (P. berghei ANKA, Py 17XL or YM), indicating that parasite associated factors as well as the host genetic background interact to determine lethality [13]. In the PbA model, the genetic background of the murine host is extremely important and modulates the disease outcome. For instance, the Th-1 biased C57BL/6 mouse is susceptible to the development of CM, whereas the Th-2 biased BALB/c mouse is resistant [12]. Although PbA infection is regarded as a standard model of experimental CM, there have been conflicting results using the Py 17XL parasite as a CM model. Contrary to PbA, Py 17XL has been described to induce high parasitemia, massive anemia and kidney failure without CM (for review see Engwerda et al., [11]). On the other hand, other studies report that Py 17XL induces clear signs of CM and is a useful model of this condition in the laboratory setting [13], [21], [36]. We detected high parasitemia (32%) at day 7 after Py 17XL infection and these animals exhibited signs of cerebral dysfunction when submitted to the SIRPA protocol. Because we were able to establish sensitive and reproducible methods by which CM could be unequivocally demonstrated by performing tests described in the SHIRPA protocol [19], [37], our findings are consistent with previous literature indication that Py 17XL induces important dysfunctions in the central nervous system. Based on previous studies [19], C57BL/6 mice with CM develop a wide range of behavioral and functional alterations as the syndrome progresses, and significant impairment in all functional categories when assessed 36 hours prior to death. Reflex, sensory function and neuropsychiatric state are altered in the early phase of malaria infection, and muscle tone, strength and autonomic functions are affected in animals with CM exclusively. We confirmed these findings in several models of CM. We also observed that C57BL/6 mice treated with chloroquine are rescued to basal locomotor activity when tested by the SHIRPA protocol (data not shown). Nevertheless, a cognitive deficit persists and was clearly demonstrated when the animals were subjected to specific tasks, such as the memory habituation open-field test performed 15 and 30 days after CM, indicating that cure of the parasitic infection does not prevent the development of late cognitive sequelae once CM is established. Furthermore, C57BL/6 mice infected with Pch developed clinical signs of infection but failed to develop CM and cognitive damage, indicating that cognitive impairment is not an unavoidable consequence of systemic malarial infection in C57BL/6 mice, but rather is associated with the development of clinically detected CM. We also found that severe infection without clinically established CM is not sufficient to trigger cognitive impairment using the PbA-infected BALB/c mice model. Taken together, these data document a strict correlation between development of CM and long-lasting cognitive impairment in surrogate models of malaria. It is our working hypothesis that acute cerebral malaria that leads to death in the absence of rescue therapy and long term cognitive dysfunction in animals that have been rescued with chloroquine share some of the same cellular and molecular mechanisms, and that the substrate for long term cognitive dysfunction is initiated by cerebral injury during the acute period of untreated cerebral malaria. We do not exclude, however, the possibility that long term cognitive dysfunction may also have complex mechanisms that are independent of those that cause neurologic injury and death during acute cerebral malaria, and that these undefined mechanisms only operate if the animal survives. Future investigations are aimed to clarify this point. Memory function is vulnerable to a variety of pathological process including neurodegenerative diseases, strokes, tumors, head trauma, hypoxia, cardiac surgery, malnutrition, attention-deficit disorder, depression, anxiety, medications, and normal aging [38]. One of the most elementary nonassociative learning tasks is that of behavioral habituation to a novel environment [39]. We identified deficits in memory habituation in open-field test analysis, which revealed long-term memory defect in mice with experimentally-induced CM. This deficit was unrelated to changes in basic exploratory or motor processes. Rather, it is likely to be directly related to impaired hippocampus-dependent memory processes [40], [41]. Additional target areas such as prefrontal cortex could also be involved, since reduced density of neuronal cells in this area is known to lead to orientation disturbances and memory problems in complex tasks [42], [43]. Memory habituation impairment was not the only late consequence of CM in our models as, in fact, several other cognitive deficits were documented in PbA-infected C57BL/6 mice. Step-down inhibitory avoidance learning triggers biochemical events in the hippocampus that are necessary for the retention of this task. The events are similar in many ways to those described for different types of long-term potentiation and other forms of neural plasticity [44], [45]. They are triggered by glutamate receptor activation and involve at least four different cascades led by different protein kinases (PK), including protein kinase G, PKC, calcium-calmodulin-dependent protein kinase II (CaMKII), and PKA. Several steps in these cascades have been implicated in other forms of learning that also involve the hippocampus (reviewed by Izquierdo & Medina [45]). Step-down inhibitory avoidance involves learning, acquired generally in one single trial, and long-term aversive memory retention. C57BL/6-infected mice lack long-term memory retention ability (24 hours post-stimulus) (Figure 4B) and have deficits in learning even when they are submitted to multiple trials (Figure 5A). The inhibitory avoidance task relies heavily on the dorsal hippocampus, but also depends on the entorhinal and parietal cortex and is modulated by the amygdala [44], [45]. CM may, therefore, be affecting distinct areas in the brain to interrupt different facets of memory and task performing ability. We found that object recognition is also impaired after CM. This task, originally developed by Ennaceur and Delacour [46], is based on the tendency of rodents to explore a novel object more than a familiar one. Because no rewarding or aversive stimulation is used during training, the learning occurs under conditions of relatively low stress or arousal [46]. We observed that PbA-infected C57BL/6 mice rescued from CM with chloroquine had significant impairment in novel object recognition memory compared with sham-infected mice. These findings are important since the novel object recognition task in rodents is a nonspatial, nonaversive memory test, in contrast to other tests performed in this study (habituation and aversive memories) [47]. Recognition of objects is thought to be a critical component of human declarative memory that is mainly dependent on the hippocampus. Object recognition is commonly impaired in human patients affected by neurodegenerative diseases, or who have suffered brain injury [47], [48]. We do not know if the cognitive defects are reversible, but our experiments indicate that they persist for at least 30 days after rescue from CM with chloroquine alone. Experiments are in progress to determine the duration of CM induced cognitive deficiency imposed by CM in these models. The mechanisms for cognitive impairment in CM are not completely characterized, but inflammation and vascular dysfunction appear to be the basis of cerebral involvement [9]. During malarial infection, the host and parasites are under severe oxidative stress with increased production of reactive oxygen species (ROS) and NO by activated cells in the host [14]. When produced in large amounts, ROS and nitrogen intermediates may cause damage to the host tissue including the vascular endothelium. Endothelial damage may lead to increased vascular permeability and leukocyte and platelet adherence, all seen in cerebral malaria [49]. Despite being generally accepted, this view has been challenged by observations showing that gp91phox deficient mice and inhibitors of iNOS fail to modify the development of cerebral malaria in appropriate murine models [50], [51]. We have performed preliminary experiments using NOS deficient mice and observed that those animals, despite being susceptible to high parasitemia and early death with CM symptoms, were protected from the cognitive damage if treated with chloroquine at day 6 post infection. Together, these observations may suggest that the pathology leading to mortality during CM may occur via different mechanisms than that leading to cognitive dysfunction after successful rescue therapy. In this pathophysiologic milieu, antioxidants may be an effective strategy to counteract damage in CM, and metal chelators may be of particular interest [52]. Products of lipid peroxidation are markers for oxidative stress in several diseases and experimental models [53]. To characterize the oxidative damage during early events of CM we measured TBARS and conjugated diene formation on days 3 and 6 post infection. Our findings indicate a significant increase in oxidative stress in the brains of PbA-infected mice on day 6 post-infection, further suggesting antioxidants as a potential additive therapy to reduce cerebral damage and cognitive dysfunction during CM. Oxidative stress is associated with the development of neurodegenerative diseases and is important to the development of multiple organ dysfunction syndromes during sepsis [25], [26], providing a precedent for this approaches. In fact, combined antioxidant therapy with N-acetylcysteine and desferoxamine improves survival in sepsis induced by cecal ligation and puncture (CLP) in rats by decreasing oxidative stress and limiting mitochondrial dysfunction [54]. Barichello and coworkers [55] showed that the combined therapy also prevents late memory impairment in experimental sepsis. N-acetylcysteine is a well-known thiolic antioxidant that acts as a precursor for glutathione synthesis [56]. The reducing thiol group in N-acetylcysteine also reacts directly with ROS, leading to cellular protection against oxidative damage in vitro and in vivo [57]. Desferoxamine is a powerful iron chelator that can inhibit iron dependent free radical reactions and has been shown to diminish oxidant damage in several animal model of human disease [58], [59]. Previous studies have demonstrated that desferoxamine protects against brain ischemic injury in neonatal rats when administered after an ischemia-reperfusion insults [60]. In adult rats, desferoxamine protects against focal cerebral ischemia when given as a preconditioning stimulus 72 h before the ischemic insult [61]. In agreement with the protective effect of antioxidants in sepsis-induced brain dysfunction, we found that combined treatment with N-acetylcysteine, desferoxamine and chloroquine in PbA-infected C57BL/6 mice or Swiss mice infected with PyXL prevented cognitive damage as detected by the open-field task test, further indicating a role for oxidative stress in the development of cognitive dysfunction in experimental CM and providing an approach to modify this consequence of cerebral injury. In addition, our initial experiments indicate that antioxidants are effective as additive treatment in combination with artesunate as well. Because N-acethylcysteine and desferoxamine have been used in clinical treatment of human subjects and their pharmacologic profile and side effects are known, we suggest that these drugs should be examined as additive therapy for antimalarial drugs in clinical trials to investigate their potential to decrease, or prevent, cognitive damage after CM. 6–8 weeks old C57BL/6, BALB/c (n = 10/group per experiment) and Swiss webster (SW, n = 8/group) mice from the Oswaldo Cruz Foundation breeding unit, weighing 20 to 25 g, were used for the studies. The animals were kept at constant temperature (25°C) with free access to chow and water in a room with a 12 hour light/dark cycle. The experiments in these studies were approved by the Animal Welfare Committee of the Oswaldo Cruz Foundation under license number L033/09 (CEUA/FIOCRUZ). The guidelines followed by this Committee were created by the same institution that provided ethical approval. N-acetylcysteine (Zambom Group S.p.A., Italy), desferoxamine (Novartis Bioscience S.A., Brazil), and chloroquine (Farmanguinhos, Oswaldo Cruz Foundation, Brazil) were directly dissolved in saline solution (NaCl 0.9%, w/v). The solutions were prepared immediately before use and were protected from the light before administration to the animals. Uncloned parasite lines of Plasmodium berghei ANKA, Plasmodium chabaudi chabaudi and Plasmodium yoelii were used in this study. Plasmodium berghei ANKA (PbA) parasitized red blood cells (PRBC) from BALB/c or C57BL/6 mice, Plasmodium chabaudi chabaudi (Pch) in C57BL/6 PRBC, Plasmodium yoelii non-lethal (PyNXL), and Plasmodium yoelii lethal (PyXL) in Swiss Webster PRBC donor strains were kept in liquid nitrogen and were thawed and passed into normal mice that served afterwards as parasite donors. 6–8 weeks old C57BL/6, BALB/c and Swiss webster (SW) mice were inoculated intraperitoneally with 0.2 mL suspension of 106 parasitized red blood cells (n = 8–10/group). As a control group for infection, mice were inoculated with 106 non parasitized red blood cells (RBC). Parasitaemia on days 3, 5, 7 and 10 and survival rate were recorded. On day 7 post-infection, animals of different groups (control, PbA-infected, and PbA-infected rescued with chloroquine and antioxidant; n = 3 per group) were transcardially perfused with 0.9% saline solution and 4% paraformaldehyde in PBS. The brains were carefully dissected, cryoprotected in 10%, 20%, and 30% sucrose at 4°C, and embedded in O.C.T. (Tissue-Tek) for frozen sectioning on a cryotome (Leica Microsystems). Parasagittal sections were cut at 12 µm and placed on slides for staining with haematoxylin and eosin (H&E – VETEC, Rio de Janeiro) for histological analysis by a blinded expert pathology. Sections were examined on an Axioplan light microscope (Zeiss, Germany). Mice were infected as described above. On day 3 and 6, they were subjected to SHIRPA protocol testing (see below) to identify neurobehavioral signs of CM. Animals that were positive for clinical signs of CM detected in this fashion were immediately started on chloroquine treatment (25 mg/kg b.w., orally) and were treated daily for 7 days (15 days analisis) or 21 days (30 days analysis). At day 15 post infection, the animals were subjected to a battery of behavioral tests to access cognitive function. As a control, uninfected mice received saline (when indicated) or chloroquine. An additional group of animals received additive antioxidant therapy with desferoxamine and N-acethylcisteine, (each 20 mg/kg b.w., intraperitoneally) from day 6 to 12 post infection, concomitant with chloroquine, and then were subjected to behavioral tasks on day 15. In an additional experiment, mice were infected with 106 PbA-PRBC. At day 6 mice were intraperiotoneally treated with a single dose of 0.5 mg anti-CD8 Mab obtained from Hybridomas 53-6.7 [62] and orally treated with chloroquine (25 mg/kg b.w.) during 7 days. The behavioral testing was performed according to the SHIRPA protocol [63], 1997). The primary screen was performed as described for detection of CM by Lackner and coworkers [19]. Individual tests are described in Table 1. Habituation to an open-field was carried out as described by Vianna and coworkers [39]. Animals were gently placed on an open field apparatus and allowed to explore the arena for 5 minutes (training session). 24 h later they were submitted to a similar open-field session (test session). Crossing of the black lines and rearing performed in both sessions were counted. The step-down inhibitory avoidance test was performed as described by Quevedo et al., [64]. In the training trial, animals were placed on the platform and their latency to step down on the grid with all four paws was measured with an automatic device. Immediately after stepping down on the grid, the animals received a 0.4 mA, 2.0 seconds foot shock. A retention test trial was performed 24 h after training and permanence on the grid is recorded. Continuous multiple-trials step-down inhibitory avoidance task testing was performed in the same step-down inhibitory avoidance apparatus, however, in the training session, animals were placed on the platform and immediately after stepping down on the grid, received a 0.3 mA, 2.0 seconds foot shock. 1 h 30 min later, this procedure was repeated until the mice remained on the platform for 50 seconds and the number of training trials required was recorded. On the following day the retention test was performed and the result was given by latency period on the platform, with a cut-off at 180 seconds [65], [66]. The object recognition task was carried out as described in previous studies [67]. Briefly, animals had the opportunity to explore the open field for 5 min. On the following day, a training session was conducted by placing individual mice for 5 min into the field in the center of the arena, in which two identical objects (object A1 and A2; Double Lego Toys) were positioned in two adjacent corners at 10 cm from the walls. In a short-term memory (STM) test (1.5 h after training), the mice explored the open field for 5 min in the presence of one familiar (A) and one novel (B) object. In a long-term memory (LTM) test (24 h after training), the mice explored the field for 5 min in the presence of the familiar (A) and different novel (C) object. Objects had only distinction in shape. The exploratory preference was defined as percentage of the total exploration time animal spent investigating the object A or the novel object and calculated for each animal by the ratio TB or C/(TA+TB or C) [TA = time spent exploring the familiar object A; TB or C = time spent exploring the novels objects B or C). To characterize the oxidative stress in murine brains, lipid peroxidation levels were measured by assays of thiobarbituric acid reactive species - TBARS [68] - and the formation of diene-conjugated species [69]. Brains from animals dying of CM were homogenized in cold phosphate buffer, pH 7.4 with BHT (final concentration 0.2%). Briefly, the samples (0.5 mL) were mixed with equal volume of thiobarbituric acid 0.67% (Sigma Chemical, St. Louis, MO) and then heated at 96°C for 30 min. TBARS were determined by the absorbance at 535 nm. To analyze diene-conjugate formation, lipids were extracted by partition on chloroform∶methanol (2∶1, v∶v) and the organic phase was submitted to espectrophotometric analysis at 234 nm. Results were expressed as malondialdehyde (MDA, ε = 1,56×105M−1cm−1) and diene equivalents (ε = 2,95×104M−1cm−1) per milligram of protein (BCA assay) [68]. Data were expressed as mean ± SEM. Statistical significance of survival curves were evaluated by Log-rank (Mantel-Cox) and Gehan-Breslow-Wilcoxon tests. Statistical analysis from SHIRPA was performed by nonparametric test (Wilcoxon rank-sum test). Data from the open-field task were analyzed with ANOVA followed by Tukey post hoc and Student's T tests and expressed as mean ± SEM. Data from the inhibitory avoidance task, object recognition task and the number of training trials from continuous multiple trials step-down inhibitory avoidance are reported as median and interquartile ranges and comparisons among groups were performed using Mann–Whitney U tests. The variations within individual groups were analyzed by Wilcoxon's test. Difference in amounts of MDA and diene-conjugates were evaluated by Student's T test. In all comparisons, p<0.05 or less was taken as statistical significance.

10.1371/journal.pntd.0004204

Comparison of Leishmania killicki (syn. L. tropica) and Leishmania tropica Population Structure in Maghreb by Microsatellite Typing

Leishmania (L.) killicki (syn. L. tropica), which causes cutaneous leishmaniasis in Maghreb, was recently described in this region and identified as a subpopulation of L. tropica. The present genetic analysis was conducted to explore the spatio-temporal distribution of L. killicki (syn. L. tropica) and its transmission dynamics. To better understand the evolution of this parasite, its population structure was then compared with that of L. tropica populations from Morocco. In total 198 samples including 85 L. killicki (syn. L. tropica) (from Tunisia, Algeria and Libya) and 113 L. tropica specimens (all from Morocco) were tested. Theses samples were composed of 168 Leishmania strains isolated from human skin lesions, 27 DNA samples from human skin lesion biopsies, two DNA samples from Ctenodactylus gundi bone marrow and one DNA sample from a Phlebotomus sergenti female. The sample was analyzed by using MultiLocus Enzyme Electrophoresis (MLEE) and MultiLocus Microsatellite Typing (MLMT) approaches. Analysis of the MLMT data support the hypothesis that L. killicki (syn. L. tropica) belongs to the L. tropica complex, despite its strong genetic differentiation, and that it emerged from this taxon by a founder effect. Moreover, it revealed a strong structuring in L. killicki (syn. L. tropica) between Tunisia and Algeria and within the different Tunisian regions, suggesting low dispersion of L. killicki (syn. L. tropica) in space and time. Comparison of the L. tropica (exclusively from Morocco) and L. killicki (syn. L. tropica) population structures revealed distinct genetic organizations, reflecting different epidemiological cycles.

Leishmania killicki (syn. L. tropica) was discovered in 1986. Few studies have been conducted on this parasite exclusively described in Maghreb. Consequently, many elements on its epidemiology, transmission, population structure and dynamics remain unknown. To better understand the evolution of this parasite, its population structure has been compared with that of L. tropica populations from Morocco using Multilocus Enzyme Electrophoresis (MLEE) and MultiLocus Microsatellite Typing (MLMT) typing. MLMT data support the hypothesis that L. killicki (syn. L. tropica) belongs to the L. tropica complex despite the strong genetic differentiation between them. Despite the probable recent divergence between L. killicki (syn. L. tropica) and L. tropica, they seem to evolve differently. Indeed, L. killicki (syn. L. tropica) appears slightly polymorphic and highly structured in space and time, while L. tropica was genetically heterogeneous, slightly structured geographically and temporally. The different population structures revealed distinct genetic organizations, reflecting different epidemiological cycles. Several parameters could explain these opposite epidemiological and genetic patterns such as ecosystems, vectors and reservoirs.

Leishmaniases are vector-borne diseases caused by several Leishmania species that cycle between their phlebotomine sandfly vectors and mammalian reservoir hosts [1]. Leishmania parasites, like many other microorganisms, have a high adaptation capacity that allows them to invade and survive in various ecosystems. The spread of a parasitic genotype or group of genotypes in new ecosystems can lead to population differentiation. Consequently, new Leishmania taxa have regularly been described during the last decades [2–4]. Leishmania killicki could be considered as a typical example of this evolutionary process. Rioux et al. [5] identified this parasite in the Tataouine province (South Eastern Tunisia) for the first time in 1980. Then, sporadic cases were reported in Kairouan and Sidi Bouzid (Center of Tunisia), Gafsa (South Western Tunisia) and Séliana (Northern Tunisia) [6–8]. Besides Tunisia, this taxon was described in Libya [9] and Algeria [10–12]. The probable zoonotic transmission of this parasite, with the Ctenodactylus gundi rodent as reservoir and Phlebotomus (P.) sergenti as vector, was suggested but needs to be confirmed [13–17]. Data on L. killicki are scarce and the few available studies mainly focused on the detection and identification of this taxon using isoenzymatic or genetic approaches (PCR-RFLP, PCR-sequencing and PCR-SSCP) [18–21]. The isoenzymatic characterization using the MultiLocus Enzyme Electrophoresis (MLEE) technique identified four zymodemes for L. killicki. Zymodeme MON-8 (the most frequently identified) was found in isolates from Tunisia and Libya [5, 9]; zymodemes MON-301 and MON-306 were identified in Algeria [10, 11, 18], and MON-317 was characterized in Tunisia for the first time [22]. In a recent taxonomic study, we confirmed that L. killicki is included within the L. tropica complex and we suggested calling it L. killicki (syn. L. tropica) [22]. Nevertheless, L. killicki (syn. L. tropica) epidemiology, transmission dynamics and why it is essentially described in Tunisia are still not well understood. The specific objective of this study was to provide new insights on the molecular epidemiology and transmission of L. killicki (syn. L. tropica). To this aim, we carried out a genetic study based on the analysis of nine microsatellite loci by MultiLocus Microsatellite Typing (MLMT) in a sample of 198 isolates from different Maghreb regions to explore the population structure of L. killicki (syn. L. tropica) and to compare the data with those of L. tropica populations from Morocco. In the “Materials and Methods” and “Results” sections, L. killicki has been used at the place of L. killicki (syn. L. tropica) for easy reading. A total of 198 samples were included in this study. They were composed by 154 Leishmania strains selected from the Leishmania collection of Montpellier, France (BRC-Leish, BioBank N° BB-0033-00052) and 44 samples collected by the research group of the Laboratoire de Parasitologie—Mycologie Médicale et Moléculaire (Monastir, Tunisia) during epidemiological investigations. These samples belonged to L. killicki (n = 85) and L. tropica (n = 113) and were identified over a period of 34 years (from 1980 to 2013). L. killicki samples were collected in Algeria (n = 7), Tunisia (n = 77) and Libya (n = 1). All the L. tropica strains were from Morocco since we have recently suggested that L. killicki and L. tropica from Morocco could have originated from a same L. tropica ancestor. Among the 198 samples, 168 were isolates from infected patients (Morocco [n = 113]; Tunisia [n = 47]; Algeria [n = 7]; Libya [n = 1]), 27 were DNA samples from human skin lesion biopsies (Tunisia), two were DNA samples from Ctenodactylus gundi bone marrow (Tunisia) and one was a DNA sample from a Phlebotomus sergenti female (Tunisia) (see supplementary data S1 Table). The L. killicki samples from Tunisia (n = 77) and the L. tropica samples Morocco (n = 113) were classified according to the area and period of isolation (S1 Fig). Although some of the isolates included in this study were previously characterized [5, 7, 9, 10, 18, 20, 23, 24], they were all (n = 168) analyzed again at the Centre National de Référence des Leishmanioses (CNRL), Montpellier (France) using the MLEE technique and 15 enzymatic systems, according to Rioux et al. [25]. Genomic DNA was extracted from the isolates using the QIAamp DNA Mini Kit, according to the manufacturer’s instructions, and eluted in 150 μl of AE buffer. The DNA samples from the 27 human skin biopsies, the two C. gundi and the P. sergenti were identified by polymerase chain reaction (PCR) amplification followed by digestion with BstU1 and Taq1, according to Haouas et al. [19]. The produced fragments were separated by electrophoresis on 3% agarose gels and compared with those of the WHO reference strains of L. major MON-25 (MHOM/MA/81/LEM265), L. infantum MON-1 (MHOM/FR/78/LEM75) and L. killicki MON-8 (MHOM/TN/80/LEM163). First, few randomly selected L. killicki (n = 10) and L. tropica (n = 25) strains were genotyped by amplifying the 21 microsatellite loci already used by Schwenkenbecher et al. [26] in order to select the best markers. All 21 loci could be amplified in the L. tropica samples. Conversely, only nine loci (six described by Schwenkenbechet et al. [27] and three by Jamjoom et al. [28]) were amplified in the tested L. killicki strains. These nine loci were used for genotyping the 198 samples under study (see supplementary S2 Table). All samples were amplified using the PCR conditions described by Schwenkenbecher et al. [27]: 2 min at 94°C and then 40 cycles of 94°C for 30 s, annealing temperature of each locus-specific primer set (2) for 30 s, 72°C for 1 min and a final extension step of 72°C for 10 min. The amplification products were visualized on 1.5% agarose gels. Multiplex genotyping was done using 1 μl of PCR-amplified DNA added to the Genescan 500LIZ internal size standard and 13.5μl of formamide in an automated sequencer. Genotyping data were analyzed with the Genemapper software v.4.0 to determine the fragment sizes. Fstat v. 2.9.3.2 [29], updated from Goudet [30], was used for statistical analysis of the sample genetic polymorphism based on Nei’s unbiased estimator of genetic diversity (Hs) [31], the number of alleles per locus (N) and the mean allelic richness. The same software was also used for calculating the Wright’s F statistics [32] according to the Weir and Cockerham’s method [33]. The Fst coefficient reflects the inbreeding that results from the subdivision of the population into sub-populations of limited size, and measures the genetic differentiation between sub-populations. It varies between 0 and 1; values > 0.25 reflect a high genetic differentiation [34]. Fst is considered significant when the p-value is ≤ 0.05. The Fis coefficient estimates the inbreeding of individuals due to the local non-random union of gametes in each subpopulation. Fis values range between -1 and 1. A negative value indicates an excess of heterozygotes, a positive value corresponds to heterozygote deficiency. Genotypes obtained from the concatenated sequences of the nine microsatellite loci were used to calculate the global genotypic diversity Dg (Dg = number of genotypes per population/total number of genotypes). The Neighbor-Joining (NJ) phenetic tree was constructed using the MEGA 5.10 software [35] from a Cavalli-Sforza and Edwards [36] genetic distance matrix obtained using the POPULATIONS software (http://www.legs.cnrs-gif.fr/bioinfo/populations). Leishmania strains were obtained from the Leishmania collection (BRC-Leish, Montpellier, France, BioBank N° BB-0033-00052) which is part of the French network of Biological Resources Centres for Microorganisms (FBRCMi). This parasite collection is isolated over a period of many years and is completely independent of patients from which strains were isolated. All samples taken from humans were anonymized. Isoenzymatic characterization of the L. killicki (n = 55) and L. tropica (n = 113) strains was performed to confirm (n = 166) or to identify (n = 2) their zymodemes. Ten zymodemes were obtained (three for L. killicki and seven for L. tropica). L. killicki was represented by three zymodemes. MON-8 was identified in 44 Tunisian isolates and in the Libyan sample, while MON-301 was found in the seven Algerian isolates (see S1 Table). The newly described zymodeme MON-317 was identified in three strains (MHOM/TN/2009/MET122, MHOM/TN/2010/MET300 and MHOM/TN/2010/MET301) isolated from the focus of Gafsa (South West of Tunisia) (S1 Table). L. tropica was mainly represented by the zymodeme MON-102 (n = 76), followed by MON-113 (n = 22) and MON-107 (n = 6). The four remaining zymodemes were found only in few strains: MON-109 (n = 3), MON-112 (n = 3), MON-264 (n = 1) and MON-311 (n = 2) (S1 Table). The 198 samples were amplified using primers for the nine investigated loci. Clear electropherograms and two alleles per locus and per sample were obtained (see supplementary data S3 Table). Twenty nine alleles were obtained, ranging from two for the GA1, GM2 and LIST7027 loci to five for the GA11 and LIST7036 loci (mean: 3.22 alleles per locus). The global genetic diversity was moderate (Hs = 0.261) and the global genotypic diversity was high (Dg = 0.53). The Fis values were positive at all loci and ranged from 0.120 for the LIST7040 locus to 0.920 for the GA6 locus (mean value = 0.664) (S2 Table). Analysis of the genotyping data concerning all the L. killicki samples (n = 85) revealed 22 alleles ranging from a single allele for the GA1 and LIST7027 loci to five for the GA11 locus. The mean number of alleles per locus was 2.55 and the value of the mean allelic richness was 1.23. The global genetic diversity was low (Hs = 0.185) and the global genotypic diversity was moderate (Dg = 0.38) (Table 1). Comparison of the data for the L. killicki samples from Tunisia (n = 77) and from Algeria (n = 7) indicated that their genetic diversity was low (Hs = 0.215 for the Tunisian strains and Hs = 0.15 for the Algerian strains) and that the genetic differentiation between these populations was low, but significant (Fst = 0.11, p = 0.03) (Table 2). The Hs and Fst values were not calculated for L. killicki from Libya because only one specimen was available. Moreover, estimation of the genetic differentiation between the different Tunisian populations (strains from Gafsa, Tataouine and Kairouan-Séliana) and the Algerian samples showed that the genetic differentiation was important between the populations from Tataouine and Algeria (Fst = 0.34, p = 0.005) and lower but still significant between the samples from Gafsa and Algeria (Fst = 0.09, p = 0.01). No genetic differentiation was found between the Kairouan-Séliana and Algerian populations (Fst = 0.1, p = 0.18), possibly due to the small number of specimens from Kairouan-Séliana (n = 3) (Table 2). Analysis of the genetic diversity within the different L. killicki populations from Tunisia showed that the Gafsa strains (Hs = 0.22) were more polymorphic than the Tataouine strains (Hs = 0.15). The Hs value for the Kairouan-Séliana population was certainly biased because of the low number of strains and was not considered in this analysis. Finally, the genetic differentiation between the Gafsa and Tataouine populations was also high (Fst = 0.3, p = 0.002) (Table 2). Analysis of the genetic diversity of the L. killicki samples classified based on the time of isolation indicated that specimens isolated during the 1980–1989 period were less diversified (Hs = 0.13) than those isolated between 2000 and 2009 (Hs = 0.24) or 2010 and 2013 (Hs = 0.2). The Hs value was not calculated for the 1990–1999 period because only one strain was collected during that time window. Genetic differentiation was important between the population isolated during the 1980–1989 period and the other populations. Conversely, no genetic differentiation was found between the populations collected between 2000 and 2009 and between 2010 and 2013 (Table 2). The Fst value was not estimated for the 1990–1999 window because only one strain was isolated in that period. Analysis of the genetic diversity of the L. killicki samples classified based on the region and time of isolation revealed relatively higher Hs values for the specimens collected in Gafsa at different times (Hs Gafsa [2000–2009] = 0.26, Hs Gafsa [2010–2013] = 0.28) than for those collected in Tataouine (Hs Tataouine [1980–1989] = 0.13, Hs Tataouine [2000–2009] = 0.16). The Kairouan-Séliana strains isolated at different periods and the Tataouine strains collected during the 1990–1999 period were not included in this analysis due to their limited number. Analysis of the genetic differentiation between these populations showed high Fst values that reflected temporal and geographical differences. However, a moderate genetic differentiation was found in the samples collected in Tataouine between 1980 and 1989 and between 2000 and 2009 and no genetic differentiation was observed between the strains isolated in Gafsa and Tataouine during the 2000–2009 period (Table 2). Finally, thirty-six genotypes were found. Genotype 24 was the most frequent (17.95%) in the samples from Gafsa and Tataouine (Fig 1). Analysis of the genotype distribution in each focus and according to the time of isolation showed that most genotypes were specific to a locality or to a period of isolation (see Figs 1 and 2). Twenty six alleles were identified ranging from two for the GA1 and LIST7027 loci and five for LIST7036. The mean number of alleles per locus was 2.88 and the mean allelic richness was 1.98. The global genetic diversity (Hs = 0.38) and genotypic diversity (Dg = 0.63) were high (Table 1). Genetic diversity was also high when strains were classified according to the area of isolation in Morocco (Hs Azilal = 0.34, Hs Essaouira = 0.44, Hs Ouarzaezate = 0.44, Hs Taza = 0.38). For the strains from the locality of Salé, the Hs was not estimated because of their limited number (n = 3). Genetic differentiation was mainly not observed between strains from different localities; however, few Fst values were significantly different, although they were very low (from Fst = 0.025, p = 0.035 to Fst = 0.05, p = 0.05) (Table 2). Genetic diversity was high also when the L. tropica strains were classified according to the period of isolation ([1980–1989] Hs = 0.35; [1990–1999] Hs = 0.35; [2000–2009] Hs = 0.43), whereas genetic differentiation was moderate but significant (Table 2). Comparison of the genotyping data showed strong genetic links between the L. killicki and L. tropica populations with 19 shared alleles among the 29 detected. Moreover, the NJ tree showed that L. killicki forms a monophyletic cluster inside the L. tropica complex (see Fig 3). Overall, the L. killicki population was characterized by lower genetic and genotypic diversity, fewer alleles per locus and lower allelic richness than the L. tropica population (Table 1). Analysis of the population structure showed an important genetic differentiation between the L. tropica population and the entire L. killicki sample (Fst = 0.53, p = 0.01) and also between the L. tropica population and the L. killicki populations from Tunisia [Fst = 0.53, p = 0.01] and from Algeria [Fst = 0.5, p = 0.01]) (Table 3). This result was confirmed also when the L. tropica population was compared with the L. killicki populations from the different locations in Tunisia (Gafsa, Tataouine, Kairouan Séliana) (Fst > 0.4, p < 0.05) (Table 3). Despite a great knowledge on Leishmania parasites, many taxa, such as L. killicki (syn. L. tropica), are still not completely characterized. The main objective of this study was to understand the epidemiology and transmission dynamics of L. killicki (syn. L. tropica) by analyzing its population structure and by comparing the genetic patterns of L. killicki (syn. L. tropica) and L. tropica populations in Maghreb. The comparison of L. killicki (syn. L. tropica) and L. tropica revealed a strong genetic differentiation associated with a lower genetic polymorphism within L. killicki (syn. L. tropica). Furthermore, the NJ tree showed that L. killicki (syn. L. tropica) creates a homogeneous and monophyletic cluster within the L. tropica complex. These data support the recently obtained MultiLocus Sequence Typing (MLST) results [22] suggesting that L. killicki (syn. L. tropica) emerged from L. tropica by a founder effect. The strong genetic differentiation indicates an independent evolution and an absence of gene flow between the two taxa after the founder event. The geographic distance and the ecological barriers between Morocco (area of isolation of all L. tropica specimens) and Tunisia, Libya and Algeria (regions of origin of all L. killicki (syn. L. tropica) samples) as well as the different transmission cycles can explain this diversification. Maghreb countries are essentially separated by mountains and the Sahara desert that could prevent the circulation and migration of Leishmania vectors and reservoirs. Furthermore, L. killicki (syn. L. tropica) transmission cycle is most probably zoonotic [14, 15], whereas that of L. tropica appears to be both zoonotic or anthroponotic [24]. The comparison of L. killicki (syn. L. tropica) samples from Tunisia and Algeria revealed also a differentiation within this taxon, but lower than the one detected with L. tropica. This result supports the idea that L. killicki (syn. L. tropica) spread recently and may be still spreading between the different countries after the founder event. It is not known yet where the L. tropica subpopulation emerged to generate L. killicki (syn. L. tropica), but the number of reported cases suggests Tunisia. Despite the low sample size from Algeria, we detected a strong and significant genetic differentiation between the population from Tataouine and the samples from Algeria and a low genetic differentiation between the Gafsa and Algerian populations. These results seem to indicate a more recent diversification between the Gafsa and Algerian populations, supporting the hypothesis of a recent L. killicki (syn L. tropica) dispersion from Gafsa to Algeria. Conversely, the only isolate from Libya is genetically closer to the Tataouine than to the Gafsa population. This pattern is in agreement with the geographical distances/characteristics of these regions. Indeed, the mountains in the Gafsa area, where the probable reservoir(s) of L. killicki (syn L. tropica) live(s), belong to the Atlas Mountain chains, while mountains in the Tataouine region are connected to the Libyan mountains. Concerning the L. killicki (syn. L. tropica) populations from Gafsa and Tataouine, despite their low genetic diversity indices, they show a strong and significant genetic differentiation with a lower genetic diversity among the Tataouine samples. These data suggest that the Tataouine population is more recent and that these two populations are genetically isolated. The presence of geographical barriers separating the South West and South East of Tunisia (the Sahara desert and the Chott Djerid salt lake) could explain this structuring. Analysis of the spatio-temporal evolution of L. killicki (syn. L. tropica) in Tunisia shows a low circulation of genotypes between the different populations not only in space, but also in time within a region. Based on this observation and because most isolates were from infected humans, we can hypothesize that L. killicki (syn. L. tropica) mainly circulates in the reservoir host C. gundi and humans are accidentally infected. This is in agreement with the zoonotic character of L. killicki (syn. L. tropica) compared to L. tropica, which is known to be an anthropozoonotic or zoonotic pathogen, according to the infection focus [24]. Comparison of L. tropica from Morocco and L. killicki (syn. L. tropica) from Tunisia revealed that the population structures of these two taxa are different. Indeed, L. killicki (syn. L. tropica) populations from Tunisia showed an important genetic differentiation and differences in terms of genetic diversity, whereas the L. tropica populations from Morocco were genetically more homogeneous and only slightly differentiated. These data suggest that L. killicki (syn. L. tropica) poorly disperses (except for rare migration events from a region to another) compared to L. tropica from Morocco. This finding might reflect different ecological patterns, such as epidemiological cycles, infection of the reservoirs or vector behavior. To conclude, this detailed study on L. killicki (syn. L. tropica) population genetics allowed exploring the evolutionary history of this parasite and highlighting its different genetic patterns compared to L. tropica. Despite the probable recent divergence between these taxa, they seem to evolve differently in terms of epidemiological cycle and thus transmission dynamics. Particularly, this study supports the hypothesis of a zoonotic transmission cycle for L. killicki (syn. L. tropica). Our data also suggest that Gafsa could be the historical focus of L. killicki (syn. L. tropica), although the sample size from the other regions was too small to firmly validate this hypothesis. It is now essential to study the P. sergenti vector populations in Tunisia and their susceptibility to L. killicki (syn. L. tropica) and the parasite biology in C. gundi to better understand the transmission cycle of this parasite. Although for the moment, L. killicki (syn. L. tropica) should be still considered a L. tropica subpopulation, our analyses indicate that in the future, this taxon position may have to be reconsidered.

10.1371/journal.pgen.1008332

Reshuffling yeast chromosomes with CRISPR/Cas9

Genome engineering is a powerful approach to study how chromosomal architecture impacts phenotypes. However, quantifying the fitness impact of translocations independently from the confounding effect of base substitutions has so far remained challenging. We report a novel application of the CRISPR/Cas9 technology allowing to generate with high efficiency both uniquely targeted and multiple concomitant reciprocal translocations in the yeast genome. Targeted translocations are constructed by inducing two double-strand breaks on different chromosomes and forcing the trans-chromosomal repair through homologous recombination by chimerical donor DNAs. Multiple translocations are generated from the induction of several DSBs in LTR repeated sequences and promoting repair using endogenous uncut LTR copies as template. All engineered translocations are markerless and scarless. Targeted translocations are produced at base pair resolution and can be sequentially generated one after the other. Multiple translocations result in a large diversity of karyotypes and are associated in many instances with the formation of unanticipated segmental duplications. To test the phenotypic impact of translocations, we first recapitulated in a lab strain the SSU1/ECM34 translocation providing increased sulphite resistance to wine isolates. Surprisingly, the same translocation in a laboratory strain resulted in decreased sulphite resistance. However, adding the repeated sequences that are present in the SSU1 promoter of the resistant wine strain induced sulphite resistance in the lab strain, yet to a lower level than that of the wine isolate, implying that additional polymorphisms also contribute to the phenotype. These findings illustrate the advantage brought by our technique to untangle the phenotypic impacts of structural variations from confounding effects of base substitutions. Secondly, we showed that strains with multiple translocations, even those devoid of unanticipated segmental duplications, display large phenotypic diversity in a wide range of environmental conditions, showing that simply reconfiguring chromosome architecture is sufficient to provide fitness advantages in stressful growth conditions.

Chromosomes are highly dynamic objects that often undergo large structural variations such as reciprocal translocations. Such rearrangements can have dramatic functional consequences, as they can disrupt genes, change their regulation or create novel fusion genes at their breakpoints. For instance, 90–95% of patients diagnosed with chronic myeloid leukemia carry the Philadelphia chromosome characterized by a reciprocal translocation between chromosomes 9 and 22. In addition, translocations reorganize the genetic information along chromosomes, which in turn can modify the 3D architecture of the genome and potentially affect its functioning. Quantifying the fitness impact of translocations independently from the confounding effect of base substitutions has so far remained challenging. Here, we report a novel CRISPR/Cas9-based technology allowing to generate with high efficiency and at a base-pair precision either uniquely targeted or multiple reciprocal translocations in yeast, without leaving any marker or scar in the genome. Engineering targeted reciprocal translocations allowed us for the first time to untangle the phenotypic impacts of large chromosomal rearrangements from that of point mutations. In addition, the generation of multiple translocations led to a large reorganization of the genetic information along the chromosomes, often including unanticipated large segmental duplications. We showed that reshuffling the genome resulted in the emergence of fitness advantage in stressful environmental conditions, even in strains where no gene was disrupted or amplified by the translocations.

Genetic polymorphisms are not restricted to base substitutions and indels but also include large-scale Structural Variations (SVs) of chromosomes. SVs comprise both unbalanced events, often designated as copy number variations (CNVs) including deletions and duplications, and balanced events that are copy number neutral and include inversions and translocations. Both have a phenotypic impact; however, the prevalence and the fitness effect of balanced SVs has been less documented than CNVs, partly because they are much more challenging to map than CNVs and also because quantifying their fitness contribution independently from the confounding effect of base substitutions remains challenging. Natural balanced chromosomal rearrangements result from the exchange of DNA ends during the repair of Double Strand Breaks (DSBs) either through Homologous Directed Repair (HDR) between dispersed repeats or intact chromosomes carrying internal repeat sequences homologous to the DNA ends [1,2] or through Non-Homologous End Joining (NHEJ) [3]. Artificial balanced rearrangements are classically engineered by inducing targeted DSBs and promoting repair through both HDR and NHEJ. However, inducing targeted DSBs and engineering scar-less chromosomal rearrangements has remained difficult. In early studies structural variants were obtained through recombination between heteroalleles or I-SceI-induced DSB repair between split alleles of a selection marker [4–7]. In later developments, the use of the I-SceI endonuclease was combined to a “COunter-selectable REporter” or CORE cassette in the frame of the delitto-perfetto technique, allowing the generation of a reciprocal translocation in a scar-less fashion [8]. Other techniques based on Cre/Lox recombination were used to make the genomes of Saccharomyces cerevisiae and Saccharomyces mikatae colinear and generated interspecific hybrids that produced a large proportion of viable but extensively aneuploid spores [9]. Cre/Lox recombination was also used to assess the impact of balanced rearrangements in vegetative growth and meiotic viability [10–12]. A novel approach using yeast strains with synthetic chromosomes allowed extensive genome reorganization through Cre/Lox-mediated chromosome scrambling [13–16]. This approach proved to be efficient to generate strains with a wide variety of improved metabolic capacities [15,17–19]. Muramoto and collaborators recently developed a genome restructuring technology relying on a temperature-dependent endonuclease to conditionally introduce multiple rearrangements in the genome of Arabidopsis thaliana and S. cerevisiae, thus generating strains with marked phenotypes such as increased plant biomass or ethanol production from xylose [20]. Methods using Zinc Finger Nucleases (ZFNs) and Transcription Activator-Like Effector Nucleases (TALENs) were also developed to generate targeted rearrangement in yeast, mammalian and zebrafish cells [21–24]. Although these technologies provide very useful insights, they are often difficult to implement and/or rely on the use of genetic markers. For this reason, the development of the CRISPR/Cas9 (Clustered Regularly Interspaced Short Palindromic Repeats/CRISPR associated) system has boosted the field of genome engineering [25–27]. This system, initially derived from immune systems of bacteria, consists of an endonuclease encoded by the Cas9 gene of Streptococcus pyogenes and a short RNA that guides the endonuclease at the targeted genomic locus. The gRNA can be easily designed to target any genomic locus proximal to a “NGG” Promoter Adjacent Motif (PAM). This technology is now routinely used to introduce targeted DSBs in genomes from a wide variety of species [28]. In yeast, CRISPR/Cas9 induced DSBs can be repaired with high efficiency by providing homologous repair DNA cassettes, allowing a variety of genome editions. Previous studies achieved the introduction of point mutations, single and multiple gene deletions and multiplexed genome modifications at different loci by transforming cells with plasmids bearing single or multiple gRNAs and linear DNA repair templates [29–32]. CRISPR-based approaches have also been developed to add centromeres and telomeres to chromosome fragments [33] concatenating chromosomes [34,35] and for massively parallel genome editing to generate large libraries of genetic variants [36–38]. Interestingly, it has been noticed that multiplex genome editing can often result in undesirable chromosomal translocation suggesting that such rearrangements are likely programmable [39]. The ability to generate genome rearrangements with CRISPR/Cas9 has been known for some time [40] and more recently used to engineer translocations in mammalian cells with high efficiency. The principle was to introduce two DSB in two distinct chromosome with CRISPR, then repair the DNA ends in trans by HDR with donor DNA carrying a selection marker, lost in a second step by Cre/Lox recombination leaving a single loxP element at the chromosomal junction [41]. However, unequivocally determining the contribution of SVs to phenotype variations in a large variety of environmental conditions and independently from the contribution of single nucleotide variants cannot be envisioned in mammals because of overwhelming technical and ethical difficulties. In this study, we developed two CRISPR-Cas9 genome editing strategies in yeast to generate markerless and scarless SVs with high efficiency in a control and unique genetic background, thus providing a mean to quantify their fitness impacts by high-throughput phenotyping independently from marker or background specific effects. The first strategy allowed generating on-demand any translocation at a base-pair resolution and we recapitulated the phenotypic consequences associated with known rearrangements. The second approach allows generating multiple SVs simultaneously leading to an important diversity of karyotypes. We showed that reshuffling the chromosome architecture between dispersed repeated sequences often results in the formation of large segmental duplications at translocation breakpoints. Strains with reshuffled genomes show large fitness diversity in various stress conditions, even in the absence of duplication when no gene or promoter sequence was directly disrupted by the translocations. We used a single-vector which encodes both the Cas9 nuclease gene and a gRNA expression cassette. This cassette allows cloning either a DNA fragment of 430 base-pairs reconstituting two different gRNAs in tandem or a unique 20 bp fragment corresponding to the target sequence of a single gRNA (Fig 1A). This system is versatile as it allows generating either a single targeted or multiple translocations at once. Targeted translocations can be induced by a pair of gRNAs generating two concomitant DSBs in two different chromosomes. Any pair of gRNAs can be cloned in the vector in a single ligation step using as insert either a synthetic DNA fragment or a PCR product (see Methods). The two DSBs are repaired in trans upon homologous recombination with chimerical donor DNA resulting in targeted reciprocal translocations with a single-nucleotide precision (Fig 1B). Multiple translocations are induced by generating several DSBs using a single gRNA that targets repeated sequences such as the Long Terminal Repeats (LTRs) scattered on different chromosomes. The multiple DSBs are repaired by recombination using the uncut copies of the LTRs as homologous templates (Fig 1C). Both targeted and multiple translocations are engineered in a scar-less fashion and without integrating any genetic marker in the genome. We first engineered a reciprocal translocation between two reporter genes leading to phenotypes easy to observe upon disruption. Mutation in the ADE2 gene involved in purine nucleotide biosynthesis results in the accumulation of a red pigment while mutating the CAN1 gene which encodes an arginine permease confers canavanine resistance to the cells. To generate two concomitant DSBs on chromosomes V and XV carrying CAN1 and ADE2, respectively, we cloned two previously described gRNA target sequences, namely CAN1.Y and ADE2.Y [29]. To repair the DSBs, we used donor DNA fragments of 90 base-pairs each composed of two homology regions of 45 bp identical to the sequences flanking CRISPR cutting sites (S1 Table). Two combinations of DNA repair donor fragments were used. As a control, we used donors called Point Mutation-donors (PM-donors), promoting the intra-chromosomal repair of DSBs in cis and mutating PAM sequences into stop codons, thus preventing further Cas9 activity. The donors promoting inter-chromosomal repair in trans, thus leading to a Reciprocal Translocation were called RT-donors. No point mutation needed to be introduced into the PAM sequence in this case because the translocation generates chimerical target sequences not complementary to the original gRNAs (Fig 2A). Transformation with the plasmid bearing two gRNAs and either RT or PM-donors resulted in 95% and 76% of the colonies showing both the pink and resistance to canavanine phenotypes [ade2, CANR], respectively (Material and Methods, Fig 2B). In the RT-donor experiment the other 5% of colonies were all white and sensitive to canavanine, probably resulting from the transformation of the LEU2 marker without induction of any DSB. In the PM experiment, we recovered 21% of white colonies, 83% of which being resistant to canavanine, likely resulting from a single DSB/repair in the CAN1 gene. We also recovered 3% of pink and sensitive colonies likely resulting from a single DSB/repair in the ADE2 gene (Fig 2B). We then confirmed by PCR the presence of the chimerical chromosomal junctions of 16 [ade2, CANR] strains recovered from the RT experiment. We further validated the translocation by karyotyping two [ade2, CANR] strains by PFGE. No other visible chromosomal rearrangements could be observed apart from the expected translocated chromosomes VtXV and XVtV (Fig 2C). Sanger sequencing of 250 bp around the chimerical junctions of these two strains confirmed that the translocation occurred right at the position defined by the sequence of the RT-donors with no additional mutations (S1 Fig). We next showed that our system allows to sequentially generate several targeted translocations. To illustrate this, we performed the reverse translocation to restore the WT chromosomes V and XV and functional ADE2 and CAN1 genes in the strains that carried the ADE2-CAN1 translocation (YAF190 and YAF192 in Fig 2C). We took advantage of the high instability of the Cas9 plasmid that can be easily cured from the strains (Material and Methods). We first cured the plasmid from the two translocated strains. Then, we cloned a pair of gRNAs that target the chimerical junctions formed by the ADE2-CAN1 translocation in the Cas9 plasmid and designed repair donors to restore ADE2 and CAN1 to their original configuration (Fig 2A, S1 Table). Co-transformation with the gRNAs plasmid and donor fragments resulted in 94.2% of colonies which restored the white and canavanine sensitive phenotypes [ADE2, CanS]. We performed pulse field gel electrophoresis (PFGE) karyotyping of two [ADE2, CanS] strains. No difference could be observed between the karyotype of these two strains and that of the original BY4741 strain (YAF194 and YAF199 in Fig 2C). The chromosomal junctions of the two de-translocated strains were Sanger-sequenced and were found identical to BY4741 natural junctions (S1 Fig). These results demonstrate that sequential chromosomal translocations can be engineered at base-pair resolution with high efficiency. Finally, we also showed that our system can be used to simultaneously generate a deletion of a few nucleotides and a reciprocal translocation. We used the same CAN1.Y and ADE2.Y gRNA target sequences as above and designed new donor fragments inducing deletions of 27 and 23 bp, including the PAM sequences, on chromosome V and XV, respectively (S2A Fig, S1 Table). As above, we obtained a high proportion of [ade2, CANR] transformants (96%). PCR of the junctions and karyotyping of 4 strains showed that chromosome V and XV underwent the expected reciprocal translocation (Fig 3A). The genome of one translocated strain was Nanopore sequenced and de-novo assembled (Material and Methods, S2 Table, S3 Table) to check whether off-target activity of the Cas9 nuclease could result in unexpected additional rearrangements. The translocated and reference genomes were entirely collinear except for the expected translocation and associated deletions on chromosomes V and XV (Fig 3B). No other rearrangement was observed, suggesting no major off-target activity of the Cas9 nuclease in this strain. In addition, the junction sequences are identical to the sequences of the chimerical donor fragments (S2B Fig). These experiments demonstrate that a deletion and reciprocal translocation can be concomitantly engineered at base-pair resolution with CRISPR/Cas9 in the yeast genome. It was previously reported that a reciprocal translocation between the promoters of ECM34 and SSU1, a sulphite resistance gene, created a chimerical SSU1-R allele with enhanced expression resulting in increased resistance to sulphite in the wine strain Y9J [42]. This translocation resulted from a recombination event between 4 base-pair micro-homology regions on chromosomes VIII and XVI. We therefore engineered the same translocation into the BY4741 background. We designed two gRNA target sequences as close as possible to the micro-homology regions (S3A Fig, S1 Table). The first gRNA targeted the SSU1 promoter region, 115 base-pairs upstream of the start codon. The second gRNA targeted the promoter region of ECM34, 24 base-pairs upstream of the start codon. To mimic the translocated junctions, present in the wine strains, we designed 2 double stranded synthetic DNA donors of 90 base-pairs centered on the micro-homology regions but not on the cutting sites. In addition, each donor also contained a point mutation in the PAM sequences to prevent subsequent CRISPR recognition (S3A Fig). The transformations with the Cas9 plasmid containing the two gRNAs and the donor DNA yielded on average 202 transformants. We tested natural and chimerical junction by colony PCR for 16 transformants and found the expected chimerical junctions in 15 of them. This translocation was not visible by PFGE karyotyping because the size of the translocated chromosomes were too close to the size of the original chromosomes (XVI 948 kb, VIII 563 kb, VIIItXVI 921 kb and XVItVIII 599 kb). To further validate the rearrangement, we checked the junctions by Southern blot for one translocated strain with probes flanking the two cutting sites on chromosome VIII and XVI (S3B Fig). This experiment validates the presence of the chimerical junction fragments in the rearranged strain YAF082 as compared to the WT. Finally, we sequenced the junctions of the translocated strain and found that the rearrangement occurred within the expected micro-homology regions (S3C Fig). In addition, the system was designed such that both mutated PAMs ended up into the promoter of ECM34 after the translocation, therefore unlikely to have any impact on the expression of the sulphite resistance gene SSU1 (S3C Fig). We then compared sulphite resistance between the lab strain in which we engineered the translocation (YAF202 corresponding to YAF082 cured for the Cas9 plasmid, see Methods), the non-translocated parental lab strain (BY4741) and wine isolates that carry or not the translocation of interest (Y9J and DBVPG6765, respectively). Surprisingly, we found that the engineered strain with the translocation was the least resistant of all strains with a Minimal Inhibitory Concentration (MIC) of 1 mM (Fig 4). By comparison, the reference strain and the wine isolate without the translocation both had a MIC of 2 mM. This suggests that the promoter of ECM34 is weaker than the SSU1 promoter in the BY background. As expected, the wine isolate with the translocation (Y9J) was the most resistant of all (MIC > 20 mM, Fig 4), However, it was reported that this wine strain also had four tandem repeats of a 76 bp segment in the promoter region of SSU1 originating from the ECM34 locus and it was shown that the number of repeats positively correlated with sulphite resistance [42]. This suggested that the translocation would not be per se responsible for increased resistance. Resistance would in fact result from the repeats in the ECM34 promoter region that were brought in front of the SSU1 gene by the translocation. There is only one copy of this 76 bp sequence in the BY background. To test whether promoter repeats were responsible for the resistance phenotype we PCR amplified the repeat-containing promoter of the translocated wine strain Y9J (introducing a point mutation in the PAM sequences to avoid subsequent recognition by Cas9, S1 Table) and used the PCR product as donor in a CRISPR experiment to introduce the promoter repeats in front of the SSU1 gene in the translocated BY4741 strain. We designed a gRNA targeting the region between the single copy motif and the beginning of SSU1 (S1 Table). We obtained hundreds of transformants and found that 5 out of 8 transformants tested by PCR contained the four tandem repeats and were subsequently validated by Sanger sequencing (S3D Fig). The addition of the repeats in the SSU1 promoter in the translocated lab strain (YAF158) resulted in increased sulphite resistance with a MIC of 7 mM (Fig 4) therefore confirming the effect of the repeats on the resistance phenotype. However, the chromosomal configuration and the promoter repeats are the same in YAF158 and the Y9J wine strain, yet the wine strain is much more resistant than the lab strain suggesting that additional polymorphisms must contribute to the phenotype in the wine isolate. It has been previously shown that DSBs give rise to chromosomal rearrangements when they fall in dispersed repetitive elements such as Ty retrotransposons [1]. We reasoned that generating in a single step multiple DSBs targeting Ty repeated sequences should also result in genome reshuffling through multiple translocations without altering any coding sequence or promoter region. We chose a gRNA sequence flanked by a PAM that targets 5 different Ty3 LTR copies located in chromosomes IV, VII, XV and XVI (four of which comprise a region identical to the gRNA target sequence while the fifth copy differs from the target sequence by a single mismatch at its 5' end, Fig 5A). In addition, there are 30 other complete copies of solo Ty3 LTRs dispersed throughout the genome but they contain several mismatches or indels relatively to the sequence of the gRNA and/or are devoid of PAM, suggesting that Cas9 will not cut at these sites (Fig 5A). We hypothesized that these uncut LTRs would be used as internal donor templates for DSB repair. We chose to target these five Ty3 LTRs because inducing DSB at these locations should allow significant genome reshuffling without compromising to much the viability of the cells by overloading the repair system by too many DSBs. DNA ends originating from the five DSBs can be repaired both in cis, i.e. the two ends from the same DSB are repaired together or in trans, i.e. the two ends from two different DSBs are repaired together. A WT-like karyotype is expected when all DSBs are repaired in cis and without any crossover with the internal Ty3 donor templates (type A in Fig 5B). In addition, we can predict 23 viable combinations of rearranged karyotypes when all DSBs are repaired in trans without crossover (types B to X in Fig 5B). Note that only reciprocal translocations are expected because the two targeted Ty3 LTRs on chromosome XV are in the same transcriptional orientation and therefore cannot induce the inversion of the intervening segment. In addition, all rearranged karyotypes comprising acentric or dicentric chromosomes are supposedly non-viable and thus will not be recovered. Other combinations of viable karyotypes resulting from both cis and trans repair with crossovers within uncut LTR donors could also be generated but are hard to predict because of the large number of possible LTR donors that can be used as template for repair and therefore are classified as 'unpredicted' in the following. We transformed BY4741 and BY4742 cells with the Cas9/gRNA plasmid targeting the 5 Ty3 LTRs and recovered a total of 211 and 159 transformants, respectively. We PFGE karyotyped 69 transformants (37 BY4741 and 32 BY4742) out of which 30 showed clear chromosomal rearrangements on the gels, representing 18 different karyotypes in total (Fig 5C). This result demonstrates that genomes are efficiently reshuffled via our strategy. In total 23 strains showed predicted rearrangements, representing 10 distinguishable profiles (B; D; (F,M,T); G; H; J; (K,N); R; V and X in Fig 5B and 5C). We validated the presence of all predicted junctions by colony-PCR in only 7 out of the 23 strains. For the other 16 strains, some junctions could not be validated suggesting that additional rearrangements might have occurred (see below). We sequenced all chromosomal junctions in 2 strains that show the most frequently observed rearranged profile (type J in Fig 5B and 5C) and found that all junctions, from both chimerical and un-rearranged chromosomes, were mutated in their PAM compared to the original target sequence (S4 Fig). This shows that during shuffling, all targeted sites were cut and repaired using as donor uncut Ty3 LTRs that had no PAM. Additional mutations in the region corresponding to the gRNA target sequence were also observed for 3 junctions, but were too few to identify which copy of the uncut Ty3 LTR was used as donor (S4 Fig). We tested any possible over- or under-representation of the predicted rearranged types by randomly sampling with replacement 15 draws (corresponding to the 15 characterized strains) in a uniform distribution of 18 types (24 types excluding type A (not rearranged) and types (F,M,T) and (K,N) that are not discernible). We performed 10,000 realizations and counted the number of times we sampled the different types. The probability to observe 5 times the predicted type J was 0.017 which suggests that this type could be overrepresented. No other type showed deviation from expectation. However, given the relatively small number of characterized strains, we cannot exclude that DSBs would be repaired in a random way. Moreover, we obtained 7 strains with distinct unpredicted karyotypes involving chromosomes other than the 4 targeted ones (Fig 5C). For instance, chromosomes XI and XIV that have no PAM sequence associated with their Ty3 LTRs (Fig 5A) are absent from several karyotypes showing that they can be rearranged in the absence of DSB (Fig 5C). Surprisingly, some of the karyotypes showed an apparent genome size increase suggesting the presence of unexpected large duplications (see YAF064, YAF126, YAF143 and YAF150 in Fig 5C). Using Oxford Nanopore MinION, we sequenced and de-novo assembled the genome of YAF064. We characterized in this strain an unequal reciprocal translocation between chromosomes VII and XV (S5A Fig). The junctions corresponded to the targeted CRISPR cut site on chromosome XV but not on chromosome VII. In the chimerical chromosome XVtVII, the junction occurred away from the expected site resulting in a 30 kb increase in DNA content. In the chimerical chromosome VIItXV, the junction also occurred away from the expected site but was accompanied by a truncated triplication of a 110 kb region. The missing part of the triplication corresponds to the 30 kb region found in the reciprocal chromosome XVtVII. The sequencing coverage relatively to the reference genome clearly confirmed the triplication of the complete 110 kb region (S5B Fig). Interestingly the 110 kb segment, composed of regions a and b, is flanked by two Ty3 LTRs. Moreover, two full length Ty2 retrotransposons are found, one in chromosome XV directly flanking the targeted Ty3 LTR and the other one in chromosome VII at the junction between regions a and b (S5C Fig). The presence of this triplication as well as partial junction-PCR validation for 16 strains suggested that segmental amplifications might have been missed for strain with predicted types. We Nanopore sequenced and de novo assembled the genomes of another 5 strains with different predicted profiles on the CHEF gels ((F,M, T), G, K, H, B). We characterized all rearrangements by read mapping onto the reference genomes and breakpoints were validated by at least 2 reads spanning the junctions. Only one strain showed the predicted genome organization without any additional rearrangement (YAF129, type G). The 4 other strains carried unanticipated rearrangements, ranging from the presence a single additional duplication to more complex rearrangements including multiple events (S6 Fig). All additional rearrangements were flanked by transposable elements, either full length or solo LTRs. Note that these additional rearrangements resulted in karyotypes that were hardly distinguishable from their predicted profiles (S6 Fig) showing that CHEF gels are not resolutive enough to precisely predict the chromosomal architecture of reshuffled strains. Finally, we also recovered 37 type A strains that had a WT karyotype (Fig 5B and 5C). For 35 of them we validated the presence of all un-rearranged junctions by PCR. We sequenced the junctions in 4 independent strains. We found that these strains underwent two different paths within the same experiment. Firstly, 3 clones had all junctions identical to the original LTR sequences with intact PAM, suggesting that Cas9 did not cut the target sites. Secondly, in the fourth clone, the PAM sequences of three sequenced junctions were mutated, showing that the corresponding chromosomes were cut and repaired yet without any rearrangement (S4 Fig). In this clone, one junction could not be amplified, possibly because of a small-sized indel that could not be observed on the PFGE profile. It is well known that strains bearing heterozygous translocations have impaired meiosis resulting in low spore viability [10,43–45]. We checked that crosses between rearranged and parental strains indeed produced few viable spores (Fig 5D). We tested 5 different strains with predicted rearranged karyotypes and PCR-validated junctions, therefore most probably devoid of unanticipated rearrangements. These strains were chosen because they encompassed various rearrangements including reciprocal translocations between 2 and 3 chromosomes (types D, F and G, respectively) and transpositions (types J and K). The 2 control strains without any rearrangement show 81% and 89% of viable spores with most tetrads harbouring 3 to 4 viable spores (Fig 5D). By contrast, all heterozygous diploids had a severely impaired spore viability ranging from 9% to 37% with predominantly 0, 1 or 2 viable spores per tetrad. We observed no clear correlation between the type of rearrangement and the impact on fertility. One strain carrying a single reciprocal translocation between 2 chromosomes (type F) showed the lowest viability, inferior to that of a more rearranged strain with 2 translocations between 3 chromosomes (type G, Fig 5D). These results show that knowing the type of rearrangements is not sufficient to predict their quantitative impact on meiotic fertility, although they all have a drastic negative effect. Secondly, 22 rearranged strains (the 15 predicted and 7 unpredicted karyotypes presented on Fig 5C) were phenotyped in 40 different growth conditions impacting various physiological processes and in complete synthetic media as the reference condition. These 40 conditions include various types of stresses, different drugs interfering with replication, transcription and translation as well as compounds impacting several subcellular structures (S4 Table). In total, we performed 943 phenotypic measurements (Fig 5E). We identified about twice as many cases where reshuffled strains grew significantly faster compared to cases where they grew slower than the WT strain (91 and 48 cases, respectively), suggesting that genome shuffling might be advantageous in many stressful conditions. The strongest phenotypic advantages of all corresponded to the strain with the unpredicted karyotype harboring the 110 kb triplication (YAF064, see above) when DNA synthesis is impaired (in the presence of the pyrimidine analog 5-fluorouracile) and in starvation (low carbon concentration 0.01% of galactose or glycerol, Fig 5E). However, none of 36 triplicated genes with a known function is directly involved either in DNA synthesis or starvation, suggesting that the other 18 uncharacterized genes present in this region could be involved in these phenotypes. More generally, for the conditions that produce the greatest effects, most of the strains tended to react in a similar way. For instance, in the presence of 6-azauracil (6AU), 4NQO and high glucose concentration all the strains that showed a significant phenotypic variation grew slower than the WT while in the presence of galactose, caffeine, cycloheximide and fluconazole all the strains that showed a significant variation grew faster than the WT (Fig 5E). Most strains (19 out of 22) showed variations in at least 2 different conditions showing that genome shuffling is efficiently broadening the phenotypic diversity. The most variable strain, YAF132, presented significant growth variations in 17 out of the 40 conditions (faster and slower than the WT in 15 and 2 conditions, respectively). By opposition, the 2 type J strains (YAF021 and YAF040) as well as one strain with an unpredicted karyotype (YAF135) showed no phenotypic variation in nearly all the 40 conditions. The type G strain devoid of additional rearrangement as validated by sequencing (YAF129, see above) had significant growth variations in 13 conditions, including fitness advantage in many environmental conditions. There are mounting evidences that SVs play a major role in phenotypic variation [46–48]. However, these genetic variants are the most difficult to interpret with respect to their functional consequences. In this study, we developed a versatile CRISPR/Cas9-based method allowing to engineer, with the same efficiency as point mutations, both uniquely targeted and multiple reciprocal translocations. Our method therefore provides the possibility to quantify the role played by structural variants in phenotypic diversity in a wide range of environmental conditions and in any genetic background. The versatility of our approach also allows to untangle the phenotypic impact of SVs from that of the genetic background. For instance, sulphite resistance remained lower in the laboratory as compared to the wine isolate even after engineering both the same translocation and promoter repeats at the SSU1 gene. This strongly suggests that additional polymorphisms such as SNPs also contribute to sulphite resistance in the wine isolate. For instance, there are 4 non-synonymous SNPs in the coding sequence of SSU1 between the 2 genetic backgrounds. These findings provide a striking example of the advantages brought by our technique to untangle the phenotypic impact of SVs from that of the genetic background. Reshuffling the genome with multiple translocations show that the genotypic space accessible by our approach is probably very large. The observed distribution of all predicted types, except type J, was very similar to the expected frequency of each types, suggesting that DSBs were repaired in a random way and that no specific chromosomal contact would be favored during repair. It is interesting to note that type J corresponds to the only chromosomal combination amongst all discernible predicted types where chromosome XVI remains non-rearranged (Fig 5B). The target sequence on chromosome XVI is the only one that contains a mismatch with the gRNA sequence (Fig 5A), suggesting that even a single mismatch located on the 5' end of the target sequence could be sufficient to decrease the cutting efficiency contrarily to in vitro results showing that mismatches at the last two nucleotides at the PAM distal region show similar or even higher cleavage efficiency than that of WT target [49,50]. We also isolated strains with rearranged karyotypes different from the 23 predicted profiles (Fig 5C). Translocations involving chromosomes devoid of CRISPR target site are believed to result from crossovers involving uncut LTRs (Fig 1C). The large number of LTRs that can be used for each repair event explains why these karyotypes would be difficult to predict. Moreover, we cannot exclude that some of the unpredicted karyotypes would also result from untargeted LTRs being cut by CRISPR/Cas9. However, this scenario seems very unlikely because all other LTRs are devoid of PAM and/or contain multiple mutations and indels in the target sequence (Fig 5A). Another possible explanation to the presence of unpredicted translocations would be multipartite ectopic recombination between cut and uncut chromosomes where a single end of a DSB can invade multiple sequences located on intact chromosomes during its search for homology leading to translocations between uncut chromosomes by multi-invasion-induced rearrangement [51]. Surprisingly, more complex rearrangements involving large duplications were frequently recovered, always flanked by transposable sequences, either solo LTR or full-length elements (Fig 5C, S5 Fig, S6 Fig), thereby expanding the genomic space accessible by our approach but making the prediction of accessible karyotypes less reliable. These events are consistent with previous finding showing that inducing DSB can trigger the formation of large duplicated segments [52] and that BIR can occur by several rounds of strand invasion, DNA synthesis and dissociation within dispersed repeated sequences such as LTRs, leading to chromosome rearrangements [53]. We also showed that genome reshuffling generates a great phenotypic diversity under specific environmental conditions including many cases of fitness advantage (Fig 5E). Similar findings were previously described both in S. cerevisiae [54,55] and Schizosaccharomyces pombe [10,48]. However, for the first time in this study all rearrangements are completely markerless and scarless (no CRE/Lox site) and at least for the predicted karyotype G (YAF129) no gene was disrupted nor duplicated suggesting that balanced rearrangements between repeated sequences such as Ty3 LTRs, simply reconfiguring the chromosome architecture are sufficient to create fitness diversity. Ty3 LTRs control expression of the Ty3 genomic RNA and have an organization similar to other yeast promoters. Ty3 LTRs contain positive control elements for pheromone induction, negative control elements for mating-type and also pheromone-independent transcriptional activity [56]. The transcriptional activity of Ty3 LTRs could possibly explain the observed phenotypic diversity associated with genome reshuffling at these sites. Interestingly, the presence of the nucleotide-depleting drug 6AU induces the strongest growth defect of all tested conditions (Fig 5E). Sensitivity to 6AU is a well-documented phenotype associated with transcription-elongation mutants, reducing both the elongation rate and processivity [57–60], suggesting that transcription might be globally affected in the reshuffled strains. Previous works in S. cerevisiae and in S. pombe revealed that reconfiguring the chromosome structure can alter transcription throughout the genome and not specifically at the rearrangement loci [10,11]. Further work is needed to understand the molecular mechanisms at the origin of the phenotypic diversity associated with genome reshuffling. Our genome reshuffling procedure is reminiscent of the restructuring technology developed by Muramoto and collaborators [61] that uses a temperature-dependent endonuclease to conditionally induce multiple DSBs in the genome of yeast and A. thaliana. However, here by only targeting solo LTRs our reshuffling method has some limitations in the number and types of chromosomal rearrangements that can be produced as all rearrangements that were produced were flanked by transposable element sequences. In addition, we mainly recovered translocations and duplications but deletions and inversions were almost never observed. By comparison, the Cre/Lox based SCRaMbLE technique can generate a much greater genomic diversity by concomitantly inducing recombination at thousands of sites and generating all types of variations, including small and large rearrangements encompassing a single or many genes at a time [13–16]. Despite these limitations, one advantage of the system presented here is that it can be used in any genetic background and not only in yeast strains with synthetic chromosomes. The strains of Saccharomyces cerevisiae BY4741, (MATa, his3Δ1, leu2Δ0, ura3Δ0, met15Δ0) and BY4742 (MATα, his3Δ1, leu2Δ0, ura3Δ0, lys2Δ0) were used for generating both targeted and multiple translocations. Pre-cultures were performed in YPD (yeast extract 10 g.l-1 peptone 20 g.l-1, glucose 20 g.l-1) or in SC (Yeast Nitrogen Base with ammonium sulfate 6.7 g.l−1, amino acid mixture 2 g.l−1, glucose 20 g.l−1) for high-throughput phenotyping experiments. After transformation, cells were selected on SC medium depleted in leucine. The SK1 strains SKY1513 (MATα, ho::LYS2, ura3, leu2::HISG, lys2, arg4(Nnde1)-Nsp, thr1-A, SPO11-HA3-His6::KanMX4) and SKY1708 (MATa, ho::LYS2, ura3, leu2::HISG, lys2, arg4-Bgl::NdeI-site1°, CEN8::URA3) were used for their high sporulation efficiency compared to BY strains [62] to perform crosses and to quantify spore viability of the rearranged strains. Sulphite resistance of our engineered strains (YAF202 and YAF158) was compared in liquid assay to that of wine strains DBVPG6765 (provided by Gianni Liti (IRCAN, Nice)) and Y9J_1b and to the BY4741 strain. Cells were grown overnight in YPD broth prior to inoculation at a concentration of 104 cells/mL in 8ml of YPD broth buffered at pH = 4 with tartaric acid and containing Na2SO3 concentrations ranging from 0 to 20 mM [63]. The 50 ml culture tubes were tightly closed to avoid evaporation of sulphites and incubated at 30°C. Photographs and OD600 were taken after 48 hours to quantify cell growth. Plasmid cloning steps were performed in chemically competent Escherichia coli DH5α. Ampicillin resistant bacteria were selected on LB medium supplemented with ampicillin at 100 μg/ml. For the translocation between ADE2 and CAN1, the target sequences CAN.Y and ADE2.Y found in the literature [29] were re-used. The new targets formed by the first translocation were targeted to “reverse” the translocation and restore the original junctions. For the ECM34/SSU1 translocation, specific CRISPR/Cas9 target sequences with minimal off-targets were chosen to overlap the natural recombination site of wine strains with the CRISPOR v4.3 website (http://crispor.tefor.net) using the reference genome of Saccharomyces cerevisiae (UCSC Apr. 2011 SacCer_Apr2011/sacCer3) and the NGG protospacer adjacent motif. For multiple rearrangements, we identified 39 occurrences of Ty3-LTRs in the latest version of the genome of Saccharomyces cerevisiae S288C (accession number GCF_000146045.2), which we have aligned using MUSCLE. Four sequences were incomplete and excluded from further analysis. We then manually selected a suitable gRNA sequence targeting five Ty3-LTR elements and looked for off-targets to this target sequence with CRISPOR. Predicted off-targets had either mismatches with the chosen guide and/or were devoid of PAM indicating that they would not be recognized by CRISPR/Cas9. The original plasmid pGZ110 was kindly provided by Bruce Futcher (Gang Zhao, Justin Gardin, Yuping Chen, and Bruce Futcher, Stony Brook University; personal communication). All plasmids constructed in this study were obtained by cloning in pGZ110 either a 430 bp DNA fragment reconstituting a two-gRNA expression cassette or a 20 bp fragment corresponding to the target sequence of a single gRNA (Fig 1A). We linearized pGZ110 with the enzyme LguI (ThermoFischer FD1934, isoschizomer SapI from NEB) and gel purified the backbone. In order to clone a single 20 bp target sequence we first annealed two oligonucleotides of 23 bases with 5’ overhangs of 3 bases complementary to the LguI sites. Annealing was performed by mixing equimolar amounts of forward and reverse oligonucleotides at 100 pmol/μl with NEBuffer 4 (New England Biolabs), heating 5 minutes at 95°C and allowing the mix to cool down slowly to room temperature. We then mixed 100 ng of backbone with 20 pmol of double stranded insert and performed the ligation with the Thermo Fischer Rapid DNA ligation Kit (K1422) according to manufacturer instructions. In order to obtain plasmids with an expression cassette containing two gRNAs, we ordered a 464 bp synthetic DNA fragment composed of a 430 bp sequence containing the target sequence of the first gRNA, its structural component and its terminator sequence followed by the promoter of the second gRNA and its target sequence, flanked by two LguI sites in opposite orientation (Fig 1A). This 464 bp DNA fragment was digested with LguI to obtain the 430 bp insert with adequate 5’ overhangs of 3 bases and gel-purified for ligation in pGZ110. We mixed 100 ng of backbone and 20 ng of insert and performed ligation as explained previously. As an alternative to ordering a synthetic DNA fragment, the double gRNA cassette can easily be PCR amplified using as template any plasmid already containing two gRNAs and two 55 bp oligonucleotides. The first oligo is composed of the LguI site (15 bp), the gRNA target sequence 1 (20 bp) and a homology region to the structural component of the gRNA 1 (20 bp). The second oligo is composed of the LguI site (15 bp), the gRNA target sequence 2 (20 bp) and a homology region to the promoter region of the gRNA 2 (20 bp). The resulting PCR product is then digested by LguI and cloned in the vector as previously described. All oligonucleotides and synthetic DNA fragments were ordered from Eurofins. Refer to supplementary material for oligonucleotides and synthetic DNA sequences (S1 Table). Yeast cells were transformed using the standard lithium acetate method [64] with modifications. Per transformation, 108 cells in Log phase were washed twice in 1 mL of double distilled water, then washed twice in 1 mL of lithium acetate mix (lithium acetate 0.1 M, TE) and cells were recovered in 50 μL of lithium acetate mix. To this mix were added 50 μg of denatured salmon sperm (Invitrogen), 500 ng of Cas9 plasmid, and 5 μL of double-stranded DNA donors for repair of CRISPR-induced DSBs. Double-stranded DNA donors for repair were prepared by mixing equimolar amounts of forward and reverse oligonucleotides at 100 pmol/μl with NEBuffer 4 (New England Biolabs), heating 5 minutes at 95°C and allowing the mix to cool down slowly to room temperature. Transformation were performed by adding 300 μL of lithium acetate/PEG solution (lithium acetate 0.1 M, PEG4000 45%, TE), then by vortexing for 1 minute and finally incubating the cells for 25 minutes at 42°C. After transformation, cells were plated on YPD to check for viability and on synthetic medium depleted of leucine to select for transformants. Plates were incubated 4 days at 30°C. To determine the canavanine phenotypes of the transformants, 137 and 114 pink colonies from the RT and PM experiments, respectively, were resuspended in water and spot tests realized on SC and SC-arg plus canavanine (60 mg/L) agar plates. Plates were incubated 4 days at 30°C before scoring the CANR/CANS phenotypes. Similar tests were performed for the 7 and 30 white transformants obtained from the RT and PM experiments, respectively (Fig 2B). The pGZ110 plasmid is highly unstable when selection for the LEU2 gene is removed. To cure the plasmid cell were grown overnight in YPD at 30°C. Ten individual cells were micromanipulated on YPD plates with the MSM400 micromanipulator (Singer Instruments) and grown at 30°C for 2 days. Colonies were then serially replicated on SC-Leu and YPD. All 10 replicated colonies had lost the ability to grow on CSM-Leu. Efficiencies were calculated on 3 replicates of the ADE2/CAN1 experiments. Transformation efficiency was defined as p/T with p being the average number of transformants obtained with Cas9 plasmid bearing no gRNA and without donor DNA and T being the average number of transformed cells. Cutting efficiency (%) was defined as 100*(p-g)/p with g being the average number of transformants obtained with Cas9 plasmid with the gRNAs and without donor DNA. Repair efficiency (%) was defined as 100*d/(p-g) with d being the average number of transformants obtained with Cas9 plasmid with the gRNAs and with donor DNAs (Fig 2B). Whole yeast chromosomes agarose plugs were prepared according to a standard method [65] and sealed in a 1% Seakem GTC agarose and 0.5x TBE gel. PFGE was conducted with the CHEF-DRII (BioRad) system with the following program: 6 V/cm for 10 hours with a switching time of 60 seconds followed by 6 V/cm for 17h with switching time of 90 seconds. The included angle was 120° for the whole duration of the run. We compared observed karyotypes with expected chromosome sizes and tested the chromosomal junctions by colony PCR with ThermoFischer DreamTaq DNA polymerase. Southern blot was used to validate the translocation between ECM34 (ch. VIII) and SSU1 (ch. XVI). Genomic DNA was extracted from BY4741 and the engineered strain YAF082 using the Qiagen DNA buffer set (19060) and Genomic-tip 100/G (10243) according to manufacturer instructions and further purified and concentrated by isopropanol precipitation. Digestion of 10 μg of genomic DNA per strain/probe assay was carried out using FastDigest EcoRI (ThermoFischer FD0274). Electrophoresis, denaturation and neutralization of the gel were performed according to established procedure [66]. Transfer on nylon membrane (Amersham Hybond XL) was performed using the capillarity setup [67]. The membrane was UV-crosslinked with the Stratalinker 1800 device in automatic mode. Probes targeting the genes YHL044W and ARN1 located upstream and downstream of the cutting site in the ECM34 promoter respectively and the genes NOG1 and SSU1 located upstream and downstream of the cutting site in the SSU1 promoter respectively were amplified with ThermoFischer DreamTaq, DIG-11-dUTP deoxyribonucleotides (Roche 11 175 033 910) and gel purified. Blotting and revelation were conducted using the Roche DIG High Prime DNA Labelling and Detection Starter kit II (11 585 614 910) according to manufacturer instructions. Imaging was performed using the G:BOX Chemi XT4 (Syngene) with CSPD chemiluminescence mode. All oligonucleotides are described in S1 Table. DNA from strains YAF019 and YAF064 was extracted using QIAGEN Genomic-tip 20/G columns and sheared using covaris g-TUBEs for average reads lengths of 8 kb and 15 kb respectively. DNA was repaired and dA-tailed using PreCR and FFPE kits (New England Biolabs) and cleaned with Ampure XP beads (Beckman Coulter). SQK-LSK108 adapters were ligated and libraries run on FLO-MIN107 R9.5 flowcells. Raw signals were basecalled locally using Albacore v2.0.2 with default quality filtering. DNA from strains YAF129, YAF140, YAF153, YAF155 and YAF156 was extracted using QIAGEN Genomic-tip 20/G columns, repaired and dA-tailed using FFPE and Ultra II (New England Biolabs) and cleaned with Ampure XP beads (Beckman Coulter). EXP-NBD114 barcodes 13–17 and adapters were ligated using Blunt/TA ligase MM and Quick T4 DNA ligase (New England Biolabs). The barcoded library was then run on a FLO-MIN106D R9.4.1 flowcell and locally basecalled using Guppy v2.3.5 with default filtering. Flowcell outputs are shown in S2 Table. YAF019 was assembled using the LRSDAY v1 pipeline [68], including nanopolish v0.8.5 correction and excluding pilon polishing due to lack of illumina data. Due to only 19x coverage, the correctedErrorRate for Canu assembly was increased to 0.16. YAF064 was processed using the LRSDAY v1 pipeline [68]. Linear chromosomes were assembled by SMARTdenovo v1 using 40x coverage of the longest Canu-corrected reads and combined with a Canu assembled mitochondrial genome. For canu correction, due to 200x coverage, correctedErrorRate was set at 0.12. YAF129, YAF140, YAF153, YAF155 and YAF156 reads were pre-processed by porechop and then assembled with SMARTdenovo v1 using Canu-corrected reads. Assembly data are shown in S3 Table. All corrected reads were aligned against a reference-quality S288C genome, assembled with PacBio reads [69] using LAST-921. At least two split reads were used to highlight structural variations alongside the SMARTdenovo assemblies. Read coverage was used to calculate an increase in the number of copies of particular regions within the rearranged genomes. For YAF019, evidence of rearrangements defined by overlapping reads and changes in copy number were used to manually adjust the assembly prior to nanopolish v0.8.5 and Pilon v1.22 error correction. Colonies of rearranged strains originating from the BY background and SK1 strains of the opposed mating type were mixed and spread on the same YPD plate and left overnight at 30°C. The next day, the cells were re-suspended in distilled water and single cells were picked using the Sanger MSM 400 micro-manipulator and left to grow on YPD for 2 days at 30°C until a colony appeared. Colonies originating from single cells were replicated on sporulation medium and left for 4 to 7 days at 30°C until tetrad appeared. For each strain 2 independent diploid clones were dissected on YPD plates and left to grow for 3 days before counting viable spores. For each WT strain we dissected a total of 66 tetrads and between 52 and 56 for the 5 shuffled strains. Quantitative phenotyping was performed using endpoint colony growth on solid media. First, strains were grown overnight on liquid YPD medium then pinned onto a solid SC medium with a 1,536 colony per plate density format using robot assisted pinning with ROTOR™ (Singer Instrument) and incubated overnight at 30°C. Once sufficient growth is achieved, the matrix plate is replicated onto 40 media conditions (S4 Table) plus SC as a pinning control. Plates were incubated at 30°C (except for the 14°C phenotyping). After 24h, plates were imaged at a 12Mpixel resolution. Quantification of the colony size was performed in R using Gitter [70]. Raw sizes were corrected using two successive corrections: a spatial smoothing was applied to the colony size [71]. This allowed to account for variation of the plate thickness. Another correction was then applied to rescale colony size by row and column [71] which is important for colonies lying at the edges of the plate thus having easier access to nutrients compared to strains in the center. All calculations were performed using R. Once the corrected sizes were obtained, the growth ratio of each colony was computed as the colony size on the tested conditions divided by its size on SC. To detect the phenotypic effect of the engineered translocations, each growth ratio has been normalized by the growth ratio of BY4741 or BY4742, depending on the origin of the shuffled strain, on the 40 tested condition. As each strain was present six times, the value considered for its phenotype was the median of all its replicates thus smoothing pinning heterogeneity. Each experiment was repeated 2 times independently in the 40 growth conditions. Correlations between the two replicate experiments are presented in S7 Fig. For each replicate experiment and condition, the growth ratios of the 6 colonies of each tested strain were compared to the growth ratio of the reference colonies (BY4741 or BY4742) and a Wilcoxon test was used determined the significance of the phenotypic effect: ***, ** and * indicated that the two p-values from both replicates were lower than 10−4, 10−3 and 10−2, respectively. The Oxford Nanopore sequencing data are deposited in the Sequence Read Archive under the project number PRJNA493199.

10.1371/journal.pcbi.1000387

Drug Discovery Using Chemical Systems Biology: Identification of the Protein-Ligand Binding Network To Explain the Side Effects of CETP Inhibitors

Systematic identification of protein-drug interaction networks is crucial to correlate complex modes of drug action to clinical indications. We introduce a novel computational strategy to identify protein-ligand binding profiles on a genome-wide scale and apply it to elucidating the molecular mechanisms associated with the adverse drug effects of Cholesteryl Ester Transfer Protein (CETP) inhibitors. CETP inhibitors are a new class of preventive therapies for the treatment of cardiovascular disease. However, clinical studies indicated that one CETP inhibitor, Torcetrapib, has deadly off-target effects as a result of hypertension, and hence it has been withdrawn from phase III clinical trials. We have identified a panel of off-targets for Torcetrapib and other CETP inhibitors from the human structural genome and map those targets to biological pathways via the literature. The predicted protein-ligand network is consistent with experimental results from multiple sources and reveals that the side-effect of CETP inhibitors is modulated through the combinatorial control of multiple interconnected pathways. Given that combinatorial control is a common phenomenon observed in many biological processes, our findings suggest that adverse drug effects might be minimized by fine-tuning multiple off-target interactions using single or multiple therapies. This work extends the scope of chemogenomics approaches and exemplifies the role that systems biology has in the future of drug discovery.

Both the cost to launch a new drug and the attrition rate during the late stage of the drug discovery and development process are increasing. Torcetrapib is a case in point, having been withdrawn from phase III clinical trials after 15 years of development and an estimated cost of US $800 M. Torcetrapib represents a new class of therapies for the treatment of cardiovascular disease; however, clinical studies indicated that Torcetrapib has deadly side-effects as a result of hypertension. To understand the origins of these adverse drug reactions from Torcetrapib and other related drugs undergoing clinical trials, we introduce a systematic strategy to identify off-targets in the human structural proteome and investigate the roles of these off-targets in impacting human physiology and pathology using biochemical pathway analysis. Our findings suggest that potential side-effects of a new drug can be identified at an early stage of the development cycle and be minimized by fine-tuning multiple off-target interactions. The hope is that this can reduce both the cost of drug development and the mortality rates during clinical trials.

Identification of protein-ligand interaction networks on a proteome-wide scale is crucial to address a wide range of biological problems such as correlating molecular functions to physiological processes and designing safe and efficient therapeutics [1]. Recent protein-ligand interaction studies have revealed that protein targets involved in entirely different pharmacology can bind similar small molecule drugs [2]–[4]. Large scale mapping of polypharmacology interactions indicates that drug promiscuity is a common phenomenon across the proteome [5]. It has been found that approximately 35% of known drugs or leads were active against more than one target. Moreover, a significant number of promiscuous compounds (approximately 25%) have observed activity in completely different gene families. Such drug promiscuity presents both opportunities and challenges for modern drug discovery. On one hand, it is possible to develop high-efficacy drugs by inhibiting multiple targets [6] or to reposition existing drugs to treat different diseases [7],[8]; on the other hand, the off-target effect may result in adverse drug reactions that account for around one-third of drug failures during development [9]. As a result, there is increasing interest in the identification of multiple targets associated with a phenotype [6] and in developing combinatorial therapies to boost clinical efficacy [10]. Chemogenomics has emerged as a new discipline to systematically establish target relationships based on the structural and biological similarity of their ligands [3], [11]–[18]. However, the success of chemogenomics depends on the availability of bioactivity data for the receptors and their associated ligands. For new drug targets, such data are either insufficient or unavailable. Further, the adverse drug reaction may involve receptors that are not well characterized. Complementary to chemogenomics methods, we have developed a chemical systems biology approach to identifying off-target binding networks through their ligand binding sites. The method requires 3D-structure information for the protein but not the ligand, thereby extending the scope of existing chemogenomics approaches. Moreover, the identified off-target binding network is integrated with the reconstructed biological pathways so that the effect of the drug on the biological system can be understood at the system level. In brief (see Methods for further details), our chemical systems biology approach proceeds as follows: 1) The ligand binding site of the primary target is extracted or predicted from a 3D experimental structure or homology model and characterized by a geometric potential [19]. 2) Off-target proteins with a similar ligand binding site to the primary target are identified across the human structural genome using a Sequence Order Independent Profile-Profile Alignment (SOIPPA) [20]. The atomic details of the interactions between the drug and the putative off-targets from step 2 are characterized using protein-ligand docking methods. Based on a normalized docking score the high-ranking off-targets are further investigated. 4) The identified panel of off-targets is subject to structural and functional cluster analysis and incorporated into a network that includes multiple metabolic, signal transduction, and gene regulation pathways. The first and second steps have been implemented in the software package SMAP, available from http://funsite.sdsc.edu. In this paper, we apply this strategy to identify and analyze a panel of unknown off-targets for Cholesteryl Ester Transfer Protein (CETP) inhibitors. CETP inhibitors represent a new preventive therapy for cardiovascular disease through raising HDL cholesterol. However, clinical studies have revealed that one of the CETP inhibitors, Torcetrapib, has deadly off-target effects as a result of hypertension [21]–[25] and consequently was withdrawn from phase III clinical trial. In contrast to Torcetrapib, another CETP inhibitor JTT-705 does not have unwanted side-effects that increases blood pressure [25]. In addition, JTT-705 is able to block cell proliferation and angiogenesis through Ras and P38 kinase pathways [26]. As will be shown, the multiple off-targets of these CETP inhibitors identified here are involved in both positive and negative control of stress regulation and immune response through an interconnected metabolic, signal transduction and gene regulation network. Our predictions are strongly correlated to the observed clinical and in vitro observations, providing a molecular explanation for the difference in side-effect profiles of these two CETP inhibitors. These findings suggest that adverse drug reactions might be modulated by the fine-tuning of the off-target binding network and exemplify the role of systems biology in the future of drug discovery. The ligand binding site of CETP (PDB id: 2OBD) is assumed to be a long tunnel interacting with two cholesteryl oleates (2OB) and two 1,2-dioleoyl-Sn-glycero-3-phosphocholines (PCW) molecules in the native state (Fig. S1), however, the exact location of inhibitor binding is unknown. Docking studies using the software Surflex [27], eHits [28] and AutoDock [29] indicate that the CETP inhibitors are able to bind to all four sites, with a slight preference for the pocket occupied by PCW. Thus, all four sites were used to search for the off-target binding sites of CETP inhibitors. Although only approximately 15% of human proteins have known 3D structures deposited in the Protein Data Bank (PDB) [30] , the structural coverage of the human proteome increases to 57% if homologous proteins are included (e-value less than 1.0e-3 and aligned sequence lengths greater than 30 residues using a Blast [31] search). The structural coverage is reduced to around 40% if the aligned length is greater than 120 residues (Fig. S2). After removing structures with redundant sequences (sequence identity = 100%), 5,985 structures and models from the PDB were selected for off-target search by SMAP. Besides bactericidal/permeability increasing protein (PDB Id: 1ewf) that is classified in the same fold and Pfam [32] family as CETP (FATCAT [33] p-value = 1.26e-11, RMSD = 4.53), 273 off-fold structures are found with similar binding sites to CETP (SMAP p-value less than 1.0e-3). Reverse virtual screening of the 273 structures against JTT-705, the smallest CETP inhibitor, was carried out with Surflex [27] and eHits [28] (see Methods) to detect the binding capability of these proteins. To reduce the impact of protein flexibility, the complex structure, whenever available in PDB, is used for docking. Proteins that have steric crashes with JTT-705 were removed from the list and a panel of CETP off-targets consisting of 204 structures was constructed for further study as shown in Table S1. The majority of these off-targets have binding sites that match to one of the two sites that are adjacent to PCW in CETP. Excluding cytochrome P450s that bind drugs promiscuously, most of the putative off-targets are involved in lipid/fatty acid transport or binding, signal transduction pathways and immune response. Based on both SMAP p-values and docking scores (p-value<1.0e-3, Surflex score>3.50 and eHiTs score<−4.50), six classes of structure were consistently found at the top of the list: CD1B like antigen recognition domains (CD1B); nuclear hormone receptor ligand binding domains (NR); lipid transport proteins (LPTP); fatty acid binding proteins (FABP); EF hand-like calcium binding proteins (EF); and heme binding proteins (HEME). The first four classes of proteins are able to bind cognate ligands similar to those that bind to CETP, such as fatty acids, lipoproteins, and lipids [34]. Although these putative off-targets do not have detectable global structural similarities to CETP according to their CE Z-scores (Fig. S3), they have local structural similarity and are related to each other, forming an interconnected off-target network. As shown in Fig. 1, 76% of the putative off-targets (154/204) form the three largest clusters. The largest helix bundle cluster includes NR, EF, HEME and other proteins (Fig. S4). In this paper, we focus on the six selected classes of proteins and demonstrate how they correlate to the clinical findings. Other putative off-targets are subject to on-going computational and experimental studies. Most of the predicted ligand binding sites of CD1B, LPTP, and FABP have a similar topology to that of CETP. The drug molecule binds to a cavity formed by anti-parallel beta-sheets and capped by other structural components such as a helix. The others, NR, EF, and HEME all have alpha-helical architectures that are completely different from the secondary structure surrounding the binding site of CETP. These differences illustrate the necessity of tools like SMAP that can find local structural similarities even when global similarity is non-existent. From a functional perspective, it is not surprising that lipid binding proteins act as off-targets for CETP inhibitors since they are required to bind similar cognate hydrophobic ligands such as PCW. It is noteworthy that glycolipid transfer protein, one of the lipid binding proteins, has significant structural similarity to nuclear hormone receptors. For example, the FATCAT [33] p-value is 1.77e-3 when comparing one glycolipid transfer protein (PDB id: 1TFJ) with that of retinoid X receptor (PDB id: 1YOW), but the RMSD is 9.82 Å for a rigid superimposition. However, if the components of these structures are allowed to twist, the RMSD drops to 2.57 Å when the helices surrounding the binding site are well aligned (Fig. S5). The structural similarity between glycolipid transfer protein and other all-helical proteins increases confidence in our result that the lipid-activated nuclear receptor (NR) is one of the major off-targets of CETP inhibitors. We searched for possible functional correlations between CETP and the putative off-targets using the iHOP [35] literature network (http://www.ihop-net.org/UniPub/iHOP/in?dbrefs_1=NCBI_LOCUSLINK__ID|1071). Several top-ranked off-targets appear in the same sentences with each other more than 3 times in the literature. They include phospholipid transfer proteins, nuclear receptors, including PPAR, major histocompatibility complex class II that is similar to CD1B, apolipoprotein A-1, and angiotension I converting enzyme. The functional similarity between CETP and the off-targets is further quantitatively measured using gene ontology (GO) relationships found with the FunSimMat web server [36] (http://funsimmat.bioinf.mpi-inf.mpg.de/index.php). From 204 off-targets, 148 structures had annotated GO terms and 94 structures had detectable similarities with a Resnik score [37] larger than 0.0. Among these 94 structures, lipid transport/binding proteins, CD1B, and nuclear hormone receptors were ranked top, followed by globin-like, EF hand-like and other proteins (Table S2). To further support our off-target predictions we conducted docking studies on CETP and the identified off-targets, which also provides insights into the molecular mechanisms of off-target binding. It has been established that the binding affinity calculated from docking programs is not necessarily reliable [38]–[40]. When using an energy-based scoring function, the errors come predominantly from the inaccurate parameterization of the individual energy terms. We find that the docking scores for CETP and its putative off-targets are linearly dependent on the number of carbon atoms on the docked molecules because the hydrophobic term dominates the scoring (Fig. S6). Based on this observation we developed a procedure to minimize the systematic error in the scoring function. Rather than considering the raw docking score we used the z-score to represent the relative binding affinity. The z-score is derived from a large number of random drug-like molecules and is dependent on both the number of carbon atoms in the ligand and the nature of the protein binding site. A large negative z-score indicates a high probability of true binding. Based on this procedure, the normalized docking scores (NDS) of the six classes of off-targets are listed in Table 1. These data indicate that binding of CETP inhibitors to putative off-targets is indeed statistically significant. Furthermore, the vector distance of the carbon atom size dependent average docking score for CETP and the majority of off-targets is less than 1.0 (Table S3). This implies that the ligands are able to bind to CETP and to the off-targets with similar binding affinities, since their predicted binding affinity differences are less than 1.0, which is the standard deviation of docking scores (see Methods). Finally, the correlation of ligand binding profiles between CETP and its off-targets [4] are relatively high (Table S3 and Fig. S7). Importantly, the binding profiles for the three CETP inhibitors (Torcetrapib, Anacetrapib, and JTT-705) are different from each other across the panel of off-targets. JTT-705 is the most promiscuous inhibitor. In contrast, Torcetrapib failed to dock into some of the off-targets, and Anacetrapib is suitable to be docked into the least number of off-targets. The difference between their off-target binding profiles can be partly explained by their different complexity [41] and sizes. The molecular volumes of JTT-705, Torcetrapib and Anacetrapib are 407.31, 498.42, and 527.28 Å3, respectively. As shown in Table 1 the estimated volume of the off-target binding pockets varies greatly. Thus, the smallest ligand, JTT-705, can be accommodated in all of these pockets, but the larger-sized Torcetrapib and Anacetrapib are difficult to fit into the smaller sized pockets. It could be argued that the failure in docking Torcetrapib and Anacetrapib into the smaller sized pockets is because the induce fit of the receptor is not explicitly modeled. However, for most of the NRs, both antagonist and agonist conformations are tested. Thus it is less likely that the unfitness of Torcetrapib and Anacetrapib for some of the off-targets is a result of not specifically considering induced fit in the docking calculation. The different off-target binding profiles of these CETP inhibitors have significant implications for the observed side-effects, as discussed subsequently. By incorporating the predicted off-targets into biological pathways it is possible for us to correlate the predicted off-target interactions with the observed pleotropic effects of Torcetrapib, Anacetrapib and JTT-705. Among them, the negative effect of Torcetrapib on blood pressure in phase III clinical trials could be deduced. Also deducible was an explanation for the increased death from infection and cancer [21]. Conversely, JTT-705 has gotten encouraging safety results from phase II clinical trials and no side-effects of hypertension have been observed thus far. Similar positive results are observed for Anacetrapib during phase I clinical trials. It should be noted that at this time that JTT-705 and Anacetrapib are in clinical trials involving only a small number of patients during short term studies. Results from long term studies are needed to confirm the absence of negative effects for these two drugs. In addition, JTT-705 is found to be able to block cell proliferation and angiogenesis through Ras and P38 kinase pathways [26]. To illustrate these findings, using a survey of the literature, we constructed a hierarchical biological network that connects drugs, off-targets, pathways and clinical observations. Using this network we could explore the implications of administering CETP inhibitors on different pathways through their interactions with corresponding off-targets (Fig. S8). The network consists of several interconnected metabolic, signal transduction, and gene regulation pathways. Each component of the network is separately shown in Fig. 2, Fig. 3, Fig. 4, and Fig. S9, and is discussed in detail in the following sections. It is notable that several predicted off-targets, especially the nuclear hormone receptors, are essential components in the network, involved in both positive and negative controls of several cellular systems. Nuclear hormone receptors are known as lipid-activated transcription factors that play key roles in lipid metabolism, inflammatory processes and the hormone system. The regulatory controls of our predicted nuclear hormone receptors are on pathways involved in hypertension, inflammation and cancer development. Torcetrapib, Anacetrapib and JTT705 showed different binding affinities to these receptors and thus different clinical outcomes resulting from the combinational responses of these receptors in related pathways. In vitro, in vivo and clinical studies indicate that CETP inhibitors exhibit pleotropic effects in humans through the interaction with unknown off-targets. We have identified a panel of proteins that likely bind to CETP inhibitors leading to the observed clinical indications. The putative off-target interactions are consistent with existing experimental data and provide insights into the molecular mechanisms of the side-effect profile of CETP inhibitors. Drug promiscuity depends not only on the similarity of ligand binding pockets in the related proteins but also the complexity of the drug itself [41]. In general, smaller molecules are able to bind more targets. The same trend has been predicted for CETP inhibitors; the smallest JTT-705 is the most promiscuous and the largest, Anacetrapib, is the least promiscuous. However, in contrast to conventional wisdom that implies the more specific the binding the lesser the side-effects, the most promiscuous inhibitor, JTT-705, does not cause the side-effect of hypertension that is observed in the more specific Torcetrapib. Considering the regulation of blood pressure by NRs, it is possible that JTT-705 acts as an antagonist of NRs to down-regulate aldosterone. However, our results suggest that CETP inhibitors prefer binding to the agonist rather than the antagonist conformation of the NR. Experimental evidence also implies that JTT-705 actually activates NR to mediate Ras and p38 kinase pathways [26]. Thus, it is more likely that the side-effect of CETP inhibitors is modulated by a combination of biological controls involved in many physiological processes such as cell proliferation [81], inflammation and hypertension. In other words, JTT-705 is involved in activation of NRs that contribute to both positive and negative controls of aldosterone. Although Torcetrapib is more specific and binds less off-targets than JTT-705, it only activates those NRs that up-regulate RAAS resulting in hypertension. To fully understand how small molecules can modulate physiological or pathological processes through such combinatorial control, it is necessary to simulate the dynamic properties of the biological system. To this end, it is a critical first step to identify all of the putative molecular receptors involved in the biological process and to connect them into a logical integrated protein-ligand interaction network. The chemical systems biology approach developed here is limited by available protein structures that currently only cover approximately 50% of the human proteome, although the structural coverage of the human proteome will steadily increase with progress in structural genomics [82] and conventional structure determination. As a result, some potential off-targets may be missed because they are not included in the screening. In addition to establishing functional relationships between proteins using their sequences, structures and functional sites, there are significant efforts to relate drug targets to their ligands through chemical genomics analysis [12]. However, the chemical genomics approach is restricted by the availability of bioactivity data. When exploring off-targets that cover the whole human proteome, this limitation becomes obvious since only a small number of target families explored by pharmaceutical companies are in the bioactivity database [5]. Thus our method is complementary to existing chemical genomics approaches. Drug-target networks will be greatly expanded by combining chemical genomics data and a structural genome-wide off-target analysis. Several studies have attempted to extend the target-based method to the domain-based model through similar sequence motifs or global structures [83]. In this study we further expand the scope of the chemical genomics approach beyond sequence and fold similarity by searching for similar ligand binding sites. Hence a ligand binding site-based approach will provide an ever improving way to generate a candidate list of proteins participating in interconnected biochemical pathways and to establish their relationships to biological processes. It is hoped that these approaches will eventually provide the foundation for the in silico simulation of the influence of small molecules on biological systems. In the interim it is noted that the analysis of incomplete networks is still invaluable in making new discoveries in biomedicine as exemplified by several recent studies [3],[11]. Besides SMAP used in this study, a number of web servers for ligand binding site search are available, for example, SiteEngine [84], SitesBase [85],[86], CavBase [87]–[89], SuMo [90], PdbSiteScan [91], eF-Site [92],[93], pvSOAR [94], and pevoSOAR [95]. Compared with these servers, SMAP has several distinguishing features making it particularly suitable for identifying off-targets on a structural genome-wide scale. First, SMAP does not require prior knowledge of both the location and the boundary of the ligand binding site. Instead, whole proteins are scanned to find the most similar local patch in the spirit of local sequence alignment such as the Smith-Waterman algorithm [96]. This feature makes SMAP appropriate for practical problems since typically the boundary of the ligand binding site is not clearly defined or depends on the ligand in the complex structure. Second, SMAP integrates geometric, evolutionary and physical information into a unified similarity score akin to a sequence alignment score. However, unlike conventional sequence alignment, the SMAP alignment is sequence order independent; a necessary requirement when comparing local binding sites. Third, because SMAP uses the reduced structure representation, it is not sensitive to structural uncertainty and flexibility. Thus SMAP can be applied to homology models and handle flexible ligand binding sites. Finally, we have developed a probability model to efficiently estimate the statistical significance of the binding site similarity. The model allows us to reliably identify similar ligand binding sites in a high throughput fashion. Despite these advantages of SMAP, it is expected that the best results will come from the combination of different tools as demonstrated by many studies in bioinformatics and molecular modeling. Despite the success of ligand binding search algorithms in protein function prediction and drug design [2], [4], [20], [87], [88], [95], [97]–[99] currently no algorithm can retrieve all of the binding sites that bind a cognate ligand such as ATP. However, in the context of searching for off-targets of drug molecules, the actual number of false negatives may be limited based on the nature of the drug. False negatives in the ligand binding site search are due mainly to large conformational changes of the ligand and corresponding physical and geometric changes in the binding site. Most existing drugs are designed to selectively inhibit an exquisite target. They are more rigid and less adaptable to the changing environment of the binding site than the cognate ligand. For example, a protein kinase ATP competitive inhibitor is designed to inhibit only the ATP binding site of the protein kinase, not that of other superfamilies such as P-loop hydrolases. On the other hand, although rational drug design may take the same cognate ligand binding site into account, it rarely explores the cross-reactivity between binding sites that are not naturally designed for the same cognate ligand but are able to bind the same drug. Studies by others have shown that the drug binding site can be considered as a negative image of the drug to screen compound database [100] or vice versa to model the drug binding site [101]. Hence ligand binding site similarity search is a valuable tool to identify off-targets that accommodates only the drug molecule but not necessarily all proteins that bind to the same cognate ligand across gene families. In general, the chemical systems biology approach developed in this paper is specific in identifying potential off-targets for drug-like molecules and could be used in concert with experimental design employing in vitro screening, in vivo screening and clinical trials. Even with the current limited structural coverage of the human proteome, our predications are able to provide a testable hypothesis as to the suitability of a lead compound prior to conducting a clinical trial. Thus our findings have implications for drug discovery and development. In contrast to the conventional drug discovery process in which drug leads are optimized to reduce promiscuous binding, the possible combinatorial control of aldosterone regulation by CETP inhibitors suggests that adverse drug effects can be minimized through fine tuning of multiple off-target interactions. Although it is desirable for a drug to bind the primary target in a highly specific way, this is difficult to achieve considering the inherent similarity among protein binding pockets within and across gene families. Moreover, many biological process involve combinatorial control to provide redundancy and homeostasis [102]. In such cases it becomes very difficult to modulate the systems behavior by inhibiting or activating only one single target protein. Thus, a multiple-target approach [6] and combination therapy [10] have been actively pursued to boost clinical efficacy in the treatment of diseases such as cancer and diabetes. However, these combined approaches are rarely systematic with the purposeful intent of developing therapeutics that bind to a primary target to treat the disease, but at the same time are considered to bind to desirable off-targets that modulate side-effects. In some cases this combined goal is achieved serendipitously as would seem to be the case for JTT-705. Instead of using a single molecule, it may be more feasible to use multiple components to treat a disease state and at the same time to reduce drug side-effects. Different from conventional combination therapy where all of components target disease related proteins, here only a subset of the molecules are directly therapeutic, other molecules serve the purpose of reducing side-effects by targeting non-disease related proteins. We speculate that many drugs which failed due to off-target effects can be rescued by this target-off-target combination therapy. For example, it is expected that the side-effect of Torcetrapib can be reduced by introducing molecules that binds to molecular components involved in the negative control of aldosterone regulation. Such therapies can be only rationally designed by exploring the system properties of the biological network. 5,985 structures or models that cover approximately 57% of the human proteome were searched against CETP (PDB id: 2obd) ligand binding sites using the sequence order independent profile-profile alignment (SOIPPA) algorithm [20]. A new statistical model was introduced to the original approach to estimate the significance of the alignment score [103]. In brief, the alignment score for a given alignment length is fitted to an extreme value distribution (EVD):(1)Where:(2)where S is the raw SOIPPA similarity score. μ and σ are fitted to the logarithm of N, which is the alignment length between two proteins:(3)(4)Six parameters a, b, c, d, e, and f are 5.963, −15.523, 21.690, 3.122, −9.449, and 18.252 for the McLachlan similarity matrix used in this study, respectively. Using this statistical model, 276 off-targets are identified with p-values less than 1.0e-3. The putative 276 off-targets are subject to further investigation using more computationally intensive protein-ligand docking. After removing three structures with the same fold as CETP, JTT-705, the smallest CETP inhibitor, is docked to the remaining 273 structures using two commonly used fast docking programs, Surflex 2.1 [27] (default setting) and eHits 6.2 [28] (fastest setting). 69 structures with a Surflex docking score smaller than 0.0 or an eHits score larger than 0.0 are considered to be difficult to fit JTT-705 due to significant steric crashes (and hence the other two inhibitors based on size) and are removed from the putative off-target list. The remaining 204 structures are subject to further investigation using the docking software AutoDock4.0 [29] and other more computationally intense methods as described below. An all-against-all global structural similarity analysis between the 204 putative off-targets was computed using CE [104]. A graph is constructed with each of the structures as a node. An edge is formed between two nodes if their CE z-score is larger than 4.0 (a superfamily level similarity) [104]. The volume of the binding pocket is computed using the CASTp server [105] (http://sts-fw.bioengr.uic.edu/castp) with default settings. Drug-like molecules are downloaded from ZINC (http://zinc.docking.org) [106]. From this database, six sets of molecules are randomly selected with a fixed number, 5, 10, 15, 20, 25 and 29 carbon atoms, respectively; each set includes 100 molecules. These molecules are docked to CETP and its putative off-targets using eHiTs [28] and AutoDock4.0 [29]. The correlation of the docking score to the number of carbon atoms is derived from linear regression for each of the protein receptors. From the linear fitting curve, the average docking score for molecules with a certain number of carbon atoms can be estimated. Based on the fitted average docking score, a normalized docking score DS is calculated as a z-score:(5)Where Si is the raw docking score for the molecule with i carbon atoms, μi is the fitted average docking score for the number of carbon atoms i, σ is the standard deviation, which is not dependent on the size of molecules and is approximately 1.0 in all cases. The vector distance of the average docking score D between CETP and its off-targets is calculated from the average values of the docking scores for randomly selected molecules with fixed numbers of 5, 10, 15, 20, 25 and 29 carbon atoms as follows:(6)where SCETP and Soff are the average values of carbon atom size dependent docking scores to CETP and its off-targets, respectively. In this case study, we identify a panel of off-targets of CETP inhibitors using a chemical systems biology approach. All of the identified off-targets belong to different protein superfamilies from the primary target, but are structurally and functionally related, being mainly involved in lipid metabolism, immune response and signaling networks. Among them, CD1, nuclear hormone receptors and lipid transport proteins are the most likely off-targets with highly consistent results from multiple resources including functional correlation, ligand binding site similarity, hydrophobic scales, and predicted binding affinities. Moreover, the elucidated off-target effects from these proteins are strongly correlated to clinical and in vitro observations. Their combinatorial control of biological process plays a key role in the modulation of the adverse drug effect of CETP inhibitors. This study demonstrates that a chemical systems biology approach, which systematically explores protein-ligand interactions on a genome-wide scale and incorporates them into biological pathways, will provide us with valuable clues as to the molecular basis of cellular function. At the same time, it will help to transform the conventional single-target-single-drug drug discovery process to a new multi-target-multi-molecule paradigm.

10.1371/journal.pntd.0005914

Detecting and confirming residual hotspots of lymphatic filariasis transmission in American Samoa 8 years after stopping mass drug administration

The Global Programme to Eliminate Lymphatic Filariasis (LF) aims to eliminate the disease as a public health problem by 2020 by conducting mass drug administration (MDA) and controlling morbidity. Once elimination targets have been reached, surveillance is critical for ensuring that programmatic gains are sustained, and challenges include timely identification of residual areas of transmission. WHO guidelines encourage cost-efficient surveillance, such as integration with other population-based surveys. In American Samoa, where LF is caused by Wuchereria bancrofti, and Aedes polynesiensis is the main vector, the LF elimination program has made significant progress. Seven rounds of MDA (albendazole and diethycarbamazine) were completed from 2000 to 2006, and Transmission Assessment Surveys were passed in 2010/2011 and 2015. However, a seroprevalence study using an adult serum bank collected in 2010 detected two potential residual foci of transmission, with Og4C3 antigen (Ag) prevalence of 30.8% and 15.6%. We conducted a follow up study in 2014 to verify if transmission was truly occurring by comparing seroprevalence between residents of suspected hotspots and residents of other villages. In adults from non-hotspot villages (N = 602), seroprevalence of Ag (ICT or Og4C3), Bm14 antibody (Ab) and Wb123 Ab were 1.2% (95% CI 0.6–2.6%), 9.6% (95% CI 7.5%-12.3%), and 10.5% (95% CI 7.6–14.3%), respectively. Comparatively, adult residents of Fagali’i (N = 38) had significantly higher seroprevalence of Ag (26.9%, 95% CI 17.3–39.4%), Bm14 Ab (43.4%, 95% CI 32.4–55.0%), and Wb123 Ab 55.2% (95% CI 39.6–69.8%). Adult residents of Ili’ili/Vaitogi/Futiga (N = 113) also had higher prevalence of Ag and Ab, but differences were not statistically significant. The presence of transmission was demonstrated by 1.1% Ag prevalence (95% CI 0.2% to 3.1%) in 283 children aged 7–13 years who lived in one of the suspected hotspots; and microfilaraemia in four individuals, all of whom lived in the suspected hotspots, including a 9 year old child. Our results provide field evidence that integrating LF surveillance with other surveys is effective and feasible for identifying potential hotspots, and conducting surveillance at worksites provides an efficient method of sampling large populations of adults.

Lymphatic filariasis (LF) is caused by infection with filarial worms that are transmitted by mosquito bites. The Global Programme to Eliminate Lymphatic Filariasis aims to eliminate the disease as a public health problem by 2020. Once elimination targets have been reached, cost-effective surveillance strategies are required to ensure that any areas of ongoing transmission or resurgence are quickly identified and managed. Potential options include the integration of LF surveillance with other public health activities. In American Samoa, blood samples collected in 2010 for a research project on a different disease (leptospirosis) were used to test for evidence of LF infection, and the study found two possible areas of ongoing transmission. We conducted a follow up study in 2014 to verify whether LF infection was truly occurring in these two areas, and found that infection rates in both areas were significantly higher compared to other parts of American Samoa. Our results therefore provide field evidence that integrating LF surveillance activities with other population-based surveys are potentially effective and feasible, and provide a cost-effective method for identifying residual areas of transmission.

Lymphatic filariasis (LF) is a mosquito-borne parasitic infection caused by Wuchereria or Brugia species of helminths. Mosquito vectors vary between countries and regions, and include Aedes, Anopheles, Culex and Mansonia species. Globally, it is estimated that 68 million people are affected, comprising approximately 36 million microfilaraemic persons and 36 million with disabling complications such as severe lymphedema, including elephantiasis and scrotal hydroceles [1]. The Global Programme to Eliminate LF (GPELF) was launched by the World Health Organization in 2000, with the aim of eliminating the disease as a public health problem by 2020. The program consists of two components: i) to interrupt transmission through mass drug administration (MDA) and ii) to control morbidity and disability of affected populations [2]. As part of GPELF, the Pacific Programme to Eliminate LF (PacELF) was formed in 1999 to focus on 22 Pacific Island Countries and Territories (PICTs). PacELF focused on coordinating the elimination efforts in the PICTs, which include >3000 islands and 8.6 million people [3,4]. Since 2000, the GPELF has made impressive progress globally, with a total of 6.2 billion treatments delivered to >820 million people [2]. Once elimination targets have been reached, effective monitoring and surveillance are critical for ensuring that programmatic gains are sustained in the long-term. The World Health Organization and GPELF have identified some key operational challenges in post-MDA surveillance, including i) timely identification of residual areas of high-prevalence and/or resurgence, ii) strategies for managing these high risk areas, and iii) development of cost-effective surveillance strategies [5,6]. In American Samoa, a group of remote islands in the South Pacific, LF is caused by the diurnally sub-periodic W. bancrofti, and the main mosquito vector responsible for transmission is the highly efficient day-biting Ae. polynesiensis. Other vectors include Ae. samoanus (night-biting), Ae. tutuilae (night-biting), and Ae. upolensis (day-biting) [7–10]. Two rounds of MDA in 1963 and 1965 reduced microfilaria (Mf) prevalence from ~20% to <2% [4,11,12]. Unfortunately, transmission was not successfully interrupted, and antigen prevalence measured by rapid immunochromatographic test (ICT) had risen to 16.5% (N = 3018) when the PacELF baseline survey was conducted in 1999. Since then, American Samoa has made significant progress towards LF elimination. After seven rounds of MDA from 2000–2006, antigen prevalence dropped to 2.3% (N = 1881) in a community cluster survey in 2007 [13–15]. However, the results of the 2007 survey did not meet PacELF’s criteria for stopping MDA (<1% antigenaemia, upper 95% CI <2%), and an additional round of MDA was recommended, but no further effective rounds of MDA were successfully completed after this time [15]. The WHO currently recommends post-MDA surveillance using transmission assessment surveys (TAS), which use critical cut-off values of numbers of antigen-positive children aged 6–7 years to determine whether transmission has been interrupted in defined evaluation units [16]. In areas where W. bancrofti is endemic and Aedes is the principal vector, the target threshold for post-MDA transmission assessment surveys (TAS) is <1% antigenaemia [16]. American Samoa passed TAS-1 in 2011–2012 [17] and TAS-2 in 2015 [18], but recent human seroprevalence studies and molecular xenomonitoring studies of mosquitoes identified epidemiological and entomological evidence of ongoing LF transmission [10,19,20]. As prevalence drops to very low levels in the end stages of elimination programs, not only will it become more challenging to detect any residual hotspots of ongoing transmission, but funding and resources for programmatic activities will also generally be reduced. The current WHO guidelines therefore encourage cost-efficient methods for post-MDA surveillance, including the integration of LF surveillance activities with other population-based surveys, and opportunistic screening of large groups (e.g. military recruits, hospital patients, and blood donors) for microfilaraemia, antigenaemia, or antibodies [16]. For example, in Togo [21] and Vanuatu [22], nationwide LF surveillance has been successfully conducted by screening blood smears collected for malaria diagnosis. However, there is currently little evidence about the effectiveness of these strategies for identifying infected persons in the post-MDA setting. Lau et al previously reported a study of the seroprevalence and spatial epidemiology of LF in American Samoa after successful MDA [19]. The study used a serum bank collected from adults (aged ≥18 years) for a leptospirosis study in 2010 [23], four years after the last effective round of MDA. The study found epidemiological evidence of possible residual foci of Og4C3 Ag-positive people in two localised areas, with an average cluster size of ~1.5 km. One cluster was found in the very small village of Fagali’i, where the seroprevalence of Og4C3 Ag was 30.8% (95% CI 9.1–61.4%). Another cluster spanned the contiguous villages of Ili’ili, Vaitogi, and Futiga, where overall Og4C3 Ag prevalence was 15.6% (95% CI 5.3–32.8%) (data derived from [19]). However, the findings were not definitive because information on microfilaraemia was not available; the study was conducted using a pre-existing serum bank and the findings were based entirely on serological markers. In this paper, we report results of a follow up study in 2014 to confirm whether there was indeed ongoing transmission and/or higher infection rates in the two suspected ‘hotspot areas’ identified from the previous work. If ongoing transmission was truly occurring in these two suspected hotspot areas, our findings would provide field evidence to support WHO’s recommendations for integrating LF surveillance activities with other population-based surveys, including those that only include adults [16]. Ethics approvals were granted by the American Samoa Institutional Review Board, and the Human Research Ethics Committees at James Cook University (H5519) and The University of Queensland (2014000409). The study was conducted in collaboration with the American Samoa Department of Health, and official permission for village visits was sought from the Department of Samoan Affairs and village chiefs and/or mayors. Verbal and written information were provided to all participants (or their parent or guardian) in Samoan or English according to the participant’s preference. Signed informed consent forms were obtained from all participants, or their parent or guardian if under 18 years of age. American Samoa is a United States Territory in the South Pacific, consisting of a group of small tropical islands with a total population of 55,519 living in ~70 villages (average population ~800 per village) at the 2010 census [24]. Over 90% of the population live on the main island of Tutuila, and the remainder on the adjacent island of Aunu’u and the remote Manu’a group of islands. American Samoa has a tropical climate and is one of the wettest inhabited places in the world, with islands that include mountains, valleys, tropical rainforests, wetlands, fringing reefs, and lagoons. Field data were collected from American Samoa in 2014 from the following groups of participants: Adult workers and village residents were recruited by convenience sampling because probability-based sampling was not logistically possible with the available budget and resources. The field team was stationed at the weekly Department of Health clinic from May to December 2014 (~4 hours per visit), and invited all clinic attendees to participate. Visits to the tuna cannery, villages, and school were conducted over a 3-week period in October and November 2014. The team visited the tuna cannery on four occasions (~4 hours per visit), and all employees on duty were invited to participate. For visits to the suspected hotspot areas, permissions were sought from village chiefs and mayors, who informed residents of the team’s pre-arranged visits. During the three village visits to Ili’ili/Vaitogi/Futiga, and one village visit to Fagali’i (~4 hours per visit), the field team was positioned in a prominent and central part of the community, and all residents were invited to participate. All children in Grades 3 to 7 who attended the elementary school in Ili’ili were invited to participate; information sheets and consent forms were distributed to parents and guardians about one week beforehand, and all children who returned valid consent forms were tested. The following samples and data were collected from each participant: Data on population demographics were sourced from the 2014 American Samoa Statistical Year Book [24], and high-resolution GIS data were provided by the American Samoa GIS User Group [25]. Venous or fingerprick blood samples were tested for filarial antigen immediately after collection using the Alere BinaxNOW Filariasis immunochromatographic test (ICT). If an ICT was positive, the result was confirmed by repeating the test, and two Mf slides were prepared, each with 60 μL of blood in 3 lines of 20 μL each per slide. Once thoroughly dried, slides were dehaemoglobinized, fixed with methanol and stained with 2% Giemsa stain for 50 minutes according to WHO guidelines [16] and examined at 100x magnification. Mf densities in 60 μL were determined by counting all Mf on each slide. Each set of Mf slides were read blindly by two or three experienced parasitologists, one at James Cook University Cairns (PG) and the other(s) at the LBJ Tropical Medical Centre in American Samoa and/or at James Cook University in Townsville, Australia. Counts were converted to Mf/mL and the final Mf density recorded was the average of the counts reported by two or three parasitologists. Venous blood samples were allowed to clot before centrifuging, and serum were aliquoted and stored at -20 degrees Celsius in American Samoa. Frozen sera and dried blood spots (DBS) were shipped to Australia for serological analysis at James Cook University, Cairns, Australia. All samples were tested for Og4C3 Ag using the TropBio Og4C3 Filariasis Antigen ELISA test (Cellabs Pty. Ltd., New South Wales, Australia) using dilutions for serum and DBS recommended by the manufacturer. Bm14 Ab was measured using ELISA tests (CDC in house version) as previously described [19]. For Wb123 Ab, all the village samples and 109 of the adult worker samples were tested with in-house Wb123 ELISA as previously described [19]. The remaining 552 adult worker samples and 178 of the school children samples were tested using the InBios Wb123 ELISA [26]. Due to insufficient volumes of blood available, 149 children did not have Wb123 ELISA done by either method. All ELISAs used standard curves with kit provided standards (Og4C3 Ag) or known strong positives (Bm14 Ab and Wb123 Ab). Cutoffs for positivity were aligned between the two Wb123 Ab methods. All ICT-positive individuals were treated with albendazole (400mg) and diethyl-carbamazine (DEC) (6mg/kg) according to WHO recommended dosages, and all treatments were provided free of charge. Children were treated with informed consent from and in the presence of at least one parent or guardian. The outcome measures used were positive results for ICT, Og4C3 Ag >32 units (weak positive), Og4C3 Ag >128 units (positive), ‘antigen’ (ICT and/or Og4C3 Ag >32 units), Bm14 Ab, Wb123 Ab, and Mf. The positivity levels for Og4C3 Ag were chosen based on product information provided by Cellabs, and a previous study in American Samoa [19]. To determine whether residents of suspected hotspot areas had higher infection rates than residents of other villages, seroprevalence of adult workers who resided in other villages were used as the reference group to provide estimates of infection rates in the general population. Simple proportions were compared using Chi-squared tests or Fisher exact tests, and binomial exact 95% confidence intervals. Point estimates of antigen and antibody prevalence were calculated for residents of each of the suspected hotspot areas and adult workers who lived in other villages. Prevalence estimates were standardised for age using American Samoa’s age distribution data from the 2014 Statistical Yearbook [24], and 95% confidence intervals calculated using the ‘stdize’ option in the ‘proportion’ command in Stata 14, with ‘stdweights’ as the proportion of the population in each age group. Statistical associations between place of residence and presence of serological markers were quantified using univariable logistic regression (weighted for age distribution). Because adult workers were used as the reference population, village residents aged <15 years were excluded from the logistic regression analyses. Data were managed using Microsoft Excel (v14, 2011) and Qualtrics (Qualtrics, Provo, UT), an electronic platform for collecting and managing data. Stata 14 (StataCorp, College Station, TX) was used for data analyses, and p values of <0.05 were considered statistically significant. The study included a total of 1,132 participants, comprising 172 employment clinic attendees, 498 tuna cannery workers, 125 community members from the two suspected hotspot areas, and 337 school children who attended an elementary school in Ili’ili. The school children represented 61.6% of the total 547 students enrolled in Grades 3 to 8 at the school. Of the 337 children, 283 (84.0%) were residents of Ili’ili/Vaitogi/Futiga, one of the suspected hotspot areas. Children who attended the school but resided in other villages were also tested, but not included in the statistical analyses for residents of suspected hotspot areas because most lived near the school and were therefore not representative of the general population of children in American Samoa. The participants were classified into three groups for statistical analyses: Fig 1 shows the age distributions of residents of suspected hotspots, adult workers from other villages, and American Samoa’s general population. A summary of the representativeness of each study population is shown in Table 1. Fig 2 shows the distribution of the general population on Tutuila, the locations of the suspected hotspot areas included in this study, the elementary school in Ili’ili where children were tested, and the clinic and tuna cannery where adult workers were tested. Fig 3 shows the residential locations of all adult workers who participated in this study. Although the adult workers were sampled at one clinic and one work site, they resided across 51 villages on the main island of Tutuila and the adjacent island of Aunu’u. Considering that convenience sampling was used in this study, the adult workers provided a reasonably representative sample of the general adult population in terms of place of residence. Overall, 72.3% (95% CI 61.4% to 81.6%) of adult community members and 67.6% (95% CI 63.8% to 71.3%) of adult workers reported taking MDA in the past, either in American Samoa and/or elsewhere, while 3.6% (95% CI 0.8% to 10.2%) of adult community members and 1.3% (95% CI 0.5% to 2.5%) of adult workers reported that a doctor or other health worker had previously diagnosed them with LF. There were no statistically significant differences in MDA participation or LF diagnosis between adult community members and adult workers, suggesting that convenience sampling did not introduce any significant participation biases towards either group being more likely to have LF infection. A total of 29 antigen-positive individuals were identified from ICT and/or Og4C3 Ag (>32 units) tests, with a female to male ratio of 1.07 and age range of 9 to 73 years. A summary of antigen-positive results for each group is shown in Table 2. Of the 337 school children tested, three of the 283 residents of Ili’ili/Vaitogi/Futiga were ICT-positive (1.1%, 95% CI 0.2% to 3.1%), compared to none of the 54 who were residents of other villages (0%, one-sided 97.5% CI 0% to 6.6%). Four microfilaraemic individuals were identified out of 20 available slides examined, with Mf densities of 8, 433, 2667, and 3267 Mf/mL. These counts were the average of two or three blind readings of the slides by different parasitologists. The Mf-positive persons were aged 9, 29, 35, and 46 years, with a female:male ratio of 1. All microfilaraemic individuals lived in the hotspot areas; three in Fagali’i and one in Vaitogi. The age distributions of antigen-positive and Mf-positive individuals are shown in Fig 4. Of the 15 ICT-positive people who resided in suspected hotspot areas, Mf results were available for 14, of which 4 (28.6%) were Mf-positive. Of the 8 ICT-positive people who resided in other villages, Mf results were available for 6, of which none were Mf-positive. The difference in proportion of Mf-positive results was not statistically significant, but sample sizes were small. The number of years lived in American Samoa was not significantly associated with seroprevalence for antigen or antibodies. Antigen prevalence was 3.2% in those who had lived in American Samoa since 2000 (when MDA started), 2.5% in those who arrived after 2006 (when the last effective round of MDA was conducted), and 3.3% in those who arrived between 2000–2006 (Chi-squared test, p = 0.89). There were no significant differences in seroprevalence between the three groups for Bm14 Ab (Chi-squared test, p = 0.25) or Wb123 Ab (Chi-squared test, p = 0.65). Estimates of the seroprevalence of LF antigen (positive ICT and/or Og4C3 Ag >32 units), Bm14 Ab, and Wb123 Ab were calculated, and adjusted for age based on the population age distribution reported in the 2010 census [24]. Comparisons of the age-adjusted seroprevalence for adult residents (aged ≥15 years) of each suspected hotspot area and adult workers who resided in other villages are summarised in Fig 5. In adults who lived outside of hotspot villages, age-adjusted seroprevalence of LF antigen, Bm14 Ab, and Wb123 Ab were 1.2% (95% CI 0.6–2.6%), 9.6% (95% CI 7.5%-12.3%), and 10.5% (95% CI 7.6–14.3%) respectively. Comparatively, adult residents of Fagali’i had significantly higher seroprevalence of antigen (26.9%, 95% CI 17.3–39.4%), Bm14 Ab (43.4%, 95% CI 32.4–55.0%), and Wb123 Ab 55.2% (95% CI 39.6–69.8%) than adults living in other areas. Adult residents of Ili’ili/Vaitogi/Futiga also had higher seroprevalence of antigen (4.0%, 95% CI 1.8–8.8%), Bm14 Ab (12.5%, 95% CI 7.1–21.1%) and Wb123 Ab 18.5% (95% CI 11.4–28.5%) than adults living in other areas, but the differences were not statistically significant. Comparisons of the age-adjusted seroprevalence of antigen and antibodies in children (aged 2 to 14 years) from each of the suspected hotspot areas are shown in Fig 6. Age-adjusted seroprevalence of antigen and antibodies were higher in child residents of Fagali’i compared to those who lived in Ili’ili/Vaitogi/Futiga: 5.0% (95% CI 0.7–29.5%) versus 1.3% (95% CI 0.5–3.5%) for antigen; 30.0% (95% CI 13.8%-53.4%) versus 2.0% (95% CI 0.9–4.4%) for Bm14 Ab; and 21.5% (95% CI 7.9–45.5%) versus 5.8% (95% CI 3.1–10.5%) for Wb123 Ab. The study did not collect data from children in other villages that were sufficiently representative for meaningful comparison with the results from suspected hotspots. It should be noted that for some children, there were insufficient blood samples for all serological tests to be conducted. Wb123 Ab results were only available for 173 of the 305 (56.7%) children who lived in Ili’ili/Vaitogi/Futiga, and 19 of the 20 (95%) children who lived in Fagali’i. Odds ratios of the presence of antigens and antibodies in residents of suspected hotspot areas were calculated using univariate logistic regression (weighted for age distribution), using adult workers living in other villages as the reference group. Only participants aged ≥15 years were included in this analysis because the reference group did not include any participants aged <15 years. Table 3 shows that residents of both suspected hotspots were significantly more likely to be seropositive. Residents of Fagali’i had significantly higher odds of being antigen positive (OR 20.4 for ICT, OR 31.8 for Og4C3 Ag >32 units, and OR 23.5 for any antigen) and antibody positive (OR 5.7 for Bm14 Ab and OR 9.5 for Wb123 Ab) than adult residents of non-hotspot areas. Compared to this reference adult group, residents of Ili’ili/Vaitogi/Futiga also had significantly higher odds of being positive for Og4C3 Ag of >32 units (OR 4.1) and Wb123 Ab (OR 2.3), but not for ICT (OR1.6) or Bm14 Ab (OR 1.0). Our results confirm that adult residents of the two suspected hotspot areas were significantly more likely to be seropositive for Og4C3 Ag and Wb123 Ab compared to adult residents of other villages in American Samoa. Residents of Fagali’i were also significantly more likely to be positive on ICT and seropositive for Bm14 Ab. We confirmed the presence of ongoing transmission in the suspected hotspot areas by identifying microfilaraemic residents, including a 9-year-old child. The results of this study therefore support the previous findings of suspected hotspots using a serum bank collected in 2010 for a leptospirosis study [19], and provide field evidence that WHO’s recommendations for integrating LF surveillance activities with other population-based surveys are potentially effective and feasible. However, the current study does not allow us to determine whether the hotspots represent areas of persistent transmission that were not successfully interrupted by MDA, or newly formed hotspots after MDA was completed. Furthermore, our study confirmed ongoing transmission even though American Samoa passed TAS-1 of 6 to 7 year old children in 2011–2012 [17] and again passed TAS-2 in 2015 [18] (conducted a few months after this study). Our findings suggest that testing adults is a potentially effective surveillance strategy, particularly if performed in conjunction with TAS and used as baseline data. However, this strategy might require a sampling scheme quite different from the current WHO recommended sampling methods for TAS. In the post-MDA setting, when overall prevalence is very low and typically even lower in young children, testing adults might be more accurate for determining transmission status and more sensitive for identifying hotspots. The age-adjusted estimates of the prevalence of all serological markers were higher in adult residents of Fagali’i compared to those who lived outside of hotspot villages. The seroprevalence of antigen and Wb123 Ab (but not Bm14 Ab) were higher in Ili’ili/Vaitogi/Futiga compared to residents of non-hotspot villages, but differences were not statistically significant, either because of the small sample size or the true absence of any difference. Although seroprevalence estimates for each group were standardised for age, the variations in age distribution between the groups could have made it more difficult to identify statistical differences. Adult residents of both hotspots had significantly higher odds of being seropositive for Og4C3 Ag and Wb123 Ab than residents of other (non-hotspot) areas. Adult residents of Fagali’i also had higher odds of being seropositive for ICT and Bm14 Ab, compared to residents of other areas, but this was not found in Ili’ili/Vaitogi/Futiga. The reasons for the differences in the patterns of serological markers between the two hotspot areas are not clear, but could potentially be attributed to differences in intensity of transmission; or the timing of possible reintroduction or resurgence; or differences between persistent transmission from before MDA versus reintroduction or resurgence. The age-adjusted estimates of overall seroprevalence of all antigens and antibodies were significantly higher in Fagali’i compared to Ili’ili/Vaitogi/Futiga, suggesting higher transmission intensity in Fagali’i both recently and in the past. Previous studies in children have demonstrated the appearance of antigen and Wb123 Ab earlier in the course of infection than Bm14 Ab [27]. In Ili’Ili/Vaitogi/Futiga, the higher odds of being seropositive to antigen and Wb123 Ab (but not Bm14 Ab), together with ICT-positive 6–7 year olds in the local school in both TAS-1 and TAS-2, might reflect more recent reintroduction and/or resurgence compared to Fagali’i. Our previous study found that recent migrants to American Samoa (mostly from Samoa) had significantly higher antigen and antibody prevalence [19], but this study did not find any significant difference in seroprevalence and number of years lived in American Samoa. A possible explanation is that more time has lapsed since MDA, and there is less difference in infection risk and/or the impact of MDA between long-term residents and recent migrants. The strengths of our study include the large proportion of the population tested in the hotspot areas as well as the general population, and the wide range of age groups included. Our study population was highly stable and allowed accurate assessment of geographic variations in risk; >70% of each group had lived in American Samoa since the PacELF programme commenced in 2000. Our results should also be considered in light of the study’s limitations. Because of financial constraints and limited resources, the study was conducted using convenience sampling instead of probability-based sampling. Our reference group (adult workers who lived in other villages) were over 15 years of age but children were included in the residents of hotspot areas. Our previous study found that recent migrants who had not lived in American Samoa from the beginning of PacELF were more likely to be seropositive for antigen and antibody [19]. A lower proportion of our reference group (70.7%, 95% CI 66.8–74.3%) had lived in American Samoa since the beginning of PacELF compared to residents of Fagali’i (75.7%, 58.8–88.2%) and residents of Ili’ili/Vaitogi/Futiga (75.7%, 95% CI 66.6–83.3%). Our study only considered the place of residence, but infection could have occurred at work or elsewhere, especially when efficient day biters are present. Each of these three limitations could have weakened the associations between living in a hotspot and the presence of serological markers, but our study found statistically significant results despite the limitations. The age distributions of the hotspot residents and adult workers were significantly different to that of the general population, but the estimates of population seroprevalence were adjusted for age. Convenience sampling of hotspot residents and adult workers might have introduced bias towards people who have been diagnosed with LF, concerned that they might have the infection, or previously taken MDA. However, there was no evidence that any biases related to previous MDA or LF diagnosis were different between hotspot residents and adult workers. Our findings raise a number of questions regarding current guidelines and targets used in LF elimination programmes, strategies for post-MDA surveillance, and transmission dynamics in the post-MDA setting. Firstly, our findings indicate that the current WHO recommended TAS has limitations in detecting ongoing transmission in the American Samoa setting. Our current study and previous studies in American Samoa [10,19] found evidence of ongoing transmission despite the territory passing TAS-1 in 2010/2011 and TAS-2 in 2015. In American Samoa, the antigen prevalence threshold used in school-based TAS of young children was not sensitive enough to detect ongoing low-level transmission. Future post-MDA surveillance strategies should consider including older children and adults, and/or determining thresholds that are more specific for different ecological settings [28]. In areas with highly efficient vectors (such as Ae. polynesiensis in American Samoa), LF transmission is likely to be more intense, and might therefore require different elimination targets to successfully interrupt transmission. Furthermore, TAS in American Samoa did not provide any indication of the high antigen prevalence in the Fagali’i hotspot even though >90% of elementary schools were included in the surveys. Our study identified heterogeneity in LF transmission at very small spatial scales, and concur with findings from diverse settings including Samoa [29], Haiti [30], Sri Lanka [31], and Zanzibar [32]. Our results also corroborate findings from Sri Lanka that TAS might not be sensitive enough for identifying small hotspots [31]. However, it is currently unclear whether these small residual foci of transmission will pose any significant risk of resurgence in the broader community, but the presence of microfilaraemic young children in these hotspots suggest that transmission is unlikely to disappear without intervention, particularly in areas with highly efficient vectors and strong environmental drivers of transmission. Secondly, our findings raise concerns that in some settings, seven annual rounds of MDA might not be sufficient for interrupting transmission. Persistent transmission has been noted in Ghana [33] after up to 11 rounds of annual MDA, particularly in areas with high baseline Mf prevalence. In Zanzibar, ongoing transmission was detected six years after MDA, despite good coverage rates and Mf prevalence of <1% at sentinel sites after five rounds of MDA [32]. Thirdly, our results suggest that ICT might not be as sensitive as Og4C3 Ag or Wb123 Ab for detecting low-level transmission or resurgence, such as our hotspot in Ili’ili/Vaitogi/Futiga where Og4C3 Ag and Wb123 Ab provided warning signals, but ICT did not. In Mali, where MDA was conducted from 2002 to 2007, surveys in children in six formerly highly endemic villages found that ICT prevalence decreased from 53% pre-MDA to 0% (N = 120) after 6 rounds of MDA, and all adults tested in these villages were also antigen negative (N = 686) [34]. However, other longitudinal surveys in these villages using Og4C3 Ag and Wb123 Ab showed an increasing trend of antigen and antibody positivity in 6–7 year old children, from 0% in 2009 to 2.7% in 2011 and 4.5% in 2013, with one and three Mf positive children in 2009 and 2011 respectively [35]. Paradoxically, antigen prevalence by ICT in older children (>8 years) and adults decreased from 4.9% in 2009 to 2.8% in 2012. The results suggest that Og4C3 Ag and Wb123 Ab were more sensitive than ICT, and also raise the suggestion that in formerly highly endemic areas, adults might be immunologically protected while young children are susceptible and more rapidly infected. Further research is being conducted in American Samoa to improve understanding of LF transmission in the post-MDA setting by conducting more representative sampling of all age groups in the general population, comparing the sensitivity of school-based versus community-based surveys, identifying risk factors for infection including the role of migrants, determining the spatial distribution and clustering of infected persons, and exploring the use of molecular xenomonitoring of mosquitoes.

10.1371/journal.pbio.2006145

JMJD5 links CRY1 function and proteasomal degradation

The circadian oscillator is a molecular feedback circuit whose orchestration involves posttranslational control of the activity and protein levels of its components. Although controlled proteolysis of circadian proteins is critical for oscillator function, our understanding of the underlying mechanisms remains incomplete. Here, we report that JmjC domain–containing protein 5 (JMJD5) interacts with CRYPTOCHROME 1 (CRY1) in an F-box/leucine-rich repeat protein 3 (FBXL3)-dependent manner and facilitates targeting of CRY1 to the proteasome. Genetic deletion of JMJD5 results in greater CRY1 stability, reduced CRY1 association with the proteasome, and disruption of circadian gene expression. We also report that in the absence of JMJD5, AMP-regulated protein kinase (AMPK)-induced CRY1 degradation is impaired, establishing JMJD5 as a key player in this mechanism. JMJD5 cooperates with CRY1 to repress circadian locomotor output cycles protein kaput (CLOCK)–brain and muscle ARNT-like protein 1 (BMAL1), thus linking CRY1 destabilization to repressive function. Finally, we find that ablation of JMJD5 impacts FBXL3- and CRY1-related functions beyond the oscillator.

In mammals, circadian rhythms are generated by a molecular oscillator in which the circadian locomotor output cycles protein kaput (CLOCK)–brain and muscle ARNT-like protein 1 (BMAL1) transcription factors drive expression of the genes coding for their own repressors, the CRYPTOCHROME (CRY) and PERIOD (PER) proteins. A key feature of the oscillator is that the protein stability of its components is highly regulated. Previous studies had implicated the JmjC domain–containing protein 5 (JMJD5) in regulation of the circadian clock in plants and flies. Here, we show that cells and livers that lack JMJD5 exhibit dysregulation of circadian gene expression. Mechanistically, JMJD5 is required for CRY1 degradation, including its destabilization by AMP-regulated protein kinase (AMPK), by facilitating its interaction with the proteasome. We found that JMJD5 is needed for normal CRY1-mediated transcriptional repression, thereby uncovering an inverse relationship between CRY1 stability and circadian repression. Finally, we showed that JMJD5 impinges on non-clock roles of F-box/leucine-rich repeat protein 3 (FBXL3) and CRY1. Altogether, our studies demonstrate that JMJD5 is a novel link between the oscillator and other physiological processes.

Circadian rhythms are endogenous, approximately 24-hour oscillations in behavior and physiology that evolved as an adaptation to the day–night cycle. These rhythms are generated by a cell-autonomous timekeeping mechanism known as the molecular circadian oscillator. At its most basic, the oscillator is a transcription–translation circuit formed by two interlocked delayed negative feedback loops [1]. In one loop, the transcription factors circadian locomotor output cycles protein kaput (CLOCK) and brain and muscle ARNT-like protein 1 (BMAL1) drive expression of the genes coding for their own repressors, the CRYPTOCHROME (CRY) and PERIOD (PER) proteins, leading to alternative cycles of transcription activation and repression—the molecular basis of the clock. In a second loop, the opposing actions of REV-ERB and ROR nuclear hormone receptors (NHRs) generate strong oscillations in Bmal1 gene transcription, which contributes to robust amplitude in circadian rhythms. However, the function of the circadian oscillator involves a much larger repertoire of factors that include other transcription regulators, kinases, phosphatases, ubiquitin ligases and peptidases, and chromatin regulators. Together, this large cohort of molecules acts in concert to generate circadian rhythms, coordinate the clock with other physiological processes, and enable environmental information to be integrated into its function. A key mode by which circadian rhythms are generated and fine-tuned is by the regulation of the protein levels of the core oscillator components [2]. For instance, phosphorylation of PER proteins by Casein kinase I (CKI) decreases their stability by stimulating their interaction with and ubiquitylation by the Skip-Cullin-F box (SCF)β-TRCP1/2 ubiquitin ligase complex [3,4]. Similarly, degradation of REV-ERBα by Homologous to the E6-AP Carboxyl Terminus (HECT)-(ARF-BP1) and Really Interesting New Gene (RING)-class E3 ligases (PAM) is induced by phosphorylation by glycogen synthase kinase 3β (GSK3β) [5,6]. BMAL1 is also phosphorylated by GSK3β, leading to its destabilization [7]. Although BMAL1 ubiquitylation has been found to be catalyzed by the HECT-class E3 ligase UBE3A [8], a link between this process and GSK3β-mediated phosphorylation has not been found. Taken together, these observations show that, although the mechanisms that control the stability of the clock proteins are similar, involving coordinated phosphorylation and ubiquitylation, there is divergence in the machinery that targets different components. In mammals, CRY degradation is mediated by SCF ubiquitin ligase complexes that contain one of two closely related F-box/leucine-rich repeat proteins (FBXLs), FBXL3 and FBXL21 [9–13]. Although the two ligases can both ubiquitylate CRYs, their actions are antagonistic. In the nucleus, FBXL3 promotes K48-linked polyubiquitylation of CRYs, leading to its degradation, whereas FBXL21 binds with greater affinity to CRY yet catalyzes K48 polyubiquitylation less efficiently than FBXL3 [13]. Thus, presence of FBXL21 diminishes the overall CRY degradation. Despite its presence in the nucleus, FBXL21 localizes primarily to the cytoplasm, where it promotes CRY degradation, highlighting the complexity of CRY1 regulation. As with other clock proteins, CRY1 degradation is controlled by phosphorylation, most notably by the AMP-regulated protein kinase (AMPK) [14]. AMPK-mediated phosphorylation of CRY1 strengthens interactions with FBXL3, thereby leading to CRY destabilization. Yet, despite the fact that both mammalian CRY paralogs largely share the same degradation machinery, differences in how their levels are controlled appear to exist [15–17]. Members of the JmjC domain–containing family of proteins are characterized by a cupin-type domain of about 150 amino acids, known as the JmjC domain, which is able to confer lysine demethylase activity to some but not all proteins that harbor it [18]. In recent times, members of the JmjC family have emerged as important regulators involved in a variety of physiological processes, including control of circadian rhythms in plant, mammalian, and insect systems [19–22]. Previously, a genetic study identified Arabidopsis thaliana Jmjd5 (AtJmjd5) as a regulator of the circadian system in plants that exhibits sufficient functional conservation with its mammalian ortholog JmjC domain–containing protein 5 (JMJD5) as to exhibit reciprocal rescue of circadian phenotypes arising from genetic ablation in plants or small interfering RNA (siRNA) knockdown in U2OS cells [19]. Similarly, in Drosophila, genetic deletion of JMJD5 leads to reduced period length in locomotor activity and decreased sleep [23]. However, though these studies firmly establish JMJD5 as an evolutionarily conserved participant of the clock, its mechanism of action in the clock has remained undefined. Although JMJD5 has been suggested to be a lysine demethylase, such a function remains highly debated and is not yet firmly established [24–26]. Nonetheless, JMJD5 has been reported to influence gene transcription through several mechanisms, including modulation of protein levels, nuclear entry of transcription factors, and proteolytic processing of histone subunits [25,27–30]. We found that JMJD5 is recruited to CRY1–FBXL3 complexes, in which it facilitates CRY1 interaction with the proteasome. Furthermore, we report that JMJD5-dependent CRY1 destabilization is intertwined with the repressive function of CRY1. To determine whether JMJD5 plays a role in the mammalian oscillator, we first analyzed the impact of its deletion on the circadian clock of mouse embryo fibroblasts (MEFs). We measured gene expression levels of core circadian oscillator components in a circadian timeline from Jmjd5+/+ and Jmjd5−/− MEFs harvested at 4-hour intervals from 12 to 56 hours post synchronization with dexamethasone (Fig 1A and S1 Fig). Cells that lack JMJD5 exhibit marked down-regulation of Clock and Bmal1 mRNAs, the two central circadian transactivators. Consistent with a decrease in CLOCK–BMAL1 activity, mRNAs of their regulatory targets Dbp, Cry1, Nr1d1, and Rora showed similar decreases. In contrast, the impact of JMJD5 deficiency on Cry2, Per1, and Per2 gene expression was divergent, with decreased Per2, unaffected Cry2, and increased Per1 mRNA abundances. Real-time bioluminescence measurements from a Per2 promoter-luciferase reporter also showed circadian dysfunction (S2 Fig), as Jmjd5-null cells exhibited shortened period length and decreased amplitudes in their oscillation, which not only confirmed our gene expression observations but also were consistent with previous reports [19,23]. Next, we assessed the impact of JMJD5 on circadian clock gene expression in vivo. Full-body Jmjd5-null mutant mice are embryonic lethal [26], but hepatocyte-specific Jmjd5-ablated animals (Jmjd5 liver knockouts [Jmjd5LKO]) are viable and exhibit no overt phenotype. We generated a circadian liver timeline from wild-type and Jmjd5LKO animals at a 4-hour resolution. As observed in fibroblasts, JMJD5 deficiency disrupts clock gene expression in a similar albeit nonidentical manner. Specifically, in both cells and liver that lack JMJD5, Per1 mRNA was increased, and those of Clock, Per2, and Rora were decreased (Fig 1A and 1B). In contrast to fibroblasts, JMJD5-null livers showed no defect in the expression patterns of Dbp, Bmal1, Cry1, and Nr1d1 mRNA (Fig 1B and S1 Fig). Next, we assessed the impact of JMJD5 on CLOCK–BMAL1-mediated transcription in a series of real-time luciferase-reporter assays performed in the non-oscillating human embryonic kidney 293T (HEK293T) cell line. When coexpressed, JMJD5 decreased CLOCK–BMAL1 activation from Per1- and Per2-driven promoter-driven luciferase reporters in a dose-response manner (Fig 2A and 2B). Repression of CLOCK–BMAL1 by JMJD5 is dependent on the presence of a functional E-box (Fig 2C). We observed that in the E-box mutant promoter, CLOCK–BMAL1 had a suppressive effect; it is possible that this effect is due to sequestration of limiting factors by CLOCK–BMAL1 away from the mutant promoter construct. Further, inclusion of JMJD5 with CLOCK–BMAL1 did not change this effect. In the absence of CLOCK–BMAL1, JMJD5 had only a minor repressive effect on the wild-type Per1 promoter and no such effect in the E-box mutant construct, demonstrating that CLOCK–BMAL1-mediated activation is required for JMJD5 repression (Fig 2D). Although disputed, JMJD5 has been suggested to possess catalytic activity by virtue of its JmjC domain. To define whether any such activity is necessary for its effect on CLOCK–BMAL1 activity, we assessed CLOCK–BMAL1 repression by JMJD5H321A, a mutant construct that harbors a mutation in a conserved residue required for cofactor binding by the JmjC domain, thus precluding any enzymatic function [30–32]. JMJD5H321A repressed CLOCK–BMAL1 activation of both Per1- and Per2-luciferase reporter constructs, indicating that the circadian function of JMJD5 does not require catalytic activity (Fig 2E and 2F). The only other two JmjC proteins that have been shown to participate in the mammalian clock—JARID1A and FBXL11/lysine-specific demethylase 2A (KDM2A)—also do so in a catalytically independent manner [21,22]. JMJD5 has previously been reported to influence other transcription factors via regulation of their stability [25]. Thus, we performed cycloheximide chase assays to assess whether JMJD5-mediated repression of CLOCK–BMAL1 was due to induction of their degradation. JMJD5 did not influence CLOCK or BMAL1 degradation but instead markedly destabilized CRY1 in a catalytically independent manner (Fig 3A and 3B). In contrast, JMJD5 had no effect on other clock proteins, including CRY2 (Fig 3C and S5 Fig). Consistent with our cycloheximide assays, we found elevated CRY1 levels in both liver nuclear and whole extracts of Jmjd5-null livers compared to wild-type controls (Fig 3D and 3E and S6 Fig). In nuclear extracts from JMJD5-null fibroblasts, CRY1 was slightly increased, even though Cry1 mRNA levels were much decreased (Fig 3F), which is the exact same situation observed in livers of FBXL3 mutant animals [9]. To determine whether JMJD5 participates directly in regulation of CRY1 degradation, we interrogated its ability to associate with CRY1–FBXL3 complexes. In coimmunoprecipitation studies, we found that JMJD5 interacts with CRY1, but not CRY2, and that this association was enhanced when FBXL3 was coexpressed (Fig 4A and S9 Fig). CRY1 degradation by FBXL3 is induced via AMPK-mediated phosphorylation of CRY1 residues S71 and S280 [14]. Interaction analyses between FBXL3 and CRY1 constructs harboring phospho-null or phospho-mimetic mutations at these sites showed that although phosphorylation of CRY1 by AMPK increases its binding to FBXL3, it is not required for basal interaction between these two proteins [14]. In a series of coimmunoprecipitation experiments, we found that the binding pattern of JMJD5 to the different phosphosite mutants tested by Lamia and colleagues [14] and to a non-ubiquitylatable CRY1 mutant [13] paralleled that of FBXL3, which further confirmed the existence of CRY1–FBXL3–JMJD5 complexes (Fig 4B and 4C). Association of JMJD5 with CRY1 is dependent on the presence of FBXL3, as RNA interference (RNAi)-mediated knockdown of the latter led to a decrease of CRY1–JMJD5 association (Fig 4D and 4E). In contrast, knockdown of CRY1 did not impact FBXL3 association with JMDJ5, nor did knockdown of JMJD5 abrogate CRY1–FBXL3 interactions (Fig 4F and 4G). Together, these data suggest that FBXL3 bridges the interaction between CRY1 and JMJD5. As FBXL3 mediates CRY1 ubiquitylation, we assessed whether a defect in this process was responsible for the increased CRY1 levels we observed in a JMJD5-null genetic background. To achieve this, we treated control and JMJD5-null MEFs that expressed FLAG-tagged CRY1 with MG132 to block proteasomal degradation (Fig 4H). Ubiquitylation was not affected. At any given timepoint, the intensity of ubiquitylated CRY1 signal was greater in JMJD5-null than in control cells. However, non-ubiquitylated CRY1 was also increased so that the ratio of these two remains unaffected (S10C Fig), indicating that ubiquitylation of CRY1 was normal. We noted that total CRY1 levels in Jmjd5+/+ MEFs increased to 400% of baseline levels after 8 hours of MG132 treatment, yet no similar increase over baseline occurred in Jmjd5−/− cells (Fig 4I and S10 Fig). Quantification of the non-ubiquitylated band alone yielded similar results (S10 Fig). These results suggest a reduction in CRY1 degradation by the proteasome in JMJD5-null cells, even while the normal process of ubiquitylation is unaffected. Coincidentally, JMJD5 has been reported to copurify with 19S proteasome regulatory particle non-ATPase 1 (RPN1), the largest 19S proteasome cap subunit [33]. Of note, RPN1 constitutes a docking site for shuttling proteins that help target ubiquitylated substrates to the proteasome [34]. Based on these observations, we hypothesized that JMJD5 was required for normal CRY1 interaction with the proteasome. To test this, we transfected Jmjd5+/+ and Jmjd5−/− cells with an HA-RPN1 expression construct and assessed its ability to associate with endogenous CRY1. We found that CRY1 association with RPN1 was significantly diminished in JMJD5-null cells (Fig 4J and 4K), which argues that JMJD5 facilitates CRY1 targeting to the proteasome. The seeming paradox posed by the repressive effect of JMJD5 on CLOCK–BMAL1 while simultaneously promoting CRY1 degradation could be resolved if the repressive function of CRY1 was coupled to its degradation. To test this possibility, we performed real-time luciferase assays in non-oscillating HEK293T cells to compare the repressive potential of wild-type to the stable CRY171A/280A mutant. Consistent with the idea that the repressive function of CRY1 is linked with its degradation, CRY171A/280A repression of CLOCK–BMAL1 activation of a Per1-luciferase reporter was markedly impaired, a defect that was most pronounced when comparing conditions with similar protein levels of wild-type and mutant CRY1 (Fig 5A–5C and 5G). Repression of CLOCK–BMAL1 by CRY171A/280A was much less impacted on a Per2-luciferase reporter construct (Fig 5D–5F). As JMJD5 inhibition of CLOCK–BMAL1 was lower on a Per2 than on a Per1-luciferase construct (Fig 2A and 2B, S4B and S4C Fig), it is possible that these observations reflect differences in the regulation of these promoters, consistent with previous reports of differential regulation of PER genes [35–37]. Next, we determined whether JMJD5 could act in concert with CRY1 to repress CLOCK–BMAL1. To this end, we measured the repressive activity of wild-type CRY1 in the presence or absence of JMJD5 coexpression (Fig 5H and 5I). To be able to determine the existence of either cooperation or synergism between CRY1 and JMJD5, we transfected suboptimal amounts of CRY1 and JMJD5 plasmids and found that they co-repressed CLOCK–BMAL1. We also observed that the stable CRY171A/280A mutant cooperated with JMJD5 to an extent similar to the wild type, indicating that JMJD5 could rescue CRY1 activity (Fig 5H and 5I, S11 Fig). Consistently, JMJD5 coexpression destabilized CRY171A/280A to the same extent as wild-type CRY1 (Fig 5J and 5K). A possible explanation for the ability of JMJD5 to rescue CRY171A/280A is that increased JMJD5 availability in the context of basal FBXL3–CRY1 interaction leads to increased proteasome degradation. In all, these results strongly support the idea that the repressive ability of CRY1, at least in some contexts, is linked to its degradation. A key feature of the circadian oscillator is its ability to integrate environmental and cellular information with its machinery. This occurs via modulation of its different molecular components by different signaling pathways. AMPK is a master regulator of energy homeostasis that relays information to the circadian clock via CRY1 [14]. As JMJD5 is required for normal CRY1 degradation, we next explored whether it played a role in AMPK-induced CRY1 degradation. To do this, we assessed the effect of AMPK activation on CRY1 levels in wild-type and Jmjd5−/− MEFs [28]. In the absence of JMJD5, the basal stability of CRY1 was much greater than in wild-type cells, and AMPK activation by 5-Aminoimidazole-4-carboxamide 1-β-D-ribofuranoside (AICAR) treatment failed to induce the rapid destabilization of CRY1 seen previously by Lamia and colleagues (Fig 6A) [14]. In contrast, reconstitution of JMJD5 sensitized CRY1 levels to the effects of AMPK activation (Fig 6B). Together, these data indicate that JMJD5 has a critical role in control of CRY1 stability by AMPK–FBXL3 axis. We next asked whether JMJD5 impinges on other biological functions of FBXL3 or CRY1 besides circadian transcription. We first looked at c-MYC levels because its stability is regulated by FBXL3 in conjunction with CRY2 [38]. We interrogated nuclear extracts of control and Jmjd5−/− MEFs. Across all time points assessed, we found elevation of c-MYC levels in the absence of JMJD5 (Fig 6C). Next, since CRY1 interacts with and modulates the activity of several NHRs [39,40], we assessed whether hepatic JMJD5 ablation impacted the expression profile of genes regulated by NHR partners of CRY1 (Fig 6D). In JMJD5-null livers, we observed increased levels of genes regulated by liver X receptor α and β (LXRα/β) and liver receptor homolog 1 (LRH1) (Abcg5 and Abcg8) [41–43], peroxisome proliferator–activated receptor δ (PPARδ) (Lpl) [44], and pregnane X receptor (PXR) (Gstm3) [45,46]. On the other hand, in the absence of JMJD5, we saw decreases in expression of genes regulated by hepatocyte nuclear factor 4α (HNF4α) and PPARγ (Cyp27a1) [47,48] and by glucocorticoid receptor (GR; Angptl4, Pck1, Lipg) [39,49,50]. In contrast to increased Abcg5 and Abcg8 levels, the expression of another LXR target, Cyp51 [51], was decreased in Jmjd5LKO livers. Finally, we also found decreased expression of HNF4α, an NHR that regulates a broad range of hepatic processes and whose promoter region is bound by CRY1 [52]. In this study, we shed light on the mechanisms by which JMJD5 participates in the circadian oscillator. Specifically, we show that JMJD5 plays a role in CRY1 function and in the regulation of stability. First, JMJD5 expression destabilizes CRY1 but not other circadian proteins. Our data indicate that FBXL3–JMJD5 complexes promote CRY1 degradation by the proteasome. In agreement with this, CRY1–proteasome association is greatly diminished in the absence of JMJD5. We found that although JMJD5 is required for normal CRY1 degradation, it nonetheless cooperates with CRY1 to repress CLOCK–BMAL1, which indicates that CRY1 destabilization and function are, in some cases, positively linked. Indeed, repression of CLOCK–BMAL1 by a degradation-resistant CRY1 mutant is drastically impaired, and JMJD5 simultaneously rescues its functional and stability defects. Though the phenomenon of activation-coupled degradation has been observed in other tightly controlled transcription factors—including the Aryl hydrocarbon Receptor (AhR), Estrogen Receptor alpha (ERα), Sma and Mad homolog 2 (SMAD2), signal transducer and activator of transcription (STATS), and even CLOCK–-BMAL1 [53,54]—this is the first time it has been described for a circadian repressor. Until now, the view regarding the relationship between the repressive function of CRY1 and its protein levels has been that these have direct correlation. As this perspective has been largely shaped by studies involving mechanisms that regulate levels of both CRY1 and CRY2, it is likely that, in such context, differences in the function and regulation between the two CRY paralogs may be obscured. Since the mechanism we describe here is specific to CRY1, we are now able to better define how regulation of CRY1 levels relates to its activity. Cells deficient in JMJD5 exhibit dysregulation in circadian gene expression, albeit with a pattern diverging from simple E-box regulation, which is consistent with previous studies. A Gene Dosage Network Analysis (GDNA) by Baggs and colleagues, for instance, showed that clock gene expression responses to circadian network perturbations are complex, depend on the specific oscillator component that is being disrupted, and do not always follow predicted changes based on transcriptional relationships [55]. In that study, siRNA-mediated depletion of Clock reduces Nr1d1 and Nr1d2 levels, has a marginal impact on Per1, has no effect on Cry2 and Per2, and results in slight increases in Cry1 mRNA, all of which are canonical target genes of CLOCK–BMAL1-mediated activation. In addition, Cry1 depletion in U2OS cells clearly increases levels of Per2 and Cry2 but has no apparent impact on Per1, Nr1d1, or Nr1d2. However, our observations in both Jmjd5-deficient cells and liver do have overlap with findings by Baggs and colleagues. Specifically, the opposite changes in Per1 and Per2 mRNA expression we observed in JMJD5-null cells (increase and decrease, respectively) are consistent with the unidirectional Period paralog compensation in gene expression observed by Baggs and colleagues, in which Per1 depletion increases Per2 levels but not in the reverse [55]. Our functional assays suggest that, at least in certain contexts, JMJD5 may have a more prominent role in control of Per1 transcription than in that of Per2 (Fig 2B and S4 Fig). Consistently, increased CRY1 stability had a much greater impact to repress a Per1-luciferase construct than one driven by the Per2 promoter (Fig 5A–5F). These observations may reflect reports that control of Per1 and Per2 transcription is not identical and raises the intriguing possibility that JMJD5 has a role in the mechanisms behind paralog compensation. Similarly, complex patterns of clock expression occur in tissues of clock component knockout mice. Single knockout of CRY1 or single knockout of CRY2 affects transcription not only differently across genotypes but also across tissues within a genotype [56]. For example, Per2 mRNA levels in the liver of CRY1 knockout mice are elevated and rhythmic but are arrhythmic and mostly reduced in the cerebellum. CRY2 ablation, on the other hand, does not result in derepression of Per2 in liver, a canonical E-box driven target, but quite the opposite. In CLOCK knockout mice, Per1 mRNA levels drop in the hypothalamic suprachiasmatic nucleus (SCN) but rise in the liver, whereas the phase of Per2 mRNA rhythm is shifted without any impact on its levels [57]. In contrast, Per2 mRNA in CLOCK-null mouse liver is elevated only during the nadir of expression, whereas Dbp and Nr1d1 levels are decreased despite drastically elevated Bmal1 gene expression [57]. Finally, a recent study by Ramanathan and colleagues found that knockdown of canonical clock genes (e.g., Cry1, Per1, Per2, Nr1d1) do not always result in the same circadian effect in different cell lines [58]. Altogether, these observations indicate that deletion of a single clock regulator—even of canonical clock components—can lead to nonintuitive effects, which help explain our observations here. Hepatic ablation of JMJD5 also resulted in abnormalities in circadian gene expression. As with Jmjd5−/− cells, we observed elevation of Per1 levels with a simultaneous decrease in Per2 levels, again suggestive of paralog compensation. As with Per genes, paralog compensation in Cryptochromes occurs unidirectionally, so that knockdown of Cry1 gene expression in U2OS cells leads to an increase in Cry2 transcript but not vice versa. In JMJD5-deficient liver, we noted a slight increase in Cry2 mRNA levels, which could reflect a decrease in CRY1 function even if Cry1 transcripts remain unaltered. In contrast to cells, we observed no changes in the levels of Bmal1, Dbp, Cry1, or Nr1d1 transcripts in JMJD5-null liver tissue. A possible explanation for the differences in the impact that JMJD5 deletion has on cells and liver is that circadian clock control is not identical in all cell types and/or due to divergence in how the clock is regulated (e.g., differences in paralog levels) in cultured cells versus in vivo. Knockdown of canonical clock genes (e.g., Cry1, Per1, Per2, Nr1d1) does not always result in the same circadian effect in different cell lines [58], which lends further support to this idea. In a previous study, Huber and colleagues demonstrated that CRY2–FBXL3 specifically regulates c-MYC protein stability. Likewise, we find that c-MYC levels are affected, suggesting that JMJD5 may impact FBXL3 function beyond CRY1 degradation, yet this effect could be indirect, given that JMJD5 did not impact or interact with CRY2. Huber and colleagues also noted a moderate yet noticeable increase in overexpressed c-MYC upon Cry1 ablation [38]. Thus, elevation of c-MYC levels in the absence of JMJD5 is consistent with a deficit in CRY1 function. A possibility is that ablation of JMJD5 disturbs the balance between CRY1–FBXL3 and CRY2–FBXL3 complexes. Nonetheless, the precise mechanisms by which JMJD5 influences c-MYC function remain to be discovered. Second, CRY1 interacts with and participates in NHR-mediated transcription control. In JMJD5-deficient livers, we found abnormal expression of several genes regulated by one or more NHRs that are known to interact strongly with CRY1. The genes we assessed are known to code for important components of metabolic processes, including cholesterol metabolism (Abcg5, Abcg8, Cyp27a1, Cyp51), lipid metabolism (Angptl4, Lpl, Lipg), glucose metabolism (Pck1), and xenobiotic detoxification (Gstm3); their dysregulation suggests that JMJD5 function may have an important role in regulation of liver physiology by CRY1. We found both up-regulation and down-regulation in the expression of these genes, which is reminiscent of what we observed in oscillator components. We found decreased expression of genes regulated by NHRs that are repressed to a similar extent by both CRY1 and CRY2 (e.g., the GR target Angptl4), which is consistent with increased CRY1 levels. In contrast, genes controlled by PPARδ and PXR, which are more strongly repressed by CRY2 than by CRY1, were moderately derepressed [40]; a possible explanation is that this effect is due to a decreased association of CRY2 with PPARδ and PXR as a consequence of increased CRY1 availability. Finally, our observation that JMJD5 seems to impinge on the core clock through CRY1 but not through other core components is intriguing. Several observations suggest that CRY1 serves a unique repressive function. First, CRY1 is able to bind and repress CLOCK–BMAL1 independently of PER [59]. In liver, CRY1 has a markedly different temporal genomic occupancy pattern than that of other circadian oscillators [52]. Furthermore, a recent study found that under certain conditions, most circadian proteins are only detectable as part of a large multiprotein complex, with the exception of CRY1 and CKIδ; in that study, both were detected as uncomplexed from other clock components [60]. When considered together, these observations support the existence of CRY1-specific regulatory mechanisms and thereby suggest that CRY1 constitutes a unique node through which the molecular oscillator machinery is fine-tuned. This work involved the killing of animals by cervical dislocation, as approved by the University of Kansas Medical Center Institutional Animal Care and Use Committee (IACUC) (protocol # 2015–2292). HEK293T cells were purchased from the American Type Culture Collection (ATCC). Cells were cultured in Dulbecco’s Modified Eagle Medium (DMEM) (Corning Cat# 10-013-CV) supplemented with 10% FBS (Atlanta Biologicals Cat# S11595H) and 1% antibiotics and antimycotics (Thermo Fisher Cat# 15240062) in a 37 °C incubator maintained at 5% CO2. Transfections were performed using Trans-IT LT1 (TLT-1) (Mirus Bio Cat# MIR 2304) according to the manufacturer’s instructions (specific conditions described below). HEK293T cells were seeded out in 24-well plates at 80,000 cells per well. Twenty-four hours later, they were transfected with 200 ng of FLAG-CRY1, 40 ng of FLAG-JMJD5, 260 ng of pCDNA3.1 vector (500 ng total), and 1.5 μl of transfection reagent. Forty-eight hours post transfection, cycloheximide (Sigma Cat# C7698-1G) was added to each well to a final concentration of 100 μg/ml, and the cells were harvested every 2 hours. Jmjd5+/+ and Jmjd5−/− immortalized MEFs were obtained from Dr. Ralf Janknecht’s laboratory and have been previously described [28]. Jmjd5+/+ and Jmjd5−/− MEFs were seeded out in 6-well plates at a concentration of 350,000 cells per well. After 20 hours, the cells were transfected with 1 μg of FLAG-CRY1 and 1.5 μl of TLT-1. Forty-eight hours post transfection, the cells were treated with a mixture of cycloheximide (100 μg/ml) ± 3 mM AICAR. For rescue experiments, Jmjd5−/− were seeded out in 6-well plates at a concentration of 350,000 cells per well. After 20 hours, the cells were transfected with 1 μg of FLAG-CRY1 and 100 ng of FLAG-JMJD5 (or pCDNA 3.1+ vector). Forty-eight hours post transfection, the cells were treated with a mixture of cycloheximide (100 μg/ml) ± 3 mM AICAR and harvested at the indicated time points. Jmjd5+/+ and Jmjd5−/− MEFs were seeded out in 6-well plates at 350,000 cells per well and 24 hours later were transfected with 1 μg FLAG-CRY1. After 48 hours of transfection, the cells were treated with 10 μM MG132, and cells were harvested every 4 hours for 12 hours. FLAG-CRY1, FLAG-CRY1S71A, FLAG-CRY1S280A, and FLAG-CRY1AA in pcDNA3.1+ expression backbone were a gift from Katja Lamia. FLAG-JMJD5 and V5-JMJD5 in the pEV3S backbone and HA-JMJD5 in the pQCXIH backbone were generated by Ralf Janknecht. The CRY1(K:R)-HA construct was a kind gift of Dr. Joe Takahashi. Real-time luciferase assays were performed in a 96-well plate format by reverse transfecting HEK293T (40,000 cells per well) with a total of 250 ng of DNA (10 ng of pGL3 Per1: Luc reporter, 30 ng CMV-CLOCK, 10 ng of CMV-Bmal1, and up to 200 ng of test plasmids) and 7.5 μl of TLT-1. The cells were seeded out in phenol red–free DMEM/Ham’s F-12 50/50 mix (Corning Cat# 16-405-CV) supplemented with 10% FBS, 1% Antibiotic-Antimycotic (Life Technologies), 25 mM HEPES, and 125 μM of D-Luciferin. The plate was sealed tight with TempPlate Optical film (USA Scientific). The plate was immediately transferred to the Tecan Infinite M200 maintained at 37 °C, and luminescence was measured in kinetic mode (every 20 minutes) for at least 72 hours. To determine the relative expression of flag components, lysates were prepared at the time corresponding to the peak of CLOCK–Bmal1 activity and analyzed by western blots. Jmjd5+/+ and Jmjd5−/− MEFs were seeded out in a 35-mm dish with 350,000 cells per dish. Sixteen hours later, they were transfected with 2 μg of a Per2-Luc reporter construct. Forty-eight hours post transfection, they were shocked with 0.1 μM dexamethasone for 2 hours. The media were replaced with DMEM:F12 media without phenol red containing 1% antibiotic-antimycotic, 10% FBS, and 25 mM HEPES. The plates were sealed tight and placed in an incubating luminometer (Atto Kronos), and the luminescence was measured for 5 days. Whole-cell lysates from cells and livers were prepared using lysis buffer containing 150 mM NaCl, 50 mM Tris-HCl, 0.5% TX-100, 0.5% NP-40, 0.25% Sodium Deoxycholate 0.025% SDS along with EDTA-free protease inhibitor cocktail and phosphatase inhibitors (Roche Cat# 4693159001 and 4906845001, respectively). Briefly, cells or crushed tissue was incubated with lysis buffer for 30 minutes on ice and then spun at 10,000 rpm for 10 minutes at 4 °C. To prepare the nuclear lysates, liver tissue was homogenized in a hypotonic buffer (10 mM Tris HCl [pH 8.0], 10 mM KCl, 0.5 mM MgCl2) using a Dounce homogenizer. The homogenate was then centrifuged at 800g for 5 minutes at 4 °C. The pellet was resuspended in S1 buffer (0.25 M sucrose and 10 mM MgCl2), layered onto a sucrose cushion S2 (0.88 M sucrose and 0.5 mM MgCl2), and centrifuged at 3,000 rpm for 10 minutes at 4 °C. The supernatant was carefully discarded, and the pellet was resuspended in buffer B2 (10 mM Tris HCl [pH 8.0] and 300 mM NaCl) and incubated on ice for 45–60 minutes. This was then centrifuged at 3,000 rpm for 10 minutes at 4 °C. The resulting supernatant was the nuclear extract and was used in subsequent applications. Nuclear extracts from the MEFs were prepared using this protocol, and confluent 35-mm dishes of cells were used. For immunoprecipitations using M2 FLAG magnetic beads (Sigma Cat# M8823-1ML), cells were lysed using a lysis buffer containing 200 mM NaCl, 50 mM Tris-HCl, 1% TX-100, and 1% NP-40 supplemented with phosphatase and protease inhibitors. HEK293T cells were seeded out in 6-well plates and transfected the next day with a total of 2.5 μg of DNA (1.5 μg V5-FBXL3, 400 ng FLAG-CRY1, and 600 ng of HA-JMJD5). Forty-eight hours post transfection, the cells were lysed, incubated on ice for 30–45 minutes, and spun at 10,000 rpm for 10 minutes at 4 °C. The supernatant was incubated with the M2 beads overnight at 4 °C while tumbling. Subsequently, the beads were washed 3 times with chilled 1X TBS for 5 minutes each. The protein was eluted from the beads with equal volume of 3X flag peptide (Sigma Cat# F4799-4MG). The eluate was boiled in NuPAGE LDS Sample Buffer and reducing buffer and subjected to SDS-PAGE-immunoblot analysis. HEK293T cells were transfected with 1.5 μg V5-FBXL3, 400 ng HA-CRY1K:R, and 600 ng of V5-JMJD5. Forty-eight hours post transfection, the cells were lysed using the lysis buffer described above and immunoprecipitated with HA antibody bound to protein G beads for 1 hour at 4 °C. The beads were washed 3 times with chilled 1X TBS for 5 minutes with mild tumbling. The bound proteins were eluted by boiling in sample buffer and subjected to SDS-PAGE-immunoblot analysis. Jmjd5+/+ and Jmjd5−/− MEFs were transfected with HA-RPN1 and 48 hours later were lysed with buffer containing 400 mM NaCl, 50 mM Tris-HCl, 1% TX-100, and 0.25% Sodium Deoxycholate, supplemented with phosphatase and protease inhibitors. The lysates were incubated on ice for 30–45 minutes and spun at 10,000 rpm for 10 minutes at 4 °C. Protein G beads were prebound with anti-HA tag antibody. The lysates were incubated with the antibody–bead complexes for 1 hour at 4 °C and washed 5 times with 1X TBS containing 0.5% Triton X-100. The bound proteins were eluted by boiling in sample buffer and subjected to SDS-PAGE-immunoblot analysis to detect endogenous CRY1 levels bound to RPN1. FLAG M2 (Sigma), Anti-HA (12CA5), V5 (Abcam), and Anti-CRY1 were the antibodies used in this study. Anti-CRY1 (687) antibody was a kind gift from Satchin Panda. Jmjd5+/+ and Jmjd5−/− MEFs were seeded out in 6-well plates at 350,000 cells per well and shocked 48 hours later with 100 nM of Dexamethasone for 2 hours before supplementing with fresh medium. Thirty-six hours post shock, the cells were harvested using 1 ml TRIzol (Fisher Scientific Cat# 15596026) and stored at −80 °C every 4 hours. Total RNA was then prepared using the manufacturer’s instructions. For real-time qPCR, 1 μg of RNA was reverse transcribed to cDNA using qScript cDNA SuperMix (Quanta Biosciences Cat# 95048–025). FastStart Universal SYBR Green Master (Rox) (Roche Cat# 4913850001) was used to perform the qPCR reaction in a BIO-RAD CFX384 Touch Real-Time PCR System. All the shRNA and siRNA knockdowns were performed using transient cotransfection of the constructs along with overexpression constructs of FLAG-CRY1, HA-JMJD5, and V5-FBXL3 in HEK293T cells in 6-well plates. shRNA hFBXL3 Sigma MISSION shRNA TRCN0000369031 (5′-CCGGCTGATCAGTGTCACGGCTTAACTCGAGTTAAGCCGTGACACTGATCAGTTTTTG-3′), shRNA hCRY1 Sigma MISSION shRNA TRCN0000231065 (5′-CCGGGGAACGAGACGCAGCTATTAACTCGAGTTAATAGCTGCGTCTCGTTCCTTTTTG-3′), siRNA hJMJD5 Sigma MISSION siRNA EHU149061, siRNA universal negative control Sigma MISSION siRNA SIC001. The following gBLOCKS were ordered from IDT and cloned into pGL3 BASIC (Promega). Wild-type E-boxes: AGTGCTAGCCATCACCCACTCACCCCTTAACGACACGTGGGCCCTCAATTGCCCTTCTCTCAGGATCTGAAGGGTCAGAGGAAAGGGTTGGATTCTTTATAACAAGGCTGGGGAGAGGCCAGGGAATGTCAGTCTAGGTTTTTCTCTCTCCCACTTCCCTTGGGTAGCAGACATTTCATTCACCCGGCACCAGGACAGGTGTCTTGTTCTGCCAAGCTGGTCAGTTTAGGAAGTAGGTTTCTCTTGAGCACTTCCTGTGGCCCAGGTATCCTCCCTGAAAAGGGGTAGTTTCCCTCCCTCACTTCCCTTTCATTATTGACGGTGTGAGACATCCTGATCGCATTGGCTGACTGAGCGGTGTCTGAGGCCCTTCAGCCCAGCACCAGCACCCAAGTCCACGTGCAGGGATGTGTGTGACACAGCCCTGACCTCAGTGGGGGCCAGTAGCCAATCAGATGCCAGGAAGAGATCCTTAGCCAACCGGGGGCGGGGCCTGCGGCTCTTCGGGCAGAAGGCCAATGAGGGGCAGGGCCTGGCATTATGCAACCCGCCTCCCAGCCTCGCGGAGCTTCTGGGTTGCAAGCTTAGC. E-boxes mutated: AGTGCTAGCCATCACCCACTCACCCCTTAACGACAgGTcGGCCCTCAATTGCCCTTCTCTCAGGATCTGAAGGGTCAGAGGAAAGGGTTGGATTCTTTATAACAAGGCTGGGGAGAGGCCAGGGAATGTCAGTCTAGGTTTTTCTCTCTCCCACTTCCCTTGGGTAGCAGACATTTCATTCACCCGGCACCAGGACAGGTGTCTTGTTCTGCCAAGCTGGTCAGTTTAGGAAGTAGGTTTCTCTTGAGCACTTCCTGTGGCCCAGGTATCCTCCCTGAAAAGGGGTAGTTTCCCTCCCTCACTTCCCTTTCATTATTGACGGTGTGAGACATCCTGATCGCATTGGCTGACTGAGCGGTGTCTGAGGCCCTTCAGCCCAGCACCAGCACCCAAGTCCAgGTcCAGGGATGTGTGTGACACAGCCCTGACCTCAGTGGGGGCCAGTAGCCAATCAGATGCCAGGAAGAGATCCTTAGCCAACCGGGGGCGGGGCCTGCGGCTCTTCGGGCAGAAGGCCAATGAGGGGCAGGGCCTGGCATTATGCAACCCGCCTCCCAGCCTCGCGGAGCTTCTGGGTTGCAAGCTTAGC.

10.1371/journal.ppat.1002570

A20 (Tnfaip3) Deficiency in Myeloid Cells Protects against Influenza A Virus Infection

The innate immune response provides the first line of defense against viruses and other pathogens by responding to specific microbial molecules. Influenza A virus (IAV) produces double-stranded RNA as an intermediate during the replication life cycle, which activates the intracellular pathogen recognition receptor RIG-I and induces the production of proinflammatory cytokines and antiviral interferon. Understanding the mechanisms that regulate innate immune responses to IAV and other viruses is of key importance to develop novel therapeutic strategies. Here we used myeloid cell specific A20 knockout mice to examine the role of the ubiquitin-editing protein A20 in the response of myeloid cells to IAV infection. A20 deficient macrophages were hyperresponsive to double stranded RNA and IAV infection, as illustrated by enhanced NF-κB and IRF3 activation, concomitant with increased production of proinflammatory cytokines, chemokines and type I interferon. In vivo this was associated with an increased number of alveolar macrophages and neutrophils in the lungs of IAV infected mice. Surprisingly, myeloid cell specific A20 knockout mice are protected against lethal IAV infection. These results challenge the general belief that an excessive host proinflammatory response is associated with IAV-induced lethality, and suggest that under certain conditions inhibition of A20 might be of interest in the management of IAV infections.

Influenza virus or flu epidemics represent a recurrent threat to the public health, especially for individuals which are part of a high-risk group such as children, elderly or immune-compromised people. Sporadic pandemic flu outbreaks, such as the Spanish flu of 1918, may cause high grades of mortality among healthy persons. A better understanding of how the immune system deals with these pathogens is of key importance. The protein A20 is an important negative regulator of both innate and adaptive immune responses. We show that the specific deletion of A20 in myeloid cells, such as macrophages and neutrophils, improves the resistance against otherwise lethal influenza infections. This protective effect is mediated by an enhanced innate immune response following respiratory challenge with influenza virus. Although exaggerated pulmonary immune responses are believed to be the primary cause of often life threatening influenza virus induced pneumonia, we demonstrate that boosting the innate immune response by selectively targeting the functionality of A20 in myeloid cells is beneficial for the host survival. This finding provides us with a novel valuable approach for treating influenza and potentially other respiratory viral infections.

Viruses are a class of highly diverse pathogens which depend on the host cell for their replication. The initiation of a protective innate antiviral immune response involves the action of specialized pattern recognition receptors (PRR), which detect conserved molecular structures of the invading pathogen. Triggering of PRRs induces the production of host proinflammatory cytokines (e.g. TNF, IL-6 and IL-1) and type I interferons (interferon-α (IFN-α) and IFN-β) through activation of downstream signaling pathways that control various transcription factors such as NF-κB, AP-1, IRF3 and IRF7 [1], [2]. The presence of viral nucleic acids, such as viral RNA and DNA, viral replication intermediates and viral transcription products, can be sensed by specific intracellular PRRs [3]. Endosomal Toll-like receptors (TLRs) and cytoplasmic RNA helicase RIG-I-like receptors (RLRs) or Nod-like receptors (NLRs) detect the presence of viral single stranded (TLR7, TLR8, Nod2) or double stranded RNA (TLR3, RIG-I, MDA5). Intracellular DNA sensors that mediate antiviral immune responses to DNA viruses include TLR9, DAI [4] and the PYHIN domain containing proteins AIM2 [5], [6], [7] and IFI16 [8]. TLR mediated antiviral responses are restricted to specialized type I IFN producing plasmacytoid dendritic cells (pDC), while most other cell types, including conventional DC (cDC), macrophages and fibroblasts, depend on the cytosolic RNA and DNA sensors for the production of antiviral proteins [9]. Influenza A virus (IAV) is the etiological agent of a contagious acute respiratory disease that causes considerable mortality, which is generally believed to be due to an excessive host inflammatory response. Emergence of drug-resistant strains of influenza viruses with pandemic potential underscores the importance of developing novel antiviral strategies. In this context, understanding of the mechanisms that regulate IAV-induced immune responses is critical. IAV infection leads to the exposure in the host cell of single-stranded genomic RNA and double stranded RNA, the latter being an intermediate of viral replication. TLR3 and RIG-I have been implicated as sensors of IAV infection [10]–[12]. Both receptors contribute to the proinflammatory response to IAV, but the initiation of the innate antiviral immune response largely depends on RIG-I mediated signaling [13]. Interestingly, RIG-I deficient mice are highly susceptible to IAV [14], [15], whereas TLR3 deficient mice have a survival advantage to acute infection [16]. These results indicate that an imbalance between the beneficial and harmful effects of mediators released by immune cells is likely to contribute to the pathogenesis of influenza. RIG-I contains a C-terminal DExD/H box helicase domain, which is required for ligand recognition, and two N-terminal CARD domains. Upon ligand binding, the CARD domains of RIG-I associate with the CARD domain of the MAVS (also termed IPS-1, VISA, Cardif) adaptor protein, which subsequently translocates to and inserts in the outer mitochondrial membrane via its C-terminal transmembrane domain [17]–[20]. Signaling downstream of MAVS requires the action of various ubiquitin modifying enzymes, which both positively and negatively regulate RIG-I mediated signal transduction [21]. K63-specific ubiquitin ligases (E3s), such as TRIM25 [22] and Riplet [23], [24], have been shown to directly promote RIG-I activation. In addition, well characterized ubiquitin ligases such as TRAF6 [25], [26] and TRAF3 [27] mediate respectively NF-κB and IRF3 activation upon RIG-I stimulation. On the other hand, deubiquitinating enzymes (DUBs), such as DUBA [28], CYLD [29], [30] and OTUB1/2 [31] have been shown to negatively regulate RLR signaling by specifically removing K63-linked polyubiquitin chains from several signaling molecules. Furthermore, various K48-specific ubiquitin ligases, such as AIP4 [32] and TRIAD3A [33] mark respectively MAVS and TRAF3 for proteasome mediated degradation, thus inhibiting further downstream signaling. Additionally, the attachment of K48-specific polyubiquitin chains to the IRF3 and IRF7 transcription factors by E3s such as RAUL [34], TRIM21 [35] and RBCK1 [36] further dampens antiviral signal transduction. A20 is an ubiquitin-editing enzyme belonging to the OTU-domain family of DUBs. Interestingly, A20 also harbors atypical zinc finger dependent K48-specific E3 ubiquitin ligase activity. Both catalytic and noncatalytic mechanisms were previously shown to be involved in the negative regulation of proinflammatory signaling by A20 in response to multiple receptors such as TNF receptor I [37]–[39], TLR4 [40], and IL-1R [41]. The anti-inflammatory role of A20 is clearly demonstrated by the fact that A20 deficient mice die early after birth due to severe multi-organ inflammation and cachexia [37]. More recently, gene targeting of A20 in specific cell types was shown to be associated with autoimmunity and chronic inflammation [42]–[48], further illustrating that A20 is an important brake on the inflammatory response. The relevance of these findings for human disease has recently been illustrated through the identification of polymorphisms in the A20 locus that are associated with several autoimmune diseases and chronic inflammation [45]. In contrast to its well established function in the regulation of proinflammatory responses, the role of A20 in the regulation of antiviral immune responses is less well described and limited to a number of in vitro studies using overexpression or silencing in specific cell lines, indicating that A20 may regulate RIG-I- and TLR3-induced signaling to NF-κB and IRF-3 [49]–[52]. However, the precise in vivo role of A20 in the response to viral infection remains to be clarified. Using myeloid cell specific A20 knockout mice (A20myel-KO) that were recently generated in our lab and primary cells derived of these mice, we here provide evidence that A20 is a crucial negative regulator of IAV-induced proinflammatory and antiviral signaling in macrophages. Interestingly, A20myel-KO mice show enhanced survival and reduced morbidity in response to IAV lung infection compared to wild type mice. Protection against IAV in A20myel-KO mice is associated with increased cytokine and chemokine production, augmented recruitment of innate immune cells such as neutophils and alveolar macrophages, and enhanced viral clearance. These results suggest that boosting the innate immune response to IAV by targeting A20 activity in myeloid cells might have therapeutic potential. RIG-I signaling induces the activation of NF-κB, IRF3 and IRF7 transcription factors, which promote the expression of proinflammatory cytokines and type I IFNs that restrict further viral propagation. Previous studies have shown that ectopically expressed A20 negatively regulates NF-κB and IRF3 activation upon RIG-I stimulation [50]–[52]. Similarly, we show that A20 overexpression in HEK293T cells prevents NF-κB- and IRF3-dependent luciferase reporter gene activation induced by transfection of a truncated constitutive active form of RIG-I [53], corresponding to only the two N-terminal CARD domains of RIG-I [RIG-I (2CARD)] (figure 1A, left and middle graph). We next investigated whether A20 also inhibits IRF7 activation. Unlike IRF3, IRF7 is not or weakly expressed under naïve conditions and IRF7 protein levels are rapidly upregulated upon virus-induced IRF3 activation [54], [55]. To determine the effect of A20 on IRF7 activation, we therefore transfected minor amounts of an IRF7 expression plasmid together with plasmids encoding RIG-I (2CARD), A20 and an IRF7-specific IFNα4 luciferase reporter construct. RIG-I (2CARD) expression in the absence of IRF7 co-expression showed negligible IFNα4 promoter activation (grey bar, figure 1A, right graph), whereas significant reporter gene expression was observed in the presence of IRF7. Similar to its inhibitory effect on NF-κB and IRF3 activation, A20 also prevented RIG-I-induced IRF7 activation (figure 1A, right graph). These results demonstrate the potential of A20 to inhibit RIG-I-induced NF-κB and IRF3/7 activation. To study the effect of endogenously expressed A20 on RIG-I-induced signaling in a more immunological relevant context, we performed further experiments in A20 deficient primary macrophages. Since A20 full knockout mice die prematurely as a result of severe multi-organ inflammation [37], we generated mice carrying a conditional A20 allele in which exon IV and exon V were flanked by two loxP sites [56]. Crossing these mice with transgenic mice expressing Cre recombinase under control of the lysozyme M promoter leads to specific deletion in myeloid cells and allowed us to generate myeloid cell specific A20 knockout mice [48]. To stimulate the RIG-I receptor, we transfected A20myel-KO BMDM and wild type control cells with minimal amounts of low molecular weight (LMW) poly(I:C), which is known to preferentially bind and activate RIG-I rather than MDA5 [57]. Of note, this concentration of poly(I:C) was unable to induce significant TLR3 dependent NF-κB and IRF3 activation or cytokine production (data not shown). As expected, poly(I:C) transfection induced the rapid expression of A20 in wild type, but not in A20myel-KO BMDM (figure 1B, upper panel). At early time points, slightly slower migrating forms of A20 were observed, indicating that A20 undergoes a yet unknown modification upon poly(I:C) transfection. Compared to wild type BMBM, A20 deficient cells showed enhanced NF-κB activation as indicated by increased phosphorylation and sustained degradation of IκBα (figure 1B). Furthermore, nuclear translocation of the p65 NF-κB subunit was enhanced in poly(I:C) transfected A20myel-KO BMDM, reaching a maximum at earlier time points compared to wild type cells (figure 1C). IRF3 is known to be activated upon phosphorylation of a series of carboxyl terminal serine residues by the IKK-like kinases TBK1 and IKKε [58], leading to its dimerization and subsequent translocation to the nucleus [59]. Using immunoblotting with an antibody directed against phosphorylated Ser396, maximum IRF3 phosphorylation was detected at earlier time points in A20myel-KO BMDM compared to wild type BMDM (figure 1B). Similar to p65, IRF3 nuclear translocation reached its maximum at an earlier time point in A20 deficient BMDM compared to wild type cells (figure 1C). NF-κB controls the expression of IL-6 and TNF, and NF-κB and IRF3 control the expression of IFNβ. In line with the enhanced activation of NF-κB and IRF3 as described above, A20myel-KO BMDM secreted increased amounts of IL-6, TNF and IFNβ (figure 1D). Similar results were obtained using peritoneal macrophages (data not shown). Together, these results demonstrate that A20 plays an important role in the negative regulation of RIG-I-induced NF-κB and IRF3 activation in primary macrophages. To investigate the role of A20 in the IAV-induced proinflammatory and antiviral innate immune responses, we infected A20 deficient and control BMDM with IAV X-47 (H3N2). A20 mRNA levels were rapidly induced in wild type BMDM, but not in A20 deficient BMDM, upon viral infection (figure 2A). Furthermore, A20myel-KO BMDM show enhanced expression of IL-6 and IFNβ mRNA after IAV infection compared to control cells (figure 2A). In accordance with these data, cell culture supernatant collected from these cells contained higher levels of TNF and IFNβ (figure 2B). Upon host infection with IAV, alveolar macrophages are an important source of cytokines and chemokines, attracting innate immune cells to the lung during the primary phase of infection. To test whether A20 directly controls IAV-induced gene expression in alveolar macrophages, we isolated these cells from lungs of A20myel-KO and control littermates and infected them in vitro with IAV X-47. Expression and secretion of the proinflammatory cytokines IL-6 and TNF, the type-I IFN IFNβ and IFNα4 and the chemokines MCP-1 (ccl2) and KC (cxcl1) was significantly higher in IAV infected cells lacking A20 compared to infected wild type cells (figure 2C and figure S1). Taken together these results demonstrate that A20 negatively regulates IAV-induced proinflammatory and antiviral gene expression in alveolar macrophages, consistent with the inhibitory effect of A20 seen on RIG-I-induced NF-κB and IRF3 activation. To determine the role of A20 expression in myeloid cells during an IAV infection in vivo, we intranasally inoculated both A20myel-KO mice and control littermates with a sublethal dose of the mouse adapted IAV X-47 (H3N2) strain and monitored morbidity in terms of weight loss. A20myel-KO mice showed reduced weight loss compared to wild type control littermates and recovered faster from the viral challenge (figure 3A). Also, total protein concentration in BAL fluid, which reflects lung damage and vascular leakage, was increased significantly at 7 and 10 days post infection in both wild type and A20myel-KO mice, and was slightly lower in A20myel-KO mice (data not shown). Next, we measured pulmonary viral titers at 4, 7 and 10 days post infection. No differences in viral titers were observed in A20myel-KO mice versus wild type mice at day 4 and 7 post infection. However, after 10 days, almost no virus could be detected in the lungs of A20myel-KO mice while abundant infectious viral particles could still be isolated from lungs of all wild type mice (figure 3B). This indicates that loss of A20 in myeloid cells does not affect early viral replication but contributes to viral clearance at later stages during infection. To verify if A20 deficiency in myeloid cells affects IAV-induced gene expression in the lung, we analyzed the levels of several chemokines and cytokines in the bronchoalveolar lavage (BAL) at day 4, 7 and day 10 following infection. Levels of KC and MIP-2 chemokines, as well as IL-6 were significantly higher at day 7 p.i. in BAL from IAV infected A20myel-KO mice compared to IAV infected wild type mice (figure 3C). Unlike our observations with in vitro stimulated primary macrophages we could not detect a significant increase in MCP-1 or IFNα production in the lungs of A20myel-KO animals (figure 3C). KC is the murine orthologue of IL-8 and serves together with MIP-2 as a chemoattractant for neutrophils, while MCP-1 is mainly known as a chemoattractant for monocytes, which eventually develop into macrophages upon entering the alveolar lumen [60]. Consistent with the higher KC and MIP-2 levels in A20myel-KO mice, the number of neutrophils (CD11b+ Ly6C+ Ly6G+ F4/80−) that were recruited in the bronchoalveolar spaces upon IAV infection was clearly higher throughout infection in A20myel-KO mice compared to control animals (figure 3D). Although we could detect a significant increase in monocyte (CD11b+ Ly6C+) recruitment at day 4 post infection, this was not evident at later time points after infection (figure 3D), which is consistent with the equal expression of MCP-1 in both groups of mice. The number of resident alveolar macrophages (autofluorescent+ CD11c+ F4/80+ CD11b−) was also elevated in A20myel-KO mice but did not differ significantly between IAV infected or mock infected mice (figure 3D). Elimination of IAV infected cells depends on the clonal expansion of virus specific cytotoxic CD8+ T cells (CTL) [61], [62], [63]. To test whether A20 expression in myeloid cells regulates the antiviral CTL response, total CD8+ T cells and virus specific Granzyme B (GrB) and IFNγ expressing CD8+ T cells were measured in BAL and lung parenchyma of wild type and A20myel-KO mice. A clear increase in CD8+ T cells could be detected at day 7 and 10 post infection, but no differences were observed between A20myel-KO and wild type mice (figure S2A and figure S2C). Also, the number of GrB and IFNγ CD8+ T cells as well as IFNγ expression levels in the lungs were not altered by the absence of A20 in myeloid cells (figure S2A–C). Protection against IAV infection is also provided by the humoral immune response. To test whether loss of A20 in myeloid cells affects B cell mediated immunity, we determined hemagglutinin (HA) antibody titers in the serum of A20myel-KO and wild type littermates. However, no differences could be detected between wild type and A20myel-KO animals (figure S2D), indicating that humoral immunity is not affected by A20 expression in myeloid cells. Together, these data suggest that mechanisms other than adaptive immunity, such as an increased innate immune response, characterized by an increased influx of neutrophils and increased numbers of alveolar macrophages, contribute to the better viral clearance in A20myel-KO mice. It is generally believed that IAV-induced mortality is due to an excessive proinflammatory response in the lung. We therefore analyzed whether the increased proinflammatory cytokine production and infiltration of proinflammatory cells in A20myel-KO mice affects mortality induced by intranasal infection with a lethal dose of IAV X-47. Surprisingly, almost all A20myel-KO mice survived (10/11), while all control mice succumbed (0/11) within 16 days after infection (figure 4A). A20myel-KO mice still showed significant weight loss and lung damage (as reflected by increased total protein concentration in the BAL; data not shown) during the course of infection but were able to recover, in contrast to wild type mice that succumbed (figure 4B). Similar to our observations with sublethal IAV infection, pulmonary KC and MIP-2 production was stronger in A20myel-KO animals compared to wild type mice following lethal IAV infection (figure 4C), which correlates with the increased numbers of neutrophils in the lungs of these mice (figure 4C). Also levels of the proinflammatory cytokines IL-6, TNF and IL-1β, which are often associated with immunopathogenesis in humans [64], were increased in the lungs of A20myel-KO mice compared to control animals (figure 4C and figure S3A). Again, MCP-1 production was not increased and even lower in A20myel-KO mice (figure 4C), and also monocyte recruitment was not different between both groups of mice. We could also not detect any differences in viral clearance or antiviral adaptive immunity at 6 h post infection (figure S3B–F). Collectively these data indicate that A20 deficiency in myeloid cells is associated with an increased innate immune response and protection against a lethal IAV infection. In the present study we have investigated the contribution of A20 expression in myeloid cells in the innate immune response to IAV lung infection. In the pulmonary environment, macrophages populate both lung parenchyma and the alveolar lumen where they are referred to as alveolar macrophages. Under naïve conditions, alveolar macrophages exert important immunomodulatory functions [65], [66]. However, alveolar macrophages are also crucial in the innate immune response against IAV as they are one of the first cells that encounter infectious IAV particles [67], [68]. They are an important source of proinflammatory cytokines and chemokines that attract innate immune cells to the lung during the primary phase of infection [69], and they are the primary producers of type I IFNs [70]. Alveolar macrophages are also known to phagocytose virus infected apoptotic cells and antibody coated viral particles, further contributing to viral clearance. We could show that BMDM or alveolar macrophages derived from A20myel-KO mice express higher amounts of chemokines, cytokines and type I IFNs compared to wild type mice in response to in vitro infection. Similarly, in vivo infection with IAV resulted in higher levels of primarily neutrophil attracting chemokines such as KC and MIP-2 and several proinflammatory cytokines such as IL-6, TNF and IL-1β in the lungs of A20myel-KO mice compared to wild type mice. This was associated with a selective enhancement of neutrophil recruitment to the bronchoalveolar compartment, and resulted in improved viral clearance later on during infection. Although the role of neutrophils during viral infection is still under debate, recent evidence supports a protective function of these cells during IAV infection [71], [72]. MCP-1 levels were not affected by the absence of A20 in myeloid cells, which is consistent with the notion that airway epithelial cells are the primary source of MCP-1 production during IAV infection [73]. Mice deficient for the MCP-1 receptor CCR2, which is expressed on a subset of circulating monocytes, are protected against IAV infection and display reduced signs of immunopathology [74]–[76]. During IAV infection these monocytes develop into inflammatory dendritic cells or proinflammatory macrophages [77] and are considered major contributors to IAV-induced immunopathology [78]. A20myel-KO mice were protected against a lethal IAV infection, which is rather surprising since an excessive proinflammatory cytokine response and an excessive influx of inflammatory cells is generally believed to be associated with increased lethality [64], [79]. However, the selective effect of A20 deficiency on neutrophil recruitment, without altering inflammatory monocyte (Ly6C+ CD11b+) recruitment, further support the idea that monocytes and not neutrophils are major contributors to IAV-associated immunopathology and lethality [78]. We show that A20 deficient BMDM display enhanced NF-κB and IRF3 activation in response to RIG-I stimulation by synthetic LMW double stranded RNA. RIG-I has previously been shown to play a key role in the innate immune response to IAV [13], suggesting that the increased immune response of A20myel-KO mice to IAV lung infection reflects enhanced RIG-I signaling. We propose that A20 inhibits IAV-induced proinflammatory gene expression (as shown in our manuscript for TNF, IL-6, KC, MIP-2, and IFNβ) by negatively regulating NF-κB and IRF3 activation, which are the major pathways controlling these genes. However, this does not exclude an additional effect of A20 on other signaling pathways that may contribute to proinflammatory gene expression. A20 is believed to exert its NF-κB and IRF3 inhibitory functions by modulating the ubiquitination status of different signaling proteins [80]. In this context, it was recently shown that A20 cooperates with the ubiquitin-binding proteins TAX1BP1 and ABIN1 to to disrupt the TRAF3-TBK1-IKKε complex, thereby negatively affecting K63-polyubiquitination of TBK1 and IKKε, and their ability to activity IRF3 [81], [82]. Whether similar mechanisms are involved in the regulation of RIG-I induced NF-κB activation is still unclear. So far we were unable to clearly detect ubiquitination of TBK1 and IKKε in primary macrophages, preventing us from studying the effect of A20 deficiency on their ubiquitination status. It cannot be excluded however that A20 also targets other substrates that mediate NF-κB and IRF3 activation in myeloid cells. The identification of these substrates will be the topic of future investigations in our laboratory. Multiple other deubiquitinating enzymes (DUBs), such as DUBA [28], CYLD [29], [30], OTUB1/2 [31], and A20 [49]–[52] have been shown to negatively regulate RIG-I signaling to NF-κB and IRF-3, implicating possible redundancy. However, evidence so far was limited to in vitro data and was obtained under non-physiological conditions. The clear protective phenotype of A20myel-KO mice that we here describe indicates that A20 expression in myeloid cells is a central gatekeeper of RIG-I induced signaling in response to IAV infection and that other DUBs cannot substitute for A20 deficiency under physiological conditions. If A20 has a similar non-redundant role in other cell types that are implicated in the response to IAV, such as lung epithelial cells, remains to be investigated. Understanding and controlling the activation of innate antiviral immune responses is an important step toward novel therapies. About a fifth of world's population is infected by IAV annually, leading to high morbidity and mortality, particularly in infants, the elderly and the immunocompromised. The high mutation rate of IAV turns all attempts of vaccine and antiviral design into a never ending battle. In recent years, RNA interference, triggered by synthetic short interfering RNA (siRNA), has rapidly evolved as a potent antiviral regimen. Properly designed siRNAs have been shown to function as potent inhibitors of influenza virus replication. Although specificity and tissue delivery remain major bottlenecks in the clinical applications of RNAi in general, intranasal application of siRNA against respiratory viruses including, but not limited to influenza virus, has experienced significant success and optimism [83]. Our results suggest that not only siRNA targeting IAV components, but boosting the antiviral immune response by intranasal administration of siRNA against A20 might be a valid therapeutic approach. Also small compound inhibitors of A20 might be an interesting alternative. Finally, similar targeting of A20 might be of interest in the battle against other viral infections including RSV and SARS coronavirus. All experiments on mice were conducted according to the national (Belgian Law 14/08/1986 and 22/12/2003, Belgian Royal Decree 06/04/2010) and European (EU Directives 2010/63/EU, 86/609/EEG) animal regulations. Animal protocols were approved by the ethics committee of Ghent University (permit number LA1400091, approval ID 2010/001). All efforts were made to ameliorate suffering of animals. Mice were anesthetized by intraperitoneal (i.p.) injection of a mixture of ketamine (12 mg/kg) and xylazine (60 mg/kg). A20fl/fl mice were generated as previously described [56]. A20fl/fl mice were crossed with LysM-Cre mice [84] (provided by I. Förster, Institute of Genetics, University of Cologne, Germany) to generate A20fl/fl LysMCre transgenes and are described in detail elsewhere [48]. Mice were housed in individually ventilated cages at the VIB Department of Molecular Biomedical Research in specific pathogen-free animal facilities. Influenza infections were performed on age- (between 7 and 9 weeks old) and sex-matched littermates. A20fl/fl LysM-Cre animals were backcrossed three times to the C57Bl/6 background. A20fl/fl mice expressing or lacking the LysM-Cre transgene were termed A20myel-KO and wild type (A20myel-WT) respectively. Mouse adapted IAV X-47 (H3N2; PR8×A/Victoria/3/75) was propagated in MDCK cells. For viral inoculation, mice were anesthetized by i.p. injection with ketamine (12 mg/kg) and xylazine (60 mg/kg) and 50 µl X-47 diluted in PBS was administered intranasally. For lethal and sublethal infection, mice received respectively 2-LD50 or 0.05-LD50 X-47. To determine pulmonary viral titers, median tissue culture infectious dose (TCID50) was measured as follows: lungs were homogenized with a Polytron homogenizer (Kinematica) in PBS. Eight-fold serial dilutions of lung homogenates were incubated on MDCK cells for 5 days in DMEM supplemented with trypsin (1 µg/ml), 2 mM L-glutamine, 0.4 mM sodium pyruvate and antibiotics. For read-out, 0.5% chicken red blood cells (RBC) were added and end-point dilution of hemagglutination was monitored. TCID50 titers were then calculated according to the method of Reed and Muench [85]. To determine the HAI titers in infected mice, sera of these were treated with receptor-destroying enzyme (RDE/Cholera filtrate; Sigma) to remove sialic acids from serum proteins capable of aspecific inhibition of agglutination. After incubation overnight at 37°C, the RDE was inactivated by addition of 0.75% sodium citrate in PBS and heating to 56°C for 30 min. To remove sialic acid binding proteins, sera were cleared with 1/10 volume 50% chicken RBC. Titration was done by incubating a two-fold dilution series of sera with 4 HA units of X-47 virus for 1 hour at room temperature in 96-well U-bottom plates. Finally, an equal volume of 0.5% chicken RBC was added and titers were read 30 min later. Negative controls included PBS instead of immune serum (agglutination control) or PR8 instead of X-47 virus (control for agglutination effect of sera); as positive control, serum from a mouse infected twice with a sublethal dose of X-47 was used. Granzyme B (GrB) and IFNγ expressing CD8+ T cells were determined by treating the mice intranasally with 50 µg Brefeldin A (Sigma) as previously described [86]. 6 h later, BAL and lungs were isolated and single cell suspensions were prepared from the lung in the presence of 3 µg/ml Brefeldin A. Cells were stained, fixed and permeabilized (Cytofix/Cytoperm, BD Biosciences) according to the manufacturer's instructions. Activated CD8+ T cells were analyzed by flow cytometry based on CD62lo CD3+ and CD8+ expression. Live/Dead fixable aqua dead cell stain kit (Molecular Probes) was used to discriminate live from dead cells. HEK293T and MDCK cells were grown in DMEM (Gibco) supplemented with 10% FCS, 2 mM L-glutamine, 0.4 mM sodium pyruvate and antibiotics. HEK293T cells were transfected using the calcium phosphate precipitate transfection method with specific expression vectors (pCAGGS-E-hA20 (LMBP 3778), pCAGGS-E-RIG-I-CARD (LMBP 6517), pEF-HA-IRF-7 (kindly provided by T. Taniguchi, Graduate School of Medicine and Faculty of Medicine, University of Tokyo)), NF-κB, IRF3, IRF7 reporter plasmids (respectively pConLuc (LMBP3248), pISRE-luc (LMBP4011), pGL3-IFNα4-luc (kindly provided by J. Hiscott, McGill University, Montreal, Quebec, Canada), and pACTbetagal (LMBP4341) for transfection efficiency normalization. Details of plasmids are presented along with detailed sequence maps at the BCCM-LMBP plasmid databank http://bccm.belspo.be/index.php. For the generation of BMDM, bone marrow cells were cultured 7 days in RPMI 1640 (Gibco) supplemented with 10% FCS, 2 mM L-glutamine, 0.4 mM sodium pyruvate, antibiotics and 40 ng/ml recombinant M-CSF. BMDM were of ≥95% purity as measured by flow cytometry using F4/80 and CD11b specific antibodies. For the isolation of alveolar macrophages, the trachea was canulated and the lung was flushed 4 times with HBSS containing 1 mM EDTA. Alveolar macrophages were cultured in RPMI 1640 (Gibco) supplemented with 1% FCS, 2 mM L-glutamine, 0.4 mM sodium pyruvate and antibiotics. For total lysates, cells were lysed at 4°C for 15 min in lysis buffer (200 mM NaCl, 1% NP-40, 10 mM Tris-HCl pH 7.5, 5 mM EDTA, 2 mM DTT) supplemented with protease and phosphatase inhibitors. Nuclear and cytoplasmic lysates were prepared by resuspending cells in B1 (10 mM Hepes pH 7.5, 10 mM KCl, 1 mM MgCl2, 5% glycerol, 0.5 mM EDTA and 0.1 mM EGTA supplemented with protease and phosphatase inhibitors) for 15 min at 4°C. Next, NP-40 detergent was added to a final concentration of 0.65% and cells were centrifuged at 500 g for 5 min. The nuclear fraction containing pellet was lyzed in B2 (20 mM Hepes pH 7.5, 1% NP-40, 400 mM NaCl, 10 mM KCl, 1 mM MgCl2, 20% glycerol, 0.5 mM EDTA and 0.1 mM EGTA supplemented with protease and phosphatase inhibitors) for 15 min at 4°C. The lysates were subsequently separated by SDS-PAGE and analyzed by western blotting and ECL detection (Perkin Elmer Life Sciences). Immunoblots were revealed with anti-A20, anti-IκBα, anti-p65, and anti-histon H1 (Santa Cruz), anti-IRF3 (Invitrogen), anti-phospho-IRF3 and anti-phospho-IκBα (Cell Signaling) and anti-actin (MP Biomedicals). The density of the bands was quantified (fold induction) with the ImageJ (http://rsbweb.nih.gov/ij) Gel analyzer tool. All intensities were calculated relative to the first lane ( = time 0). Lungs were dissected and incubated with collagenase type IV (1 mg/ml; Sigma) and DNAse (100 U/ml; Roche) at 37°C for 30 min. Subsequently, samples were filtered through a 70 µm and 40 µm nylon mesh. For the preparation of BAL, trachea were canulated and airway lumen was flushed 4 times with HBSS with 1 mM EDTA. Cells were stained with monoclonal antibodies directed against MHC-II (I-A/I-E) FITC (M5/114.15.2), CD11c PerCP-Cy5.5 (N418), F4/80 APC (BM8), CD62L PE (MEL-14), Granzyme B FITC (NGZB) from eBiosciences and CD3 Molecular Complex Horizon v450 (17A2), Ly6C Horizon v450 (AL-21), Ly6G PE (1A8), CD11b APC-Cy7 (M1/70), CD8α PerCP (53-6.7), IFNγ Alexa 647 (XMG1.2) and CD16/32 (2.4G2) from BD Pharmingen. Samples were acquired on a LSRII Cytometer and analyzed using FACSDiva software (BD Biosciences). Propidium iodide was used to discriminate between live and dead cells. For TNF ELISA, 96-well plates were coated with TNF coating (TN3-19, eBioscience) and detection (R4-6A2, eBioscience) antibodies. IFNα and IFNβ protein levels were determined with an ELISA kit (PBL Biomedical Laboratories). For IFNγ ELISA, 96-well plates were coated with IFNγ coating (XMG1.2) and detection (R4-6A2) antibodies (eBiosciences). Detection of MCP-1, KC, TNF, IL-1β and IL-6 in BAL fluid was performed using Bioplex (BioRad) technology according to the manufacturer's instructions. Milliplex technology (Millipore) was used for the detection of MIP-2 in BAL fluid. Total RNA was extracted using Aurum Total RNA mini kit (BioRad) and reverse transcribed into cDNA with iScript cDNA synthesis kit (BioRad) according to the manufacturer's instructions. qPCR was performed by using SYBR Green I master mix I (Roche) in the Lightcycler 480 detection system (Roche) with the following primers: HPRT: 5′-AGTGTTGGATACAGGCCAGAC-3′ and 5′CGTGATTCAAATCCCTGAAGT-3′; IL-6: 5′-GAGGATACCACTCCCAACAGACC-3′ and 5′-AAGTGCATCATCGTTGTTCATACA-3′; IFNβ: 5′-TCAGAATGAGTGGTGGTTGC-3′ and 5′-GACCTTTCAAATGCAGTAGATTCA-3′; A20: 5′-AAACCAATGGTGATGGAAACTG-3′ and 5′-GTTGTCCCATTCGTCATTCC-3′; CCL2: 5′-TTAAAAACCTGGATCGGAACCAA-3′ and 5′-GCATTAGCTTCAGATTTACGGGT-3′; CXCL1: 5′-GAGCCTCTAACCAGTTCCAG-3′ and 5′-TGAGTGTGGCTATGACTTCG-3′ and IFNα4: 5′-TGATGAGCTACTACTGGTCAGC-3′ and 5′-GATCTCTTAGCACAAGGATGGC-3′. Primers were designed with PerlPrimer (http://perlprimer.sourceforge.net). Quantification was performed using the comparative CT method (ΔΔCT). Results are expressed relative to HPRT values. Results are expressed as the mean ± SEM. Statistical significance between groups was assessed using two-way ANOVA. The differences for in vivo experiments (at least 5 mice per group) were calculated using the Mann-Whitney U-test for unpaired data. Statistical significance of differences between survival rates was analyzed by comparing Kaplan-Meier curves using the log-rank test (GraphPad Prism version 5, GraphPad, San Diego, CA).

10.1371/journal.pntd.0003635

What Drives the Occurrence of the Melioidosis Bacterium Burkholderia pseudomallei in Domestic Gardens?

Melioidosis is an often fatal infectious disease affecting humans and animals in tropical regions and is caused by the saprophytic environmental bacterium Burkholderia pseudomallei. Domestic gardens are not only a common source of exposure to soil and thus to B. pseudomallei, but they also have been found to contain more B. pseudomallei than other environments. In this study we addressed whether anthropogenic manipulations common to gardens such as irrigation or fertilizers change the occurrence of B. pseudomallei. We conducted a soil microcosm experiment with a range of fertilizers and soil types as well as a longitudinal interventional study over three years on an experimental fertilized field site in an area naturally positive for B. pseudomallei. Irrigation was the only consistent treatment to increase B. pseudomallei occurrence over time. The effects of fertilizers upon these bacteria depended on soil texture, physicochemical soil properties and biotic factors. Nitrates and urea increased B. pseudomallei load in sand while phosphates had a positive effect in clay. The high buffering and cation exchange capacities of organic material found in a commercial potting mix led to a marked increase in soil salinity with no survival of B. pseudomallei after four weeks in the potting mix sampled. Imported grasses were also associated with B. pseudomallei occurrence in a multivariate model. With increasing population density in endemic areas these findings inform the identification of areas in the anthropogenic environment with increased risk of exposure to B. pseudomallei.

Melioidosis cases are on the rise in endemic areas of northern Australia and Thailand. This potentially severe infectious disease affecting humans and animals in the tropical belt is caused by the gram negative bacterium Burkholderia pseudomallei. Domestic gardens are a common point of exposure to these environmental bacteria and B. pseudomallei are more prevalent in the dry season in gardens when compared to other areas. This is why we analysed whether common gardening practices such as regular watering (irrigation) or soil fertilizing change the occurrence of B. pseudomallei. We conducted a soil microcosm experiment with a range of fertilizers and soil types as well as a longitudinal interventional study over three years on an experimental fertilized field site in an area naturally positive for B. pseudomallei. Irrigation was the only consistent treatment to increase B. pseudomallei occurrence over time. The effects of fertilizers upon these bacteria depended on soil texture, physicochemical properties such as pH or salinity and vegetation. B. pseudomallei occurrence was also associated with imported grasses. With increasing populations in endemic areas, these findings inform the identification of areas in the anthropogenic environment with increased risk of exposure to B. pseudomallei.

Southeast Asia and tropical Australia have recently experienced a surge in melioidosis, an often fatal infectious disease caused by the saprophytic environmental bacterium Burkholderia pseudomallei [1,2]. Case numbers in the Top End of Australia have substantially increased in recent years. In the 20 years from 1989 until 2009 there was a median of 27 cases annually [3]. In the last 5 years there has been a median of 64 cases annually and in each of two recent years, 1 in every 2,000 people living in the Top End has had culture confirmed melioidosis [4]. B. pseudomallei are found in soil and water world-wide in the tropical belt with the major endemic region being southeast Asia and tropical Australia [5–10]. B. pseudomallei is an opportunistic pathogen able to infect humans [11] and a large variety of animals [12]. Humans with a compromised immune system such as from diabetes, hazardous alcohol use, chronic renal disease and immunosuppressive therapy are at particular risk of acquiring and dying from melioidosis [13]. Clinical presentations vary widely and include skin and soft tissue abscesses, pneumonia and disseminated infection with septic shock, the latter having mortality rates above 80% [14]. The Darwin area (12° S latitude) in the tropical north of Australia is endemic for melioidosis and gardening is considered to be an important recreational and occupational source of exposure to and ultimately, infection with B. pseudomallei [3]. In the 20-year Darwin prospective melioidosis study, 407 (75%) of 540 consecutive melioidosis patients had documented recreational activities such as gardening or outdoor sporting activities where exposure to B. pseudomallei was considered likely to occur [3]. Domestic gardens are not only a common ground for humans to be exposed to the environment, but B. pseudomallei might also thrive in the garden habitat. While B. pseudomallei and melioidosis predominate in the monsoonal wet season [3], previous work in rural Darwin found that in the dry season B. pseudomallei is more often present in domestic gardens than in farms or environmentally less disturbed areas [15]. This might be attributed to the widespread use of irrigation during the dry season. Being a non-spore forming, gram negative bacterium, B. pseudomallei is often, but not exclusively associated with moist soil close to a water source and with surface water or alluvial areas as well as rice fields [7,15–19]. At environmentally disturbed sites, B. pseudomallei was associated with pens or paddocks for pigs, chickens or horses with an average odds ratio of 3.8 [15]. This raises the possibility that soil aeration through digging activities or organic material and nitrogen from animal waste support growth of B. pseudomallei [15]. In this study, we addressed the hypothesis that anthropogenic manipulations associated with gardens such as the use of irrigation, fertilizers, commercial potting mix or keeping pets influence the habitat of B. pseudomallei and change its abundance and/or occurrence. We conducted a soil microcosm experiment with a selection of fertilizers as well as a longitudinal study over three years on an experimental fertilized field site in a location naturally endemic for B. pseudomallei. In August 2008 an experimental site was established on a private property in rural Darwin in an area that previously tested positive for B. pseudomallei. The soil at this site was a hydrosol [20] and the soil texture of the topsoil was clay with a subsoil consisting of grey clays and siltstone. The site consisted of two plots, 0.75 metres apart and each plot had six 1x1 metre quadrants (Fig. 1), which included a control quadrant and five quadrants with different treatments which represent common garden practices in the Darwin region (Table 1). Treatments were applied every two weeks with water application every 2nd day for three years. Timing and dose reflected local garden practices. The water used was unchlorinated water from the property’s bore with a pH of 7.5 containing 50 mg/L calcium carbonate and which repeatedly tested negative for B. pseudomallei by culture. There were 14 rounds of soil sampling and in each round 2 random soil samples were collected from each of the 12 quadrants to give a total of 336 soil samples. The first sampling round was before the start of the experiment in August 2008 followed by sampling every two months in year-1, every three months in year-2 and every four months in year-3 of the experiment, with the last round in August 2011. Soil from a depth of 20–30 cm was collected into sterile 50 mL specimen containers and auger and spade were cleaned with 70% ethanol between soil collections [21]. Soil moisture was determined as described previously using the Australian Soil and Land Survey Field Handbook [21,22]. Soil pH was measured using a soil pH field test kit (Inoculo, Australia). In the last 6 months, soil electrical conductivity (EC) was measured using the Field Scout EC Meter (Spectrum Technologies, USA). Grasses covering the experimental field site were identified by the Northern Territory Government Herbarium and were either Sorghum spp. (spear grass) or Pennisetum pedicellatum (annual “mission grass”). At the time of sampling, the presence or absence of live specimens of these grasses at the sampling hole was noted. Of 120 250-mL clean and autoclaved plastic containers, 30 were each filled with either 130 g of commercial “garden soil”, sandy clay loam, clay or sand (Table 2). The non-commercial soil was collected in rural Darwin and tested negative for B. pseudomallei by culture. Nine different treatments plus controls (no change) were applied in triplicate to the containers (Table 3). Treatments included the addition of distilled water or distilled water in combination with eight fertilizers which are commonly used in residential gardens in the Darwin region. After two weeks of soil conditioning at 32 degrees Celsius in the dark, all soils were inoculated with 5x10e4 CFU of an environmental strain of B. pseudomallei (MSHR2817) which has the commonly found multi-locus sequence type (ST) 144 [23] and incubated at 32 degrees Celsius for four weeks in the dark. Soil DNA was extracted and B. pseudomallei DNA detected as described below. Soil DNA extraction was done as previously described [15,21]. Briefly, 20 g of soil were incubated with 20 mL of Ashdown’s broth for 39 hours shaking at 37°C, the soil supernatant was centrifuged twice and the pellet processed using the PowerSoil Kit (MoBio Laboratories, USA). Modifications included the addition of 0.8 mg of aurintricarboxylic acid (ATA) and 20 μL of proteinase K (20 mg / mL). B. pseudomallei DNA was targeted by the well validated B. pseudomallei specific Type Three Secretion System-1 TTS1 real-time PCR [24,25]. For the microcosm experiment, DNA was extracted from 20 g of soil using the previously described semi-quantification method with an internal extraction and amplification plasmid control [23]. TTS1 copy numbers were normalized by dividing them by the copy number of the internal pt7 plasmid control which was added to the soil samples prior to extraction, in order to account for differences in DNA extraction and amplification efficiency as a result of varying amounts of inhibitors present in soil samples [23]. Statistical analysis was carried out using Stata (Intercooled Stata, version 12.1, USA). For bivariate analyses, Fisher’s exact test and Mann-Whitney U test were used. All tests were 2-tailed and considered significant if P values were less than 0.05. Graphs were generated in Stata and GraphPad Prism 6. For the experimental field site, a conditional logistic regression model was used to model the odds of B. pseudomallei occurrence once the experiment had started, with fixed effects for treatments and dates of sampling. Fractional treatment effects were assumed for the first 12 months (e.g. 50% of full treatment effect after 6 months) to allow the application of fertilizers to have a gradual effect on the soil environment and Burkholderia community [26]. Heat maps for the experimental field site were generated using the thin-plate-spline interpolation method in ArcGIS 10.1 (ESRI 2012). B. pseudomallei occurrence was monitored over three years on a field site with five different treatments applied in an area in rural Darwin naturally positive for B. pseudomallei (Fig. 1) The microcosm experiment was used to determine whether commercial fertilizers commonly used in gardens in the Darwin region increased B. pseudomallei load in soil. Eight different treatments of garden fertilizers were applied to each of four different soil types in triplicate. There were also triplicate water controls with only distilled water added and triplicate controls with nothing added. Four weeks after inoculation, no B. pseudomallei were detected in commercial garden soil for any of the treatments. Sand and clay contained on average 733 times more B. pseudomallei than sandy clay loam which only contained minimal B. pseudomallei cells (Fig. 5). After four weeks, no B. pseudomallei (10/12) or only minimal B. pseudomallei cells (2/12) were detected in the 12 control samples. The effect of fertilizers upon B. pseudomallei differed between soil type (Fig. 5). In sand, the addition of a fertilizer rich in urea showed the highest B. pseudomallei load compared to controls but the same fertilizer only had a moderate effect in clay and no effect in sandy clay loam. The pH and EC of urea in sand were lower with a mean of 4.6 and 56 μS/cm in comparison to sandy clay loam (pH 5.8 and EC 216 μS/cm) and clay (8.4 and 172 μS/cm). The addition of an organic and a phosphorus rich fertilizer resulted in the two highest mean B. pseudomallei counts in clay but no such effect was seen for the other soils. The addition of water alone caused a similarly high increase in B. pseudomallei load in sand and clay which had a low VSW of 0–2% before addition of water but there was no load increase in sandy clay loam with a higher initial VSW of 5%. Domestic gardens have been known for many years to be a source of acquisition of melioidosis [3]. It is not clear why this is the case other than gardens being a common meeting point between humans and the environment. However, our previous work has found an increased occurrence of B. pseudomallei in gardens in comparison to control areas in the dry season [15]. One of the main anthropogenic manipulations in gardens in the Darwin region is irrigation, i.e. regular watering during the prolonged mid-year dry season. Irrigation indeed proved to be the only treatment in this study to be associated with a significant longitudinal increase in occurrence of B. pseudomallei on the experimental field site. Furthermore, the addition of water was one of the main predictors for higher B. pseudomallei load in the microcosm experiment. This matches with previous reports that B. pseudomallei is often found on irrigated sports grounds [27,28] and golf courses; and irrigated rice fields are a known major risk factor in acquiring melioidosis in Southeast Asia [6,7,19]. B. pseudomallei occurrence generally increased across the field in 2010 which coincided with above average rainfall in the wet season 2009/2010 and high melioidosis case numbers during that time [2]. In addition to irrigation, the use of fertilizers is another common soil disturbance factor in domestic gardens. B. pseudomallei belongs to the Betaproteobacteria, a class with members including B. pseudomallei capable of ammonium oxidation, denitrification and polyphosphate accumulation, thereby providing a selective advantage over other bacteria in fertilized, eutrophic ecosystems [29]. A greater relative abundance of Betaproteobacteria was found in sediments of eutrophic reservoirs and agricultural wetland soils and abundance decreased after restoration [29]. In another study, a shift to Burkholderia spp. was evident after a change from forest to pasture vegetation [30] and tillage and fertilization have been shown to affect the Burkholderia community structure [26]. We found the impact of fertilizers upon B. pseudomallei to be complex and dependent on soil type, physicochemical soil parameters such as pH or salinity as well as biotic factors such as vegetation. In the microcosm experiment, a fertilizer rich in phosphorus or phosphates caused the highest mean B. pseudomallei load increase in clay with a neutral soil pH but only a small effect in sandy clay loam and sand. Phosphates adsorb to clay minerals due to the clay’s electrostatic surface and depending on soil pH, react with soil cations such as iron cations, making phosphates unavailable to microbes [31]. A neutral pH is the ideal range for maximum phosphate availability. Apart from essential functions of phosphates, B. pseudomallei uses phosphates to generate polyphosphates for oxidative stress response, motility and biofilm formation [32]. In the microcosm experiment, a fertilizer rich in nitrates increased B. pseudomallei growth across different soil types. These results match with a study conducted in Thailand, where B. pseudomallei was associated with soil containing more total nitrogen [9]. Nitrate is one of the biologically most important compounds in the nitrogen cycle and highly susceptible to leaching, thereby contaminating groundwater [33,34]. As a denitrifier B. pseudomallei reduces nitrates to nitrites as electron acceptors for anaerobic respiration [33]. Another nitrogen containing compound which increased B. pseudomallei growth in the microcosm experiment was a fertilizer rich in urea. Urea is hydrolysed to ammonia which is used by B. pseudomallei in biosynthetic pathways and is also oxidized to nitrates by nitrifying soil bacteria [35]. In previous studies B. pseudomallei occurrence was higher in areas where animals were kept (mainly horses and chickens) and the soil in these areas likely contained increased levels of urea [15,21]. However, neither urea nor the ammonium containing NPK fertilizer increased B. pseudomallei load on the fertilized field site. The latter might be due to the nutrient salts of the NPK fertilizer considerably increasing the soil salinity on these quadrants. B. pseudomallei is a saprophyte so it was surprising not to find an association with the organic fertilizer in the field experiment. Concentrated organic material such as found in commercial potting mix has high buffering and cation exchange capacities, resulting in increasing pH and salinity. Indeed, no B. pseudomallei was recovered after four weeks in the tested commercial potting mix which showed exceptionally high EC values of up to 1,000 μS/cm. A preference of B. pseudomallei for less saline conditions and thus, less osmotic stress has previously been reported for B. pseudomallei in water, media and a soil microcosm study [17,36–38]. Nitrogen containing fertilizers are also known to acidify the soil in the long term with the release of hydrogen ions through nitrification processes by soil bacteria oxidizing ammonium to nitrites and nitrates [39]. Soil pH controls the availability of many nutrients in soil and is one of the strongest drivers of the soil bacterial community structure [29,40,41]. For B. pseudomallei, soil pH has previously been found to be an important abiotic soil parameter [9,15,17,36,37,42]. This study confirmed the preference of B. pseudomallei for a slightly more acidic soil but also with a decline in B. pseudomallei occurrence for pH below 5 [42,43]. The preference of B. pseudomallei for a more acidic soil makes it well equipped to grow in the weathered, lateritic soil commonly seen in tropical Australia. Interestingly a similar soil environment in Gabon has recently been shown to harbor B. pseudomallei [44]. This unmasking of the potential for endemic melioidosis in central Africa has important implications for ongoing studies that are attempting to define the geographical boundaries of the environmental presence of B. pseudomallei globally [10]. High annual monsoonal rainfall leads to excessive leaching with a depletion of alkaline cations in the topsoil, leaving behind hydrogen ions which contributes to the acidity of the soil as well as its low buffering capacity [41]. The pH on the experimental field site was indeed highly acidic to start with, having a median pH of 4.5 at baseline. After 3 years, the pH on the quadrants which were irrigated every 2nd day rose to 6, most likely as a result of high ion load such as naturally occurring magnesium or calcium in the irrigation water, which was untreated bore water containing 50 mg/L calcium carbonate and with a pH of 7.5 [45]. The application of urea every two weeks also increased the pH from 4.5 to 5.5 due to urea hydrolysis releasing ammonia which was converted to ammonium at low soil pH. No inhibitory effect of garden lime (32% w/w calcium carbonate) against B. pseudomallei was observed in the microcosm study after application as per manufacturer’s instructions (1% w/w). This matches a previous finding that even with quicklime (calcium oxide) which is more caustic than garden lime, a bactericidal effect against B. pseudomallei was only observed if mixed into the soil at considerable 40% w/w leading to a pH increase above 10 [46]. As a common habitat for bacteria of the Burkholderia genus, B. pseudomallei colonizes the rhizosphere and aerial parts of various plants such as grasses of the family Poaceae [23,47,48]. In particular exotic grasses introduced to Australia for pasture such as Brachiaria humidicola and Pennisetum pedicellatum (annual “mission grass”) have been found to be colonized by B. pseudomallei [23]. On the experimental field site, mission grass started to appear during the first year of the experiment, replacing native Sorghum spp. The presence of mission grass was a significant predictor for the presence of B. pseudomallei in a multivariable model accounting for soil pH and moisture, supporting previous findings of B. pseudomallei colonizing these grasses. These results suggest that while mission grass might have influenced the occurrence of B. pseudomallei across the field; there was no evidence that this grass preferentially occurred on the irrigated quadrants and thus, the growth of mission grass could not explain the association between B. pseudomallei and irrigation. Statistical power was limited with four replicates per treatment per time point for 13 time points on the experimental field site, and three replicates per treatment per soil type in the microcosm experiment. Further studies are recommended to confirm the results. Furthermore, a small amount of mixing of treatments across quadrants of the experimental field could not be excluded; however, salinity data indicated no or only minimal mixing. Remediation measures to decrease B. pseudomallei load in gardens also need more formal study. Measures might include a reduction of irrigation and improved drainage as well as increasing the buffering capacity of the soil causing a rise in soil pH and salinity [46]. While a large amount of quick lime is needed to raise the soil pH and ultimately decrease B. pseudomallei counts [46], the use of potting mix might help increase the soil salinity due to its high cation exchange capacity. A reduction of fertilizers such as those containing nitrates might also assist in reducing load as well as restoration of native vegetation, with the latter also requiring less irrigation. It was previously reported that at a location in Western Australia B. pseudomallei was no longer detected after removal of chemical fertilizers and restoration of native vegetation [49]. In summary there was clear evidence for irrigation increasing B. pseudomallei occurrence. The effect of fertilizer application upon B. pseudomallei was more complex and was dependant on soil type and physicochemical properties as well as on vegetation, with nutrients also causing an increase in plant root development beneficial to B. pseudomallei. The use of fertilizers is causing drastic changes to the global nutrient cycle with a significant rise in supply of otherwise limiting nutrients. These changes have a major impact upon the soil and water microbial community structure and likely also upon host pathogen interactions [50], including those involving B. pseudomallei.

10.1371/journal.pbio.2005761

Reporting bias in the literature on the associations of health-related behaviors and statins with cardiovascular disease and all-cause mortality

Reporting bias in the literature occurs when there is selective revealing or suppression of results, influenced by the direction of findings. We assessed the risk of reporting bias in the epidemiological literature on health-related behavior (tobacco, alcohol, diet, physical activity, and sedentary behavior) and cardiovascular disease mortality and all-cause mortality and provided a comparative assessment of reporting bias between health-related behavior and statin (in primary prevention) meta-analyses. We searched Medline, Embase, Cochrane Methodology Register Database, and Web of Science for systematic reviews synthesizing the associations of health-related behavior and statins with cardiovascular disease mortality and all-cause mortality published between 2010 and 2016. Risk of bias in systematic reviews was assessed using the ROBIS tool. Reporting bias in the literature was evaluated via small-study effect and excess significance tests. We included 49 systematic reviews in our study. The majority of these reviews exhibited a high overall risk of bias, with a higher extent in health-related behavior reviews, relative to statins. We reperformed 111 meta-analyses conducted across these reviews, of which 65% had statistically significant results (P < 0.05). Around 22% of health-related behavior meta-analyses showed small-study effect, as compared to none of statin meta-analyses. Physical activity and the smoking research areas had more than 40% of meta-analyses with small-study effect. We found evidence of excess significance in 26% of health-related behavior meta-analyses, as compared to none of statin meta-analyses. Half of the meta-analyses from physical activity, 26% from diet, 18% from sedentary behavior, 14% for smoking, and 12% from alcohol showed evidence of excess significance bias. These biases may be distorting the body of evidence available by providing inaccurate estimates of preventive effects on cardiovascular and all-cause mortality.

In the scientific literature, reporting bias occurs when communication and publication of results are influenced by the direction of findings. Reporting bias can distort scientific evidence and may misguide subsequent clinical and public health efforts. Our study provided an assessment of the degree of reporting bias in the literature on health-related behavior (smoking, alcohol, diet, physical activity, and sedentary behavior) and statins and their association with cardiovascular disease and mortality. We analyzed recently published systematic reviews. Most of the systematic reviews (90%) had a high risk of bias related to study eligibility criteria, identification and selection of studies, data collection and study appraisal, and synthesis and findings. We found evidence of reporting bias in about one-fifth of health-related behavior meta-analyses but none of the statin-related meta-analyses. Readers should be aware of the extent of reporting bias in these research areas when interpreting meta-analytical results.

The literature on the association between behavioral risk factors (e.g., smoking, alcohol, physical inactivity, and unhealthy diet) and cardiovascular diseases—the single largest cause of death globally [1]—has grown exponentially in the last decades [2–39]. Observational epidemiological studies are the dominant design assessing these associations, since clinical trials cannot always be ethically or logistically conducted [40]. Systematic review methods are used to synthesize and evaluate this growing body of evidence. It is important to evaluate the methodological risks of bias in systematic reviews [41], as well as the impact that reporting bias can have on the findings of reviews [42, 43]. Reporting bias is one of the most common biases identified in the literature. It includes selective publication of studies or outcomes of studies [44, 45] based on factors other than the study quality, such as nominally statistically significant results (P < 0.05) [46, 47] or authors’ “pedigree” [44, 45, 48]. These practices threaten the completeness and validity of scientific evidence [46] by distorting the estimates of causal effects of interventions or exposures on diseases [49]. The extent of reporting bias could differ between bodies of evidence consisting of randomized trials, such as drug studies, compared to observational studies, such as studies of health behavior. Different levels of reporting bias in the literature on health behavior may lead to inaccurate estimates of preventive effects on cardiovascular and all-cause mortality and therefore offer incorrect guidance for policymaking. To gain a better understanding of the potential reporting bias in the literature on health-related behavior and cardiovascular disease mortality and all-cause mortality, we examined reporting and other risks of bias in a sample of systematic reviews published between 2010 to 2016. Our analysis also provided a comparative assessment of the reporting bias between health-related behavior and statins used in primary prevention. Of the 5,511 records identified while searching the databases, we selected 49 systematic reviews. All research areas (tobacco, alcohol, diet, physical activity, sedentary behavior, and statins) presented fewer than 20 eligible systematic reviews; therefore, we included all the systematic reviews within each area that met our inclusion criteria (Fig 1). Lists of excluded reviews and reasons for exclusions are described in S1 and S2 Tables. Studies were excluded most frequently for not including one of the exposures (28%) or outcomes (29%) of interest and for utilizing clinical samples (16%). Most of the included systematic reviews (n = 35, 71.4%) analyzed only one outcome eligible for our study (cardiovascular disease mortality and all-cause mortality), whereas 9 (18.4%), 4 (8.2%), and 1 (2.0%) analyzed two, three, and four outcomes, respectively. All-cause mortality (69%), cardiovascular disease mortality (29%), and stroke mortality (14%) were the most frequent outcomes investigated (Table 1). The majority of the systematic reviews exhibited a high overall risk of bias (n = 44, 90%) (Fig 2). Among the four ROBIS domains, domain 1 (study eligibility criteria) presented the best scores, with 32 (65%) out of 49 reviews showing a low risk of bias. In domain 2 (identification and selection of studies), 2 (4%) reviews were scored as unclear, 40 (82%) showed a high risk, and 7 (14%) a low risk of bias. Whereas, in domain 3 (data collection and study appraisal), 7 (14%) reviews were scored as unclear, 28 (57%) scored with high risk, and 14 (29%) with low risk of bias. Finally, in domain 4 (synthesis and findings), 2 (4%) review was scored as unclear, 30 (61%) with high risk, and 17 (35%) with low risk of bias (Fig 2 and S3 Table). Comparing risk of bias in the reviews across research areas, sedentary behavior performed worst in domain 1 (study eligibility criteria; 70% of reviews were regarded as having high risk of bias). All research areas performed poorly in domain 2 (identification and selection of studies), with high risk of bias ranging from 70% in smoking reviews to 90% in both sedentary behavior and statin reviews. Alcohol (70%) and diet (60%) reviews presented high risk of bias in domain 3 (data collection and study appraisal). Sedentary behavior (90%), smoking (70%), and diet (70%) reviews presented high risk of bias in domain 4 (synthesis and findings). Overall, statin reviews presented the best scores in the ROBIS assessment compared to other research areas. Among statin reviews, a low risk of bias was identified in 60% in domain 1, 10% in domain 2, 50% in domain 3, and 60% in domain 4 (Table 1 and Fig 3). We identified 111 meta-analyses (exposure–outcomes associations) that were performed across the 49 included reviews. On average, each meta-analysis synthesized results from 9 primary studies (ranging from 2 to 81), including 331,688 participants (ranging from 595 to 3,674,042) and 19,012 deaths (ranging from 33 to 320,252) (Table 2 and S4 Table). Of the 111 meta-analyses, 72 (65%) showed a nominally statistically significant result at P < 0.05. Nominally statistically significant results (P < 0.05) were found in 92% of the meta-analyses from sedentary behavior and 100% of the meta-analyses from physical activity and smoking. Alcohol and statin reviews had 38% and 45% of meta-analyses with P < 0.05 results, respectively (Table 2 and S4 Table). We conducted a sensitivity analysis by restricting the sample in each research area to meta-analyses with ≥10 primary studies. In this subsample (n = 29), 86% of the meta-analyses showed statistically significant results at P < 0.05, as compared to 65% in the entire sample of meta-analyses. These results varied by research area, ranging from 60% in statin meta-analyses to 100% in physical activity, sedentary behavior, and smoking meta-analyses (Table 3). Small-study effect was present in 31% of the meta-analyses. The proportions of meta-analyses in the sensitivity analysis with small-study effect were 80% for physical activity, 50% for sedentary behavior, 29% for alcohol, and 33% for smoking. Diet and statin meta-analyses had no evidence of small-study effect. Around 38% of the health-related behavior meta-analyses with ≥10 primary studies presented small-study effect, as compared to zero in statin meta-analyses (Table 3). Excess significance was identified in 27% of the meta-analyses with ≥10 primary studies: 100% of the meta-analyses for sedentary behavior, 50% for diet, 40% for physical activity, 20% for alcohol, and 17% for smoking. Around 33% of the health-related meta-analyses with ≥10 primary studies showed evidence of excess significance, as compared to zero in statin meta-analyses (Table 3). Overall, after excluding small individual studies (with <200 deaths) from meta-analyses, results from small-study effect and excess significance tests did not change (S5 Table). This study aimed to assess the extent of reporting bias among recent meta-analyses that examined the associations of health behavior and statins with cardiovascular and all-cause mortality. We found evidence of reporting bias across all health-related behavior areas. The degree of reporting bias varied by the method used to assess it. Reporting bias was present in 20% (according to excess significance test) or 18% (according to small-study effect test) of all meta-analyses included (health behavior and statins). Evidence of reporting bias was found in between a quarter and one-fifth of health-related behavior meta-analyses (22% small-study effect and 24% excess significance) but in none of the statin meta-analyses (0%). In lifestyle epidemiology, the interpretation of evidence for researchers and policymakers is challenging for several reasons [61]. As observational studies are the dominant designs in this area, spurious associations can arise due to confounding or several sources of bias. The impact of such biases on statistical findings and interpretation of findings has been poorly reported and discussed [62]. Therefore, meta-analytical synthesis of the evidence in lifestyle health behavior epidemiology may provide precise but spurious results [63]. Reporting bias is a major threat to the validity of the relevant body of evidence. Our results suggest that around 20% of the meta-analyses on health-related behavior and cardiovascular disease mortality and all-cause mortality may be susceptible to reporting biases. The existence of reporting bias in the literature has several explanations. Failure to submit manuscripts of analyses that did not produce statistically significant results (“the file-drawer problem” [46]) and the low likelihood of publication of small studies (regardless of statistical significance) [44] are two possible reasons. The selective reporting of certain analyses with statistically significant results is another likely source of reporting bias [44, 46, 47]. Each of the research areas we examined is likely to be linked to variable levels of reporting bias due to the different economics, dynamics, and conflicts of interest in each discipline [64, 65]. Interpreting the literature as a whole is challenging, considering the numerous biases that may affect the reliability and integrity of the scientific enterprise [66, 67]. To obtain a complete picture of the evidence (i.e., without reporting bias), it is important to know the results from all conducted studies on a given research question [68]. In our study, results from meta-analyses of health-related behavior and cardiovascular disease mortality and all-cause mortality were more likely to be affected by reporting bias compared to statin meta-analyses (22% and 24% versus 0%, respectively). The literature of health-related behavior is almost exclusively composed by observational studies, whereas statins are most often studied using randomized controlled trials. Reporting bias may be less frequent among trials than observational studies because several efforts to increase transparency and reproducibility of results have been adopted over the history of randomized controlled trials [69]. These include the mandatory registration of all clinical trials in humans and disclosure of all results [70]. As of more recently, data sharing statements of clinical trials are also required [71]. Observational epidemiologic studies should embrace these reproducible research practices to reduce reporting bias in the literature [68–70, 72]. These practices could involve key elements of the scientific process, including (a) methods (e.g., rigorous training in statistics), (b) reporting and dissemination (e.g., disclosure of conflicts of interest), (c) reproducibility (e.g., open data), (d) evaluation (e.g., pre- and postpublication peer review), and (e) incentives (e.g., funding replication studies) [72]. Improving methodological training involves aspects of both research design and statistical analyses—for example, correct interpretation of P values [73], acknowledging the importance of statistical power, and improving the accuracy of effect sizes [72]. Protecting against cognitive biases is another major issue that has been overlooked [72]. Protecting against conflict of interests, especially financially related, is an imperative to achieve reproducible science. In addition to disclosure of potential conflicts of interest, promoting preregistration of study procedures and analytical plan may prevent reporting bias favoring positive results [72]. Funding replication of studies and encouraging openness in science and reproducibility practices by making datasets, scripts, and software publicly available may increase transparency and credibility of scientific claims [72]. For instance, food industry–sponsored studies are more likely to report conclusions favorable to the sponsors [74] but frequently lack transparency on acknowledgment of the funding source [75]. Further examples of reproducibility practices have been described and discussed by Munafò and colleagues [72]. To our knowledge, our analysis is the first comparative assessment of reporting bias across different fields of health-related behavior and statins. Our findings were based on well-established statistical tests developed to detect different aspects of reporting bias, as well as a complementary assessment of the risk of bias of systematic reviews using the ROBIS tool. We selected the ROBIS tool as it has greater specification to assess risk of bias compared to other tools. For instance, the “Assessing the Methodological Quality of Systematic Reviews” (AMSTAR) that has been used to evaluate the methodological quality of systematic reviews has constructs that are more related to quality of reporting than risk of bias [76, 77]. Risk of bias is linked to methodological quality of systematic reviews but provides further evaluation on how methodological limitations were considered to form conclusions. In this sense, the ROBIS tool is increasingly being used to assess risk of bias not only in systematic reviews [41, 76, 78] but also in guideline committees that evaluate evidence level (e.g., Australian government, National Health and Medical Research Council). Our ROBIS tool results showed that most of the systematic reviews had high risk of bias. Similar findings have been observed in previous studies appraising risk of bias in other research areas using the ROBIS tool [76, 78]. For instance, 18 (58%) out of 31 systematic reviews evaluating the effectiveness of intra-articular hyaluronic acid injection in treating knee osteoarthritis had high (n = 16) or unclear (n = 2) risk of bias [78]. Another survey assessing systematic reviews about psoriasis found that most reviews (86%) were classified as high risk of bias [76]. It is noteworthy that high risk of bias was found even for systematic reviews exhibiting high methodological quality as assessed through AMSTAR [76]. Our ROBIS assessment indicated that identification and selection of studies (i.e., appropriate range of databases, terms and filters used, and efforts to minimize errors in selection of studies) are major concerns. These biases in the review process could explain, at least in part, reporting bias results obtained from small-study effect and excess significance tests. The synthesis and findings domain also revealed potential risk of bias due to insufficient inclusion of studies and appropriate synthesis of estimates. This domain also reflects between-study variation, robustness of findings (e.g., sensitivity analyses), and biases in synthesis findings (i.e., if evaluated by systematic reviews). We used small-study effect and excess significance tests to appraise reporting bias in the literature, which are the most commonly recommended and used methods [79]. However, results from these tests might also reflect methodological and clinical heterogeneity, or even chance [42]. In fact, most meta-analyses contained moderate to high heterogeneity (based on I2 statistic; S4 Table). Results from an Egger test (small-study effect) can give spurious false positive results due to correlation between log of effect size and its variance, especially in the presence of heterogeneity between studies in a meta-analysis. An alternative better-performing test has been proposed by Peters to identify reporting bias in meta-analyses, but it requires data from a 2 × 2 table [80]. Such data were rarely reported in individual studies in the meta-analyses of observational studies. As also noted by Tsilidis and colleagues [81], meta-analyses commonly use maximally adjusted relative risks rather than unadjusted relative risks calculated from 2 × 2 tables. For such data, the use of the Egger test is appropriate. The egger test and excess significance test have low power to detect reporting bias and do not give indication about what the sources of bias are. Therefore, we performed sensitivity analyses, retaining only meta-analyses with ≥10 primary studies. In this subsample of meta-analyses, evidence of reporting bias was higher than the entire sample (small-study effect: 31% versus 18%; excess significance: 27% versus 20%). Differences between primary results and sensitivity analyses are likely related to low power of reporting bias tests, which could lead to false negative results in the former group of meta-analyses. Therefore, our estimates of reporting bias in the meta-analyses are possibly conservative. The ranking of research areas according to levels of reporting bias was also different between the main analysis and the sensitivity analysis (i.e., meta-analyses with ≥10 primary studies). For instance, meta-analyses of sedentary behavior appeared most sensitive to this restriction, as the estimated proportion of reporting bias increased when calculated with either the small-study effects (from 9% to 50%) or excess significance tests (from 18% to 100%). A possible explanation could be the small fraction of meta-analyses with ≥10 primary studies (2 out of 12) in this relatively new research field [82]. It is important to acknowledge that certain methodological decisions we made may have introduced bias in the sample of reviews selected or may compromise the generalizability of our findings. We excluded systematic reviews on alcohol published in Chinese language (n = 2), which potentially have high risk of bias [83]. In addition, we restricted our analyses to systematic reviews published in this decade only (2010–2016), which explains the small number of included meta-analyses in some research areas. This may have limited comparisons of the extent of reporting bias between research areas investigated. Our results may not provide a complete historical assessment of reporting bias in these areas. Nevertheless, our results reflect reporting bias in the literature of recent and relevant public health topics and from a time period when reporting standards have been improving due to, e.g., the widespread use of various manuscript reporting checklists [84]. Recent systematic reviews contain a higher number of primary studies than older systematic reviews and synthesize evidence of emerging fields that have flourished only recently (i.e., sedentary behavior). In conclusion, we found evidence of reporting bias in approximately one-fifth of recent meta-analyses of observational studies of health-related behavior (physical activity, sedentary behavior, smoking, alcohol consumption, diet) and cardiovascular and all-cause mortality. Such a level of reporting bias may, to some extent at least, distort conclusions arising from this body of evidence. Contrarily, we found no evidence of reporting bias in meta-analyses of randomized controlled trials of statins. We searched Medline (through PubMed), Embase (i.e., excluding Medline), Cochrane Methodology Register Database, and Web of Science for systematic reviews published between 2010 and 2016. We restricted our search to recent systematic reviews for several reasons. These systematic reviews belong to a “birth cohort” of systematic reviews published after the launch of the Meta-analysis of Observational Studies in Epidemiology (MOOSE) [85] and Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) [86] guidelines and are expected to have lower risk of bias. As we were interested in comparing levels of bias across different research areas, this restriction may have reduced confounding due to date of publication. We restricted the search, as well as the successive phases of our study, to systematic reviews aiming to investigate the associations of health-related behavior (tobacco, alcohol, diet [fat, fruits and vegetables, salt, and sugar], physical activity, and sedentary behavior) and statins with cardiovascular disease mortality (overall cardiovascular mortality and cause-specific deaths from cardiovascular disease) and all-cause mortality. We accepted any definition for the exposures and the outcomes as defined in the original systematic reviews. The keywords used in the search are described in S1 Text, and files exported from databases during search strategy with all studies screened and selected are available at https://osf.io/wpb69/. Systematic reviews were screened and selected (by two reviewers, and disagreements solved by a third reviewer) based on the following eligibility criteria: (i) sought to investigate an exposure–outcome association in a nonclinical population; (ii) systematically searched for primary studies and performed a meta-analysis (i.e., weighted summary effect size) using results from primary studies; (iii) selected only observational studies (cohort and case-control studies) if a health-related behavior review and only randomized controlled trials if a statin review; (v) reported data from each primary study included in the meta-analysis (S1 Text). We decided a priori that a random sample of up to 20 systematic reviews per research area (tobacco, alcohol, diet, physical activity, sedentary behavior, and statins) would be included to compare levels of reporting bias in the relevant literature. If our search retrieved fewer than 20 meta-analyses in a given research area, we included them all. A similar study-selection strategy was recently used in a study evaluating publication bias in meta-analyses of individual studies [87]. These methods were decided a priori as described in the analysis plan available at https://osf.io/wpb69/ (not published prior to the identification and selection of systematic reviews). Reporting bias could be related to overall risk of bias in a review. Therefore, four reviewers (JPRL, NC, AF, LP), working in pairs, independently assessed the risk of bias in the included systematic reviews using the ROBIS tool [41]. ROBIS comprises three phases: (1) assess relevance; (2) identify concerns with review process; (3) judge risk of bias in the review. To assess relevance, we extracted the target question from each review using the PICOS acronym (participants, interventions, comparisons, outcomes) or equivalents for etiological questions (participants, exposure, comparisons, outcomes). In phase 2, we assessed the risk of bias in four domains related to the review process: (1) study eligibility criteria; (2) identification and selection of studies; (3) data collection and study appraisal; and (4) synthesis and findings. Questions included in each of the four domains are available in S3 Table. Questions were answered as “Yes,” “Probably Yes,” “Probably No,” “No,” and “No Information,” with “Yes” indicating low risk of bias. In phase 3, we summarized the concerns identified in each domain during phase 2 and risk of bias in the review as low, high, or unclear. Further details on the ROBIS tool are described elsewhere [41]. For each meta-analysis performed in the selected systematic reviews, we assessed the extent of reporting bias in the included literature via small-study effect [42] and excess significance tests [88]. To perform these tests, we extracted necessary data (e.g., effect size, confidence intervals, sample size, and number of events [deaths]) for each primary study included in the main meta-analysis performed in the systematic reviews. We also used these data to reperform the meta-analyses (i.e., using random effect models, which was used in the majority of the original meta-analyses). We did this to describe the number of meta-analyses with nominally statistically significant results at P < 0.05 (S1 Text). Small-study effect test (also known as regression asymmetry test, proposed by Egger and colleagues) evaluates whether smaller studies tend to overestimate the effect size estimates compared to larger studies. For this matter, the test evaluates whether the association between effect size (e.g., relative risk, odds ratio) and precision (standard error) is greater than might be expected by chance. We considered a P value < 0.10 as a statistical significance threshold for small-study effect bias (i.e., suggesting evidence of reporting bias), as initially proposed by Egger and colleagues [42, 89] and consistently used in the literature [42, 66, 81, 87, 90, 91]. Excess significance test evaluates whether the O differs from the E. The E in each meta-analysis was obtained from the sum of power estimates of each primary study. The power estimate of each primary study depends on the plausible causal effect of each research area (e.g., smoking and cardiovascular mortality), which was assumed to be the effect of the most precise primary study (smaller standard error) in each meta-analysis [88]. We considered P < 0.10 (one-side P < 0.05 for O > E) as a statistical significance threshold for excess significance bias [43, 88]. The excess significance is reported as a proportion of studies, with the higher proportion indicating more excess significance (O > E) and thus more evidence of reporting bias. Due to the low power of these bias tests, we performed a sensitivity analysis excluding meta-analyses with fewer than 10 studies to analyze the impact in the results. We also performed a sensitivity analysis excluding small individual studies (fewer than 200 deaths) within meta-analyses to evaluate whether results reflect reporting bias among small studies only. We performed all statistical analyses using Stata version 15.0 (College Station, TX).

10.1371/journal.pgen.1005395

Genome-Wide Reprogramming of Transcript Architecture by Temperature Specifies the Developmental States of the Human Pathogen Histoplasma

Eukaryotic cells integrate layers of gene regulation to coordinate complex cellular processes; however, mechanisms of post-transcriptional gene regulation remain poorly studied. The human fungal pathogen Histoplasma capsulatum (Hc) responds to environmental or host temperature by initiating unique transcriptional programs to specify multicellular (hyphae) or unicellular (yeast) developmental states that function in infectivity or pathogenesis, respectively. Here we used recent advances in next-generation sequencing to uncover a novel re-programming of transcript length between Hc developmental cell types. We found that ~2% percent of Hc transcripts exhibit 5’ leader sequences that differ markedly in length between morphogenetic states. Ribosome density and mRNA abundance measurements of differential leader transcripts revealed nuanced transcriptional and translational regulation. One such class of regulated longer leader transcripts exhibited tight transcriptional and translational repression. Further examination of these dually repressed genes revealed that some control Hc morphology and that their strict regulation is necessary for the pathogen to make appropriate developmental decisions in response to temperature.

Eukaryotic cells alter their developmental programs in response to environmental signals. Histoplasma capsulatum (Hc), a ubiquitous fungal pathogen of humans, establishes unique transcriptional programs to specify growth in either a multicellular hyphal form or unicellular yeast form in response to temperature. Since hyphae and yeast are specialized to function in infectivity or pathogenesis, respectively, Hc provides a clinically relevant system in which to query eukaryotic regulatory processes. Here we used next-generation sequencing approaches to annotate the transcriptomes of four distinct Hc strains in response to temperature. We found that a fraction of Hc transcripts have differential transcript architecture in hyphae and yeast, exhibiting 5’ leader sequences that differ markedly in length between morphogenetic states. To begin to understand the effect of these differential leader sequences on expression, we performed the first ribosome density and mRNA abundance measurements in Hc, thereby uncovering transcriptional and translational control that contribute to cell-type regulation.

Environmental human pathogens have evolved the ability to survive in human hosts as well as diverse environmental reservoirs. Thus a hallmark of environmental pathogens is their capacity to adapt to varied growth conditions such as differences in temperature, alterations in nutrient sources, as well as exposure to the host immune system. The dimorphic human fungal pathogen, Histoplasma capsulatum (Hc), is one such environmental pathogen that responds to an increase in temperature inside a mammalian host by dramatically altering its cellular morphology and gene expression programs to cause disease. The environmental, infectious form of Hc grows in the soil as a saprophyte in a multicellular hyphal or “mold” form that produces vegetative spores called conidia [1,2]. Infection occurs when humans inhale conidia or hyphal fragments. Upon exposure to mammalian body temperature, Hc transitions into a unicellular, budding yeast (Fig 1A) that is capable of causing primary disease in immunocompetent individuals [3–6]. The temperature-regulated differentiation of Hc between the hyphal and yeast forms can be recapitulated in the laboratory simply by switching the temperature from room temperature (RT) to 37°C, making Hc a unique organism for studying the regulation of gene expression during multicellular development, environmental signal transduction, and adaptation of a pathogen to a mammalian host. Hc cells respond to temperature through a transcriptional regulatory circuit that controls cell morphology as well as the expression of 5–15% of the Hc transcriptome that is differentially regulated between yeast and hyphal cells [8–10]. In addition, transcripts encoding many of the known Hc virulence factors are enriched in the yeast-phase of growth [11], making yeast-enriched transcripts intriguing virulence factor candidates. To this end, microarrays [8,9,12–14] and RNA-sequencing (RNA-seq) [10] have begun to elucidate the transcriptomes of yeast, hyphal, and conidial cell types to describe transcript expression patterns that specify these three major cell types important for the Hc pathogenic lifecycle. Despite these efforts, our understanding of the Hc transcriptome and cell-morphology specific regulatory programs remains incomplete partly due to challenges inherent in deciphering intron-rich, densely populated eukaryotic fungal transcriptomes. To date, Hc transcriptomics studies have been performed using two North and Central American Hc var. capsulatum strains, G217B and G186AR (see Fig 1B). Yet, Hc is found globally (North, South, and Central America, Southeast Asia, and Africa) and phylogenetic analysis of Hc clinical isolates has revealed that Hc comprises at least 8 genetically and geographically distinct lineages that span 3–13 million years of evolutionary distance [7,15] (Fig 1B). In addition to its genetic diversity, Hc can be segregated into two varieties (var.) based on clinical outcomes: Hc var. duboisii and Hc var. capsulatum. Hc var. capsulatum is represented in all of the major Hc lineages and infection with this variety predominantly manifests as pulmonary and systemic disease [4]. Hc var. duboisii, conversely, is geographically restricted to Western and Central Africa and causes cutaneous and subcutaneous skin as well as bone lesions in a systemic infection that is distinguished in name from classical disease as African histoplasmosis [16,17]. The virulence mechanisms and basic biology of the causative agent of African histoplasmosis, Hc var. duboisii, as well as Hc var. capsulatum strains from additional, more divergent lineages, remain uninvestigated. In this work, we employ recent advances in next-generation sequencing and de novo transcriptome reconstruction methodologies to refine transcript architecture and define transcript expression programs across divergent isolates of Hc that cause classical (G217B, G186AR, H143) or African (H88) histoplasmosis. Using comparative transcriptomics of two temperature-regulated states, we broaden our understanding of the basic biology and pathogenesis of Hc by defining core yeast (parasitic, disease-causing form) and hyphal (infectious form) transcript expression patterns that exhibit conserved regulatory patterns. Systematic analysis of our improved Hc transcript architecture further revealed that a subset of the Hc transcriptome (~ 2%) exhibited dramatically different leader sequence lengths (often referred to as the 5’ untranslated region; UTR) between these two Hc developmental cell states. To begin to understand the biological significance of longer leader transcripts, we probed their ribosome occupancy using ribosome profiling. Consequently, we defined several categories of genes, including those that are regulated at both the transcriptional and translational level in response to temperature-initiated developmental programs. Phenotypic analysis revealed that this exquisite regulation is necessary for appropriate cell-type specification by temperature, the key signal for infectious and parasitic states. To compare yeast and hyphal transcript expression patterns across Hc lineages, we grew each Hc strain (G217B, G186AR, H88, H143) at 37°C for yeast-phase growth and at room temperature (RT) for hyphal phase growth (Fig 1C). Total RNA was isolated from biological duplicates of yeast or hyphal cultures and strand specific cDNA libraries were created from poly(A)-enriched RNA and sequenced using paired-end deep-sequencing. The 4 Hc strains that we selected for transcriptomics analysis each have sequenced genomes; however, the Hc gene predictions vary among strains and often lack accurate models of 5’ leader and 3’ UTR regions. Therefore we constructed de novo transcript calls leveraging our paired-end RNA-seq data with the aim of improving transcript models, generating high confidence ortholog mapping across Hc strains, and directly identifying poly(A)+ transcribed regions. To do so we combined yeast and hyphal sequencing reads for each Hc strain and used these pooled reads to generate per-strain reference transcriptomes across the 4 Hc lineages (see Fig 2A; S1 Table; S1–S12 Data; Materials and Methods). Hc RNA species without polyadenylated 3’ ends (i.e., potential small non-coding RNA) were not enriched in our sequencing libraries and thus are likely not represented in these transcript models. Our genome-guided transcriptome assembly pipeline yielded 12, 175–12, 889 transcripts for the 4 Hc strains with 9, 580–9, 844 of assembled transcripts predicted to be protein coding (Table 1). This falls within the range of 9, 229–11, 329 protein coding genes predicted from Hc genomic sequences by ab initio gene prediction algorithms (Table 1). These Hc transcript models do not take into account any alternate transcript isoforms that may exist within or between Hc cell types (i.e., transcript forms that arise due to intron retention); however this is an interesting area for future refinement and investigation. The assembled transcripts were overall larger in size than the predicted transcript set (S1 Fig), which is mostly due to the improved annotation of 5’ leader and 3’ UTR transcript ends (Fig 2B and 2C). In addition to augmenting 5’ and 3’ transcript end annotations, we noticed many cases where assembled transcripts had more accurate intron boundaries than predicted transcripts resulting in improved annotation of coding sequence (CDS) regions (S2 Fig). In order to compare transcription across strains, groups of orthologous genes among the 4 Hc strains (orthogroups) were determined using two independent methods, Mercator [18] and InParanoid [19], which gave similar numbers of per-strain ortholog assignments (Table 1). For downstream analyses, we chose the Mercator method for determining Hc orthogroups since it incorporates both BLASTP [20] homology and genomic synteny for a more stringent determination of orthogroups. Using Mercator we assembled 6, 791 unique Hc orthogroups that were present in all four strains (see Materials and Methods; Table 1). Genes not assigned to orthogroups by this method could include in-paralogs (gene duplications that occurred after strain divergence), improperly assembled transcripts that led to inappropriate open reading frame (ORF) predictions, and genes truly unique to a given strain. Of the 6, 791 Hc orthogroups, 423 (6.2%) were not in any of the previously predicted Hc transcript sets and exhibited a striking pattern of short CDS regions relative to overall transcript lengths (S3 Fig). Only a small percentage of these novel orthogroups had conserved Pfam-A domains (4%) [21] or predicted secretion signals (9% as determined by Phobius [22,23]), and they exhibited lower transcript expression levels in comparison to the full orthogroup set (S4 Fig). In other eukaryotic species, poly(A)+ long non-coding RNAs (lncRNAs) often have lower expression levels than protein-encoding transcripts [24]. Thus, this set of novel orthogroups in Hc may include cases of poly(A)+ non-coding RNA species with spurious ORF predictions as well as small proteins or peptides missed by ab initio genome prediction algorithms. Overall, however, 90% of Hc orthogroups were predicted to encode at least one Pfam-A domain [21] indicating that the vast majority of Hc orthogroups determined from our RNA-seq transcriptome assemblies represent bona fide protein coding transcripts. We used eXpress [25] to determine transcript expression levels across yeast and hyphal cell morphologies for the 4 Hc strains. eXpress estimates of FPKM values (fragments per kilobase of exon per million aligned fragments) were well-correlated between biological replicates (S2 Table); therefore we took the mean of log2 FPKM values of biological replicates for downstream analyses of Hc yeast and hyphal expression patterns (S13 Data). To define a global picture of evolutionarily conserved programs of gene regulation across Hc lineages, we identified transcripts enriched in the yeast (parasitic form) or hyphal (infectious form) state from the 6, 791 Hc orthogroups (S5 Fig; S13–S15 Data). Minor lineage-specific differences in gene expression were observed, though these were largely of unknown biological significance. In terms of conserved patterns of gene expression, we found 139 yeast-phase enriched transcripts (core yeast-phase transcripts: Y/H log2 ≥ 1.5) and 291 hyphal-phase enriched transcripts (core hyphal-phase transcripts: Y/H log2 ≤ - 1.5) conserved in differential expression pattern among the 4 Hc strains (S5B and S5C Fig; S13 Data). Core yeast-phase enriched transcripts included two characterized virulence factors (CBP1, SOD3) [26,27], siderophore biosynthesis and transport genes (SID3, SID4, ABC1) important for yeast cell iron acquisition [28], and 6 putative transcription factors (CSR1, CHA4, MEA1, RYP1, RYP4, XBP1) including RYP1 and RYP4, which are master transcriptional regulators of Hc yeast-phase growth [8,9]. We also identified 8 cell wall modifying enzymes (AMY2, CFP8, CTS1, DCW1, ENG1, GEL2, OCH1, SKN1) with transcripts upregulated in the yeast phase that are predicted to be involved in the biosynthesis or remodeling of cell wall polysaccharides such as β-glucan, chitin, and mannan; the upregulation of these transcripts in yeast may reflect differences inherent to yeast versus hyphal cell division and growth. Core hyphal-phase enriched transcripts included 18 putative transcription factors (see S13 Data), many of which are uncharacterized and are members of the fungal-specific Zn2C6 DNA binding domain family of transcriptional regulators (Pfam Accession: PF00172). In addition, we noticed that many core hyphal-phase enriched transcripts encoded enzymes such as cytochrome p450s, polyphenol oxidases (also known as tyrosinases), oxidoreductases, and peroxidases (see S13 Data). The hyphal-phase enrichment of these enzymes, which are often involved in nutrient acquisition or the production of toxins, melanin, and other secondary metabolites, is likely reflective of the saprophytic lifestyle of Hc hyphal cells. Examining the relative abundances of yeast-enriched transcripts (using per-state mRNA FPKMs) indicated that some of the most abundant core yeast-enriched transcripts (CBP1, GH17/CFP4, SOD3, ENG1) encode proteins known to be secreted by yeast cells [29] (S6 Fig). Since bacterial and fungal pathogens often utilize secreted proteins as virulence effectors [30,31], we were interested in identifying additional conserved yeast-phase enriched transcripts encoding proteins with secretion signals. Prediction of secretion signal peptides for the conserved yeast-phase enriched ORFs (using Phobius [22,23]) revealed that secreted proteins were significantly more likely to be differentially expressed than non-secreted proteins (p = 0.000314), suggesting that many secreted proteins could have phase-specific roles. In addition, we noticed that there were many small (≤ 200 AAs) putative secreted proteins (Figs 3A and S6) that exhibited a conserved C-terminal, 6-cysteine spacing pattern reminiscent of some insect toxins [32]. We examined existing hidden Markov models (HMM) of cysteine-rich protein domains to determine whether this Hc 6-cysteine motif belonged to any known protein families. Through this analysis, we found that many Hc ORFs with the 6-cysteine motif had homology to a cystine knot (or knottin) gene family [34]. Knottin domains are comprised of 3 interwoven disulfide bonds that form one of the smallest known stable globular domains [35], making these proteins extremely resistant to chemical, heat, and proteolytic stresses [36]. Knottins can be found in fungi, insects, plants, and animals [34], and their 3-disulfide bond core can present constrained loop structures that mediate protein-protein interactions [37]. Functionally, knottin proteins have been shown to act as pore formers, ion-channel inhibitors, as well as protease inhibitors [38–40]. Phylogenetic analysis of proteins with identifiable knottin domains (of the Fungi1 knottin family [34,41]) in Hc as well as other related fungal species, revealed that this family appears to have undergone an expansion in the Ajellomycetaceae family of human fungal pathogens (Hc, Paracoccidioides brasiliensis, Blastomyces dermatitidis) (Figs 3B and S7; S16–S18 Data). Notably, knottins are absent from many Saccharomycetes fungi such as Saccharomyces cerevisiae and Candida albicans (S7 Fig). Hc transcripts encoding this knottin domain are predominantly yeast-phase enriched (S8 Fig) and are intriguing candidates for virulence effectors that could mediate host-pathogen interactions. While examining our improved annotation of transcript boundaries in the 4 assembled Hc transcriptomes, we noticed that a subset of transcripts exhibited differences in the size of their leader regions between yeast and hyphal cell types. For example, we observed that some transcripts in yeast cells displayed longer leader regions as compared to their cognate hyphal transcripts (Fig 4). This was a phenomenon that we had previously seen for a handful of Hc yeast-enriched transcripts while examining transcript abundance and architecture in G217B yeast and hyphal cells using Northern blotting and 5’ RACE (5’ rapid amplification of cDNA ends) [13]. Given the high-resolution nature of RNA-seq data, we were able to identify leader regions that were differential in length between Hc morphologies genome-wide across all 4 strains. To do so, we employed our existing transcript assembly pipeline to determine per-cell type (yeast and hyphae) transcript structures for each Hc strain (see Materials and Methods; S19–S26 Data). Globally, we found that most leader regions were similar in length between yeast and hyphae in all 4 Hc strains as expected (S9 and S10 Figs). However, systematic analysis of leader regions using per-cell type assembled transcript structures identified 187 transcripts with conserved, differentially sized leader regions between yeast and hyphal cells (see Materials and Methods; S10 Fig; S13 Data). The conservation of differential leader transcript structure in all 4 Hc lineages examined suggests that changes to 5’ transcript length are not a stochastic transcriptional outcome, but are evolutionarily conserved and likely regulated by the cell. We found that the majority of transcripts that exhibited differential leader length are longer in the yeast phase compared to the hyphal phase of Hc growth (138/187 differential leaders; S10 Fig; S13 Data). We did, however, find 49/187 conserved transcripts with longer leaders in hyphal cells (S10 Fig; S13 Data), indicating that differential leader length is not an artifact of sample preparation or produced exclusively by one Hc cell type. We examined the sizes of a subset of longer leader transcripts in yeast and hyphal cells via Northern blotting, which confirmed the differential transcripts sizes between Hc cell morphologies (S11 Fig). Northern blot analysis of longer leader transcripts also highlighted that some of these transcripts represent a population of transcripts of variable sizes (S11 Fig). This is analogous to emerging observations of complex transcript structure in S. cerevisiae where variations in 5’ and 3’ transcript boundaries give rise to many different transcript isoforms for a given gene [42,43]. Differential sizes of transcript leader regions between cell states or in cells encountering changing environmental conditions is a recently observed genome-wide phenomenon in eukaryotes [44–48]. While the biological function of differential leader regions is often unclear, changes in leader length have been proposed to influence the translational outcome of protein coding ORFs either by introducing regions of RNA that can serve to regulate translation (e.g., upstream ORFs; uORFS) [44] or by introducing an alternate start codon that encodes a new protein variant [42]. Thus we hypothesized that differential leader transcripts in Hc may be translationally regulated between Hc developmental states. To investigate this possibility we adapted ribosome profiling to Hc. Ribosome profiling is based on deep-sequencing of ribosome-protected RNA fragments or “footprints” and has been well-established as a method to probe translational regulation in eukaryotic cells [44,49–51]. Analogous to ribosome profiling developed in the model yeast organism S. cerevisiae [49], we digested total RNA from yeast or hyphal G217B Hc cells with RNase I and then separated and fractionated the RNA with a sucrose density gradient to collect Hc 80S monosomes (Fig 5A and 5B). Polysome profiles of undigested RNA from yeast and hyphal vegetative cells were noticeably different (Fig 5A and 5B), most likely illustrating the distinct modes of growth for these Hc cell types. Hc yeast grow and divide as unicellular budding yeast cells while Hc hyphal cells are connected by septa to form filamentous, branched hyphae that grow as highly polarized multicellular structures. In Ascomycete fungi, hyphal cells exhibit specialization and often possess differences in growth rate, cell size, or access to nutrients [52], leading to a heterogeneous mix of cell states per hyphae. We collected biological replicate samples of yeast and hyphae monosomes as well as poly(A)-enriched RNA from the same samples for ribosome profiling and RNA-seq, respectively. Ribosome footprint density and mRNA abundance measurements (calculated as the number of ribosome-protected or mRNA fragments per kilobase of coding sequence per million mapped reads; FPKM) were reproducible between yeast and hyphal biological replicates (S12 Fig). With ribosome profiling data we can also determine the translational efficiency (TE) of transcripts (calculated as the ratio of ribosome footprint density to mRNA abundance over a coding region), which is a measure of the extent to which a transcript is translationally regulated. Importantly, measurements of TE in Hc were reproducible between biological replicates (Fig 5C and 5D; S27 Data) and TE measurements for genes expressed in yeast or hyphal cells spanned similar orders of magnitude (S13A Fig). Additionally, ribosome footprint FPKMs for transcripts involved in basic cellular functions thought to be important for growth or maintenance of both cells types (ACT1, GAPDH, TEF1, CDC2, TUB1, and TUB2) were overall remarkably similar between yeast and hyphae (S13B Fig; S27 Data). However, we want to stress that the ribosome profiling data sets provide relative measures of translational efficiencies and not absolute measures of translation. Together these data indicate that ribosome profiling is a robust method to probe genome-wide ribosome occupancy as well as the translational regulation of transcripts in the developmentally distinct morphologies of Hc. Transcriptome-wide ribosome occupancy measurements did not correlate perfectly with mRNA abundance in Hc yeast or hyphae (Fig 6A and 6B). This suggested that expression of a subset of Hc genes is controlled at the level of translation and thus could reflect cell-type specific translational regulation between Hc cell morphologies. To investigate this, we first examined ribosome footprint density measurements to identify transcripts with different levels of translation between yeast and hyphal cells. Through this analysis we found subsets of transcripts that were predominantly translated only in one cell type (yeast = 200 transcripts and hyphae = 490 transcripts; Fig 6C; S27 Data) and reasoned that these transcripts encoded proteins with biological functions that are specific to a given cell type. Since protein abundance in a cell is influenced by both transcript levels as well as the rate at which transcripts are translated, we also examined the translational efficiency of transcripts in yeast and hyphal cells. Clustering of yeast and hyphal TE values for transcripts that exhibited translational regulation highlighted at least 6 categories of translational regulation between yeast and hyphae (Fig 6D; S28–S29 Data). Most noticeably, we identified a large class of genes with low translational efficiency in both yeast and hyphae (Fig 6D- Category B); not surprisingly many transcripts in this category encode proteins that are not expected to serve a function in Hc vegetative yeast or hyphae under the growth conditions used in these experiments. For example, genes predicted by homology to be involved in mating/meiosis (HOP1, MEI2, STE2, STE6), light sensing during circadian rhythm (WHC2), as well as conidial (asexual spore) development and biology (CATA, RDS1, CON132) showed low levels of translational efficiency in both yeast and hyphal cells. We also found many transcripts with robust translational efficiency in both cell types (Fig 6D- Category E) that included genes involved in routine cellular functions such as fatty acid biosynthesis (ACC1, FOX2, FAA1), cytoskeletal organization (ACT1, BUD6, ARP1, SLA1), and nuclear and mitochondrial transport (KAP95, TOM20, KAP108, TIM16). Categories where we observed translational efficiency differences between yeast and hyphae (Fig 6D- Category A, C, D, F) included some previously identified yeast and hyphal-enriched transcripts, indicating that a subset of cell-type enriched transcripts are regulated at both the transcriptional and translational level by the temperature-dependent developmental program. Having established ribosome profiling as a method to measure ribosome occupancy and translational efficiency in Hc yeast and hyphae genome-wide, we next examined the translational outcomes of differential leader transcripts. Comparing yeast and hyphal TE values for the set of longer leader transcripts revealed that many exhibited altered translational regulation compared to their cognate shorter transcript isoforms (Figs 7 and S14; S30–S33 Data). Correlation of ribosome footprint densities on differential leader regions with changes in TE values (by examining a 2 or 1.5 fold change in CDS TE) between short and long transcript forms resulted in classification of leaders into broad regulatory patterns (Figs 8 and S15). The most striking pattern we observed was differential leader ribosome density that correlated with a decrease in TE, reminiscent of uORF mechanisms of translational repression (Fig 8- Category a). We also observed longer leaders with no ribosome density and strong repression of downstream CDS translation (Fig 8- Category d) via unknown mechanisms. In addition, we found examples of robust leader ribosome occupancy that resulted in no change to the TE, which may represent cases of N-terminal protein extensions and uORFs that don’t influence translation of the CDS (Fig 8- Category b). Together these data suggest that as many as half of the observed longer leaders may be involved in translational repression of the associated CDS. The features distinguishing such regulatory leaders remain to be elucidated experimentally. While examining longer leader transcripts that exhibited strong repression of CDS translation (Fig 8- Category a, d), we noticed that many of these transcripts had longer leaders in the yeast phase and were additionally regulated at the level of transcription between yeast and hyphae (Fig 7A). Specifically, a subset of yeast-phase longer leader transcripts were transcriptionally and translationally repressed in yeast cells, yet were robustly transcribed and translated in hyphae (Fig 7A- red box). To validate whether these ribosomal profiling data corresponded to levels of protein accumulation, we expressed GFP under the control of the MS95 longer leader transcript promoter and leader region (MS95p) in Hc cells and examined GFP protein production at 37°C and RT. MS95 has a longer leader transcript form in yeast cells that exhibits lower translational efficiency compared to its shorter hyphal form (S16A Fig). Assessment of GFP protein levels by Western blot in Hc cells expressing the MS95p – GFP construct and grown at 37°C or RT recapitulated the differential protein expression pattern we inferred from ribosome profiling (S16B and S16C Fig). Namely, the MS95 upstream region restricts expression at 37°C, but allows robust transcription and translation at RT. Why Hc cells tightly restrict expression of Ms95 at 37°C is unclear as the biological function of Ms95, a homolog of the DNA damage and heat-stress responsive protein Ddr48 in S. cerevisiae, remains unknown. Much precedence exists for cells to employ translational regulation to achieve tight spatial or temporal control of protein expression during developmental processes [53]. In this vein, we noticed that some longer leader transcripts have been implicated in governing Hc cell fate decisions. Both RYP2 and RYP3, master regulators of yeast cell morphology, have longer leader regions in hyphal cells. RYP2 has ribosome density in its differential leader region and lower TE of the RYP2 CDS in hyphae versus yeast (S17A Fig). The role of the RYP3 hyphal longer leader region is unclear as there is only a slight reduction in TE of the RYP3 CDS in hyphae versus yeast (S17B Fig). Ryp2 and Ryp3 both associate with DNA to regulate transcription of yeast-phase enriched transcripts at 37°C; furthermore, disruption of RYP2 or RYP3 in Hc cells results in inappropriate hyphal growth of Hc at 37°C [9,54]. Thus, tight restriction of Ryp levels during cellular development may be a mechanism used by Hc to control appropriate morphology in response to temperature. We mined Hc longer leader transcripts for additional candidates that could be involved in regulating cell fate decisions by identifying longer leader transcripts with transcription factor domains or homology to developmental regulators in other fungal species. This approach identified 14 putative developmental regulators with longer leader regions in Hc yeast or hyphal cells (S3 Table). One putative Hc developmental regulator we identified was WET1, which we named based on its homology to the Aspergillus nidulans regulator of conidial (asexual spore) development, WetA [13,55]. WET1 has a longer leader sequence in Hc yeast cells and its CDS region is translationally repressed in yeast while the shorter hyphal transcript form has robust ribosome CDS occupancy in hyphae (Fig 9A). To explore the idea that regulators of Hc developmental programs include longer leader transcripts that are tightly regulated between yeast and hyphal cell types, we placed the CDS of WET1 under the control of a heterologous Hc promoter, ACT1p (Fig 9B). Importantly ACT1 is not differentially translationally or transcriptionally regulated between Hc yeast and hyphae. Ectopic expression of this ACT1p – WET1 construct in Hc yeast cells resulted in inappropriate hyphal growth of Hc at 37°C (as compared to vector control cells; Fig 9C and 9D), while expression of the ACT1p – WET1 construct had no discernible effect on Hc hyphal morphology at RT (as compared to vector control cells; Fig 9D). These data indicate that restricting Wet1 expression in yeast-phase cells is critical for maintenance of the yeast cell developmental program and that Hc longer leader transcripts are interesting candidates for Hc developmental regulators. Here we present improved transcript models for multiple clinical Histoplasma isolates as well as the first look at the translational landscape of this medically relevant human fungal pathogen. Through this work we find that Hc alters transcript leader length between its morphologic cell types for a subset of its transcriptome. Transcript leader length is a biologically regulated outcome of transcription in Hc as our work uncovers many examples of leader extensions (187 transcripts) that are evolutionarily conserved across Hc phylogenetic lineages. Ribosome and mRNA density measurements of longer leader transcripts reveal a class of genes that are under tight translational and transcriptional regulation. Further examination of this group of transcriptionally and translationally regulated genes indicated that some are involved in controlling Hc morphology and that their strict regulation may be necessary for appropriate developmental decisions. Our ribosome profiling experiments suggested mechanisms of translational regulation for longer leader transcript CDS regions. Namely, we observed cases where ribosome density on longer leader regions correlated with reduced translational efficiency of the downstream CDS, suggesting that this class of transcripts may be translationally repressed via well-studied mechanisms of uORF-mediated translational control [56]. Alternatively, we observed cases of translationally repressed CDS regions of longer leader transcripts with no evidence of ribosome occupancy on the leader, indicating that translational repression may be regulated via uORF-independent mechanisms. Future work will be needed to examine mechanisms of translational repression for this class of transcripts such as changes to mRNA secondary structure of the leader that inhibit ribosome scanning or cell-type specific factors that regulate ribosome function. For example, leader sequences could specify subcellular localization of mRNAs, which could inhibit (or facilitate) translation. Lastly, we identified longer leader transcripts that exhibited ribosome footprint density in the leader region directly upstream of the predicted CDS with little effect on translational efficiency. We suspect that some of these longer leader transcripts could serve to encode alternate protein variants and thus represent cases where leader extensions increase proteome diversity. We must also emphasize that some of our longer leader transcripts appear to be part of a complex mixture of transcript isoforms of varying sizes (as seen by Northern blot; see S11 Fig), which will complicate future mechanistic analyses. The mechanism by which differential leader sequences are specified by distinct cellular states is yet to be determined. One of several possibilities is that a particular transcript initiation site may be occluded in one state by protein complexes that associate with the DNA. Intriguingly, we observed that 13 of the 49 transcripts with a longer leader sequence in the hyphal phase (including Ryp2; see S17A Fig) show association of the corresponding DNA with a yeast-enriched protein complex (p = .0015, given 785 total Ryp associated transcripts). Specifically, in the yeast-form at 37°C, these genes are bound by the Ryp proteins [9], which are yeast-enriched transcription factors that are required for yeast-phase development. It is possible that the physical presence of the Ryp complex on the DNA at 37°C could interfere with production of the longer-leader transcript; consequently the longer transcript isoform appears only in hyphal cells at room temperature. This model is appealing because it ties the presence of state-specific transcription factors to state-specific transcript architecture. It is accepted that metazoans employ translational regulation to precisely tune gene expression during developmental processes [53]. Additionally, pathogens can regulate virulence factors at the translational level using host cues, such as temperature, to induce virulence factor expression once in the host environment [57,58]. Thus, translational control is a mechanism that can be used to quickly or precisely regulate gene expression during cellular development and pathogenesis. However, the biological function of many of the translationally regulated longer leader transcripts in Hc is unknown, making the context and significance of their translational regulation difficult to deduce at this time. We hypothesize that longer leader transcripts that are translationally repressed in the yeast phase could play a role in regulating hyphal growth or promoting morphogenesis in response to a change in temperature from 37°C to RT. Repression of such genes during yeast-phase growth would be critical to prevent inappropriate hyphal growth, which we have demonstrated for the longer leader transcript WET1. Alternatively, longer leader transcripts that are translationally repressed in hyphae could be tightly regulated for appropriate expression during Hc pathogenesis (since yeast cells are the disease-causing host form of Hc). Ultimately, a better understanding the biological function of translationally regulated longer leader transcripts will inform ideas for how Hc cells use translation (or appear to use changes in transcript structure) to regulate expression of a subset of the transcriptome. In addition to uncovering examples of extended leader transcripts that exhibit distinct modes of translational regulation, our work provides a better foundation for understanding the biology and pathogenesis of the human pathogen Histoplasma. For example, our discovery of an expansion of knottin family proteins was enabled by the identification of small transcripts that show conserved enriched expression in the parasitic phase of this organism. Additionally, as with many organisms, it is becoming apparent that the regulation of Hc gene expression and transcript architecture is more complex and nuanced than previously appreciated. For example, our work demonstrates that both transcriptional and translational regulation contribute to the yeast and hyphal developmental programs. By deciphering the transcript structures, transcript expression patterns, as well as identifying novel orthogroups that likely encode short peptide products and regulatory RNAs across Hc lineages, we have laid the groundwork for a deeper understanding of the Hc strains that cause a wide variety of disease complications in the human host. Histoplasma capsulatum (Hc) strains G217B (var. capsulatum; ATCC 26032), G217Bura5Δ (var. capsulatum; WU15), G186AR (var. capsulatum; ATCC 26029) were all gifts from the laboratory of William Goldman, University of North Carolina, Chapel Hill. H88 (var. duboisii; ATCC 32281) and H143 (var. capsulatum; CBS 287.54) were obtained from the American Type Culture Collection and Centraalbureau voor Schimmelcultures, respectively. Hc strains were routinely grown in HMM (Histoplasma-macrophage medium) broth or plates [59] or Sabouraud dextrose (Difco, BD, San Jose, CA) agar plates. Media was supplemented with 200 μg/mL of uracil (Sigma-Aldrich, St. Louis, MO) where indicated. Hc cultures were grown at 37°C under 5% CO2 for yeast-phase growth or at room temperature (RT) for hyphal-phase growth with continuous shaking of liquid cultures on an orbital shaker. G217B, G186AR, H88, and H143 yeast cells were inoculated from HMM agarose plates into HMM liquid medium for yeast-phase growth. Yeast cells were passaged at a 1:25 dilution three times and 2-day cultures of each strain were harvested for RNA collection after the third passage. For hyphal cells, G217B, G186AR, H88, and H143 hyphae were inoculated from Sabouraud dextrose agar plates grown at RT into HMM liquid medium. Hyphal cells were grown for 4–6 weeks with passaging 3 times (1:5 dilution) into fresh HMM medium at RT before reaching a sufficient density of cells for harvesting. Hyphal and yeast cells were collected by centrifugation or filtration and total RNA was isolated using a guanidine thiocyanate lysis protocol as previously described [13]. Paired-end RNA-seq libraries were made from biological duplicate cultures of G217B, G186AR, H88, and H143 for both yeast and hyphae. DNA was depleted from 10–20 μg of total RNA by 1 hour of DNaseI digestion at 37°C (Ambion by Life Technologies, Carlsbad, CA). Poly(A)+ RNA was selected from DNA-depleted RNA on oligo-dT25 DynaBeads (Invitrogen by Life Technologies) using 2 rounds of selection per the manufacturer’s instructions. Approximately 50 ng of poly(A)+ RNA was used as input for generating sequencing libraries with the ScriptSeq v2 RNA-Seq Library Preparation Kit (Epicentre, Madison, WI) per the manufacturer’s instructions with the following modifications. After 11 cycles of the ScriptSeq v2 PCR cDNA amplification step, cDNA libraries were resolved on an 8% TBE gel (Invitrogen by Life Technologies) and a 350–500 bp range of cDNA library was excised, eluted, and precipitated from the gel. Seven additional PCR cycles were resumed following the ScriptSeq v2 Kit protocol. cDNA libraries were purified with AMPure XP beads (Beckman Coulter, Brea, CA) to deplete primers. Libraries were multiplexed and subjected to 100 bp paired-end sequencing using the Illumina HiSeq2000 sequencer (Illumina, San Diego, CA). Throughout this work, the following sources of Hc genomic sequences and gene predictions were used: version 2 of the H88, H143, and G168AR genome assemblies and predicted gene sets from the BROAD Institute, retrieved on 6/15/2011 from http://www.broadinstitute.org/annotation/genome/dimorph_collab/MultiDownloads.html and the 11/30/2004 version of the G217B genome assembly and 9/21/2005 predicted gene set from the Genome Sequencing Center at Washington University (GSC) as mirrored at http://histo.ucsf.edu/downloads/. We used a genome-guided approach to assemble Hc transcriptomes independently for each of the 4 Hc strains from our paired-end sequencing data. The bulk of our analysis was based on combining reads from all yeast and hyphal biological replicates in a given strain. For the longer leader analysis, we also assembled state-specific transcriptomes using only yeast or hyphal reads. In all cases, the same assembly pipeline was applied, as outlined in Fig 2A and described below. Paired-end reads were pre-processed to remove high copy sequences that were not of interest; specifically, all read pairs were searched against the mitochondrial genome, a representative full length MAGGY transposon, and the 45S rDNA locus using bowtie2 version 2.0.0-beta7 [60] with a minimum fragment length (-I) of 50 and a maximum fragment length (-X) of 800. Read pairs that did not align to the above sequences were aligned to the appropriate genome assembly (GSC assembly for G217B, BROAD assembly for G186AR, H88, and H143) using GSNAP [61] with parameters based on the alignReads.pl script from TRINITY [62] (allowing novel splice sites (-N 1), up to 20 alignments per pair (-n 20), and up to 10 kb introns (-w 10000)). Transcribed fragments (transfrags) were initially defined as contiguous genomic regions with GSNAP-based read coverage > = 8 reads/base. Initial transcripts were then assembled by joining transfrags separated by no more than 2300 bases and spanned by at least 6 single reads or read pairs. Introns were annotated as GSNAP-identified splice sites falling within the initial transcripts. Where multiple splice sites spanned the same genomic location, only the splice site supported by the greatest number of aligned reads was annotated as an intron. Transcripts were identified as antisense artifacts and removed if they met either of the following criteria: 1) all splice junctions in the transcript were CT-AC (antisense to the canonical GT-AG); 2) the transcript was completely spanned by a higher abundance transcript on the opposite strand and contained either no introns or at least one non-canonical splice junction. Initial coding sequences (CDS) were then annotated as the largest open reading frame (ORF) in each transcript, with a minimum ORF size of 60 codons. This initial assembled transcript set contained fusions due to adjacent, same-strand transcripts separated by intergenic regions that were small or non-existent (e.g., due to overlapping 3' and 5' ends). We addressed this assembly artifact by identifying initial transcripts spanning multiple gene predictions. Specifically, we identified all ORFs greater than 60 bp within each initial transcript, then looked for same-strand overlaps between these ORFs and the upstream predicted protein sets (from the GSC or the BROAD); where an initial transcript contained ORFs that mapped independently to at least two different sets of predicted proteins, such that an untranslated subsequence of the transcript divided the two sets without overlapping either, the transcript was split at the midpoint of the dividing subsequence, and the CDS of the resulting transcripts were reannotated as the largest ORF of at least 60 bp. For the pooled transcriptome assembly, transcripts were named as ucsf_hc.01_1.(strain).(gene_id), where strain is G217B, G186AR, H88, or H143 and gene_id is a five digit, zero padded integer. For the state-specific transcriptome assemblies, transcripts were named as ucsf_hc.01_1.(strain)(state).(gene_id), where state is Y or H for yeast and hyphae, respectively, and the remaining fields are as above. gene_id is unique within a given transcriptome assembly, and there is no deliberate correspondence among gene_id values from different assemblies. For ortholog assignment by InParanoid [19], InParanoid version 1.35 was run with default parameters and no outgroup for each pair of Hc strains, using either predicted or assembled protein sequences as inputs. For orthologroup assignment by Mercator [18], Mercator version 0.4 (compiled from commit 991c85a of the cndsrc git repository: http://www.biostat.wisc.edu/~cdewey/software/cndsrc.git) was run in draft mode on the 4 Hc genome assemblies using all-against-all BLASTP searches of the assembled protein sets as anchors. This resulted in 5509 complete orthogroups (containing one gene from each strain) and 2356 incomplete orthogroups. In order to find orthologous genes present in the transcript assemblies but missing from the BLASTP searches due to inappropriate CDS annotations, protein sequences from incomplete orthologs were searched against the genome assemblies of missing strains using TBLASTN [20] with soft masking and an expect threshold of 1e-6, then finding the assembled transcript with greatest same-strand overlap to the top TBLASTN hit; transcripts detected this way were added to the existing orthogroups, as long as TBLASTN searches with each member of the orthogroup did not detect different transcripts and multiple orthogroups did not detect the same transcript. Application of this TBLASTN-based strategy resulted in an additional 1282 complete ortholog groups, for a total of 6791. BLAST runs for both InParanoid and Mercator were carried out with version 2.2.26 of NCBI BLAST Relative abundances (reported as FPKM values [63,64]) for each transcript in each sample were estimated by aligning read pairs to the transcriptome assembly for the corresponding genome with bowtie2 version 2.0.0-beta7 [60] with a maximum fragment length (-X) of 800, allowing all alignments (-a), and submitting the output to eXpress version 1.3.1 [25] specifying strand-specific read pairs (—fr-stranded). Transcripts with FPKM values ≤ 1.0 in yeast or hyphae were clipped to a value of 0. FPKM Y/H values were median normalized per strain using transcripts with FPKM values ≥ 10.0 Assembled CDS were annotated with Pfam domains by searching all protein sequences against version 27.0 of Pfam-A [21] with hmmscan from HMMer 3.0 (http://hmmer.org/). A Pfam domain was considered to be conserved if it was matched by at least 3 ORFs in a given orthogroup. Assembled CDS were annotated as secreted if a signal peptide was detected in the corresponding protein sequence by Phobius version 1.01 [23]. The curated Fungi1 alignment was downloaded from the KNOTTIN database [34] and trimmed to the conserved cysteine residues plus 10 residues of padding on either side (alignment positions 423 through 483, counting from 1). Hmmer 3.0 (http://hmmer.org) was used to build an HMM from the trimmed alignment and to search the resulting model against the Hc assembled transcriptomes, predicted gene sets for 41 fungal genomes (see S18 Data), and the curated Fungi1 sequences using an expect threshold of 1 e-2, yielding 167 total hits, 121 of which corresponded to unique (non-redundant) genes. Non-redundant hits were aligned to the HMM with hmmalign, and a phylogenetic tree was inferred from the aligned positions using fasttree2 [65]. The phylogenetic tree is available as supplementary file 16 and the protein alignment is available as S17 Data. To generate the tree in Fig 3B, the full tree was pruned to the Hc G217B assembled transcripts plus additional sequences representative of the diversity of the G217B-containing clades, maintaining the topology and branch lengths of the full tree. To generate the reduced alignment shown in Fig 3C, the full alignment was reduced to just the Hc G217B assembled transcripts and AVR9, and adjusted with the following manual improvements: removing gap-only columns, removing staggered gaps at the N-terminus to improve the alignment of the aliphatic/aromatic position at +3 relative to the first cysteine, removing gaps at the C-terminus of ucsf_hc.01_1.G217B.08018 to align the final cysteine, and trimming all sequences to the aligned AVR9 positions. Biological duplicates of G217B yeast cells were grown at 37°C in HMM liquid medium to mid-log phase. Biological duplicates of G217B hyphal cells were grown for 4–6 weeks with passaging three times (1:5 dilution) into fresh HMM medium at RT before reaching a sufficient density of cells for harvesting. Hyphal and yeast cells were treated with 100 μg/mL cycloheximide for 2 minutes (MP Biomedicals, Santa Ana, CA) before harvesting of cells by filtration. Total RNA for mRNA library preparation was isolated from a small fraction of the total yeast or hyphal cells collected for ribosome profiling using a guanidine thiocyanate lysis protocol as previously described [13]. Ribosome profiling and matched mRNA-seq sample preparation and library building were performed as described [49], except that the 3′ linker ligation strategy was used instead of poly(A) tailing for marking and capturing the 3’ RNA end [50]. Libraries were multiplexed and subjected to 50 bp single-end sequencing with an Illumina HiSeq2000 sequencer (Illumina). The strand-specific, single-end reads from the matched mRNA and ribosome footprint samples were processed identically, except as noted. Due to the relatively short lengths of these reads, alignment steps were performed with bowtie version 0.12.7 [66], rather than the bowtie2/GSNAP approach used for the paired-end data. Reads were pre-processed by stripping 3' sequence matching the primer linker sequence, allowing 20% mismatches, filtering any reads that were less than 11 bp after linker stripping, and additionally filtering for matches to the mitochondrial, MAGGY, and rDNA sequences described above (see paired-end read assembly) using bowtie with default parameters. The remaining, linker-stripped sequences were mapped to the genome with bowtie, restricting the output to unique alignments (-k1) and post-filtered for full length alignments of the query sequences. We noted a strong bias for T mismatches in the first aligned position, consistent with previously observed terminal nucleotidyl transferase activity of reverse transcriptase [67]; to address this, we removed all first position T mismatches, treating the second aligned position as the true 5' of the sampled fragment for the remaining steps. Other than this special treatment of the first position, all alignments were required to be perfect matches between query and genome sequence. Aligned sequences were required to be at least 22 bp and, for ribosome footprint samples only, no more than 32 bp. For alignments passing the above criteria, the 13th aligned position, inferred to correspond to the ribosomal P-site for ribosome footprint samples, was taken as the location of the mapped read. Assembled transcript FPKMs for each sample were calculated as the number of reads located in the CDS of that transcript, divided by the length of that CDS in kilobases and the total number of CDS-mapped reads for the sample in millions. The same formula was used for quantifying ribosome footprints in leader regions, defining the leader as all transcript sequence 5' of the first CDS position, and normalizing by the length of the leader and the total number of CDS-mapped reads (to allow direct comparison of CDS and leader FPKMs). Yeast and hyphae per-state assembled transcript calls were used to identify transcripts with differential leader regions. Leader regions were calculated as the distance from the start of the 5’ transcript end to the beginning of the predicted CDS. Leaders were defined as differential between Hc cell types if the change in size of the yeast and hyphal leader lengths was measured to be ≥ 100 bp in 3 out of 4 Hc strains in the per-state transcript assemblies. Transcripts meeting these criteria were further evaluated manually to determine whether the observed mapped read density supported a 5’ transcript extension in one cell type. For assessment of translational efficiency and ribosome occupancy of longer leader transcripts, the set of conserved longer leader transcripts were further manually evaluated against the G217B mRNA read coverage from the ribosome profiling experiments to ensure that the observed mapped read density supported a 5’ transcript extension in one cell type. Cluster 3.0 was used to perform hierarchical clustering of genes using an uncentered Pearson correlation [68,69]. Clustered data was visualized using Java Treeview 1.1.4r4 (available at http://jtreeview.sourceforge.net) [70]. 2781 bp of the MS95 promoter/leader region (MS95p), 1045 bp of the GAPDH promoter (GAPDHp), and 729 bp of the CATB terminator (CATBt) were amplified from G217B gDNA and assembled into a Gateway entry vector pDONR/Zeo (Life Technologies) containing enhanced GFP (eGFP) using a CPEC cloning strategy [71]. This generated BAS1464 (MS95p –eGFP–CATBt) and BAS1514 (GAPDHp–eGFP–CATBt) constructs. All primer sequences are included in S4 Table. Using LR Gateway cloning (Life Technologies) each pDONR/Zeo entry vector was recombined into the Hc episomal expression vector pDG33 (pDG33 is a derivative of pWU55 [72] with Hc URA5 added for selection and made Gateway compatible). The episomally-maintained positive control (GAPDHp–eGFP–CATBt) and MS95p –eGFP–CATBt constructs were electroporated into G217Bura5Δ as previously described [72]. G217Bura5Δ strains transformed with the GAPDHp and MS95p eGFP constructs were grown at 37°C to late log phase, diluted 1:25 into 5 mL HMM medium for growth at 37°C or 1:10 into 10 mL HMM medium for growth at RT. At 1 d, 2 d, and 3 d post-inoculation, cells were harvested by centrifugation and protein and RNA was isolated simultaneously from each 37°C or RT culture using Qiazol (Qiagen, Netherlands) following the manufacturer’s instructions. The Hc WET1 coding sequence, 1035 bp of the ACT1 promoter (ACT1p), and 729 bp of the CATB terminator (CATBt) were amplified from G217B gDNA and assembled into the Gateway entry vector pDONR/Zeo (Life Technologies) using restriction enzymes to generate BAS1504. A vector control construct, BAS252, was generated identically except lacking the WET1 CDS. Using LR Gateway cloning (Life Technologies) each pDONR/Zeo entry vector was recombined into the Hc episomal expression vector pDG33 (pDG33 is a derivative of pWU55 [72] with Hc URA5 added for selection and made Gateway compatible). The episomally-maintained vector control (ACT1p –CATBt) and ACT1p –WET1 –CATBt constructs were electroporated into G217Bura5Δ as previously described [72]. Protein and RNA was isolated simultaneously from each 37°C or RT culture using Qiazol (Qiagen, Netherlands) following the manufacturer’s instructions. Cell morphology of vector control and WET1-expressing cells was determined using differential interference contrast (DIC) microscopy with a Yokogawa CSU-X1 (Yokogawa, Tokyo, Japan) spinning disk confocal mounted on a Nikon Eclipse Ti inverted microscope (Nikon, Tokyo, Japan) with a PLAN APO 40X objective (Nikon) and an Andor Clara digital camera (Andor, Belfast UK). Images were acquired by and processed in NIS-Elements software 4.10 (Nikon). 10–20 μg of protein was resolved on a 4–12% NuPAGE Bis-Tris gel (Life Technologies) in MOPS buffer. For detecting endogenous levels of Wet1, a rabbit polyclonal peptide antibody was raised (Covance, Princeton, NJ) against Wet1 (epitope: KTKARREQEAREKRRKLS; ID: CA2890). GFP was detected using a mouse anti-GFP antibody (Roche Applied Bioscience, Indianapolis, IN; 11814460001). Equivalency of protein levels between samples was assessed with a rabbit antibody against Cdc2 (Santa Cruz Biotechnology, Santa Cruz, CA; sc-53). 5–10 μg of total RNA was separated on a 1.5% denaturing agarose-formaldehyde gel and transferred to a positively charged nylon membrane (Roche Applied Bioscience or PerkinElmer, Waltham, MA). Northern probes were generated with gene specific primers by amplifying ~ 200–500 bp of transcript from genomic DNA. Primer sequences are given in S4 Table. 100–150 ng of each DNA probe was labeled using the High Prime DNA Labeling Kit (Roche Applied Bioscience) and [α – 32P] dCTP (PerkinElmer). Membranes were blocked in UltraHybe hybridization buffer (Ambion, Life Technologies) for at least 30 minutes and denatured probe was added to the same blocking buffer and incubated at 42°C overnight. The next day membranes were washed twice in 2X SSC, 0.1% SDS for 5 minutes and twice in 0.1X SSC, 0.1% SDS for 15 minutes at 42°C before exposure to a phosphorimager screen (GE Life Sciences, Pittsburgh, PA). Screens were scanned with a Fuji FLA-5100 imager and analyzed with Multi Gauge Software (ver. 3.1; Fujifilm/GE Life Sciences). Patterns of ribosome occupancy on G217B yeast and hyphal longer leader transcripts were used to determine categories of translational regulation. Categories were defined as follows: a: Ribosome density on longer leader region and decreased differential TE for CDS of longer leader transcript, b: Ribosome density on longer leader region and no change in TE for CDS of longer leader transcript, c: Ribosome density on longer leader region and increased TE for CDS of longer leader transcript, d: No ribosome density on longer leader region and decreased TE for CDS of longer leader transcript, e: No ribosome density on longer leader region and no change in TE of CDS for longer leader transcript, f: No ribosome density on longer leader region and increased TE for CDS of longer leader transcript. g: Ribosome profiling reads used to calculate TE CDS values are below the limit of accurate measurement (< 128 counts) in at least one cell type. We examined these categories using two different fold cut-offs for CDS TE values: a change in TE was defined as ≥ 1.5 fold (shown in S15 Fig) or ≥ 2 fold (shown in Fig 8). FPKM values and raw counts of ribosome density on leader regions were used to assist in determining the presence of ribosome density in differential leader regions after manual examination of the location of the footprint density. After manual confirmation of the location of footprint reads on each leader region, leaders with greater than 10 footprint reads in the differential leader region were considered to have ribosome density. MochiView [73] and IGV [74] were used to visualize sequencing reads. All statistical calculations were carried out in R 2.15.1 [75]. We wrote custom scripts and generated plots in Python 2.7, using the following open-source libraries: NumPy 1.6.2 and SciPy 0.10.1 [76], NetworkX 1.7rc1 [77], and Matplotlib 1.1.1rc2 [78]. IPython notebooks were used for interactive data exploration and collaboration [79]. Transcriptome assembly data for each Hc strain (G217B, G186AR, H88, and H143) are available as genomic features gff3 files (generic feature format version 3) (S1–S4 Data; S19–S26 Data). For high-throughput sequencing data, the raw data are available at the NCBI Sequence Read Archive (SRA) and Gene Expression Omnibus (GEO) databases [80,81] under GEO SuperSeries accession GSE68707.

10.1371/journal.pntd.0006026

RNA-seq transcriptional profiling of Leishmania amazonensis reveals an arginase-dependent gene expression regulation

Leishmania is a protozoan parasite that alternates its life cycle between the sand-fly vector and the mammalian host. This alternation involves environmental changes and leads the parasite to dynamic modifications in morphology, metabolism, cellular signaling and regulation of gene expression to allow for a rapid adaptation to new conditions. The L-arginine pathway in L. amazonensis is important during the parasite life cycle and interferes in the establishment and maintenance of the infection in mammalian macrophages. Host arginase is an immune-regulatory enzyme that can reduce the production of nitric oxide by activated macrophages, directing the availability of L-arginine to the polyamine pathway, resulting in parasite replication. In this work, we performed transcriptional profiling to identify differentially expressed genes in L. amazonensis wild-type (La-WT) versus L. amazonensis arginase knockout (La-arg-) promastigotes and axenic amastigotes. A total of 8253 transcripts were identified in La-WT and La-arg- promastigotes and axenic amastigotes, about 60% of them codifying hypothetical proteins and 443 novel transcripts, which did not match any previously annotated genes. Our RNA-seq data revealed that 85% of genes were constitutively expressed. The comparison of transcriptome and metabolome data showed lower levels of arginase and higher levels of glutamate-5-kinase in La-WT axenic amastigotes compared to promastigotes. The absence of arginase activity in promastigotes increased the levels of pyrroline 5-carboxylate reductase, but decreased the levels of arginosuccinate synthase, pyrroline 5-carboxylate dehydrogenase, acetylornithine deacetylase and spermidine synthase transcripts levels. These observations can explain previous metabolomic data pointing to the increase of L-arginine, citrulline and L-glutamate and reduction of aspartate, proline, ornithine and putrescine. Altogether, these results indicate that arginase activity is important in Leishmania gene expression modulation during differentiation and adaptation to environmental changes. Here, we confirmed this hypothesis with the identification of differential gene expression of the enzymes involved in biosynthesis of amino acids, arginine and proline metabolism and arginine biosynthesis. All data provided information about the transcriptomic profiling and the expression levels of La-WT and La-arg- promastigotes and axenic amastigotes. These findings revealed the importance of arginase in parasite survival and differentiation, and indicated the existence of a coordinated response in the absence of arginase activity related to arginine and polyamine pathways.

Leishmania are auxotrophic for many essential nutrients, including amino acids. In this way, the parasite needs to uptake the amino acids from the environment. The uptake of amino acids is mediated by amino acid transporters that are unique for Leishmania. As part of polyamine pathway, the arginase converts L-arginine to ornithine and furthermore to putrescine, products which are essential for parasite growth. On the other hand, the absence of arginase activity could alter the metabolism of the parasite to surpass the external signals during the life cycle and the fate of infection. The transcriptional profiling of La-WT and La-arg- promastigotes and axenic amastigotes revealed 8253 transcripts, 60% encoding hypothetical proteins and 443 novel transcripts. In addition, our data revealed that 85% of the genes were constitutively expressed. Among the 15% (1268 genes) of the differentially expressed genes, we identified genes up- and down-regulated comparing the transcript abundance from different life cycle stages of the parasite and in the presence or absence of arginase. We also combined the transcriptional with metabolic profile that revealed a proportional correlation between enzyme and metabolites in the polyamine pathway. The differentiation of promastigotes to amastigotes alters the expression of enzymes from polyamines biosynthesis, which modulates ornithine, L-glutamate, proline and putrescine levels. In addition, the absence of arginase activity increased the levels of L-arginine, citrulline and L-glutamate and decreased the levels of aspartate, proline, ornithine and putrescine in promastigotes by differential modulation of genes involved in its metabolism. Altogether these data provided additional insights into how Leishmania is able to modulate its biological functions in the presence or absence of arginase activity to survive during environmental changes.

Leishmania is a protozoan parasite that causes widespread human disease known as leishmaniases, characterized by cutaneous, mucosal or visceral manifestations. Leishmania alternates its life cycle between the sand-fly vector (promastigote form) and the mammalian host (amastigotes form) [1]. This alternation involves environmental changes and submits the parasite to dynamic modifications in morphology, metabolism, cellular signaling and regulation of gene expression to allow for a rapid adaptation to new conditions. The parasite has also developed resistance mechanisms to evade sand-fly digestive enzymes and the host innate immune response, such as the mammalian complement system and macrophage defense mechanisms involving nitric oxide (NO). NO is produced by nitric oxide synthase 2 (NOS2) using the amino acid L-arginine as substrate [2, 3]. On the other hand, arginase is an immune-regulatory enzyme that can reduce NO production by activated macrophages, limiting the availability of L-arginine to NOS2, supporting Leishmania resistance to host defense mechanisms. Arginase uses L-arginine to produce urea and ornithine, a precursor of the polyamine pathway [4]. The success of the Leishmania infection depends on the parasite ability to subvert the host defense mechanisms [3, 5]. Leishmania also expresses arginase, which supplies the metabolic precursors for parasite replication, an essential step for the establishment of the infection [4, 6]. Our research group has been studying the role of arginase in L. amazonensis during the parasite life cycle and its role in the establishment and maintenance of the infection in mammalian macrophages [7–9]. Transcriptional profiling has been used in expression studies of several model organisms, including Leishmania [10–13]. Holzer et al. (2006) used microarray analysis to determine that 3.5% of the genes were differentially expressed between promastigotes and lesion-derived amastigotes of L. mexicana, and 0.2% were differentially expressed between promastigotes and axenic amastigotes. The reduced number of regulated genes was a consequence of an increase in the magnitude of the transcript levels in cells under axenic conditions [14]. Leifso et al. (2007) also demonstrated differential gene expression between promastigote and lesion-derived amastigote forms of L. major [15]. These data indicated that the Leishmania genome is mostly constitutively expressed during the parasite life cycle, but there are still some genes that are differentially expressed to adapt to different environmental changes [14–16]. In addition, Goldmann et al. (2007) demonstrated with transcriptome analysis that arginase has an important role in the establishment of infection with Streptococcus pyogenes. Arginase type II was up-regulated in the infection of macrophages with S. pyogenes after 1, 4 and 16 h. However, NOS2 did not show differential gene expression. The same profile was observed in macrophages stimulated with γ-interferon and lipopolysaccharide [17]. In this work, through RNA-seq of La-WT and La-arg- promastigotes and axenic amastigotes, we identified 8253 transcripts, from which 60% encoding hypothetical proteins and 443 novel transcripts that did not match any previously annotated gene. The transcriptional profiling revealed that 85% of the genes were constitutively expressed. Among the 15% (1268 genes) that were DE, we identified genes up- and down-regulated. Interestingly, we showed 100 genes differentially expressed in La-WT promastigotes and 908 genes differentially expressed in La-arg- promastigotes. Additionally, we identified 183 genes differentially expressed in La-WT axenic amastigotes and only 34 genes differentially expressed in La-arg- axenic amastigotes. In summary, our results showed that L. amazonensis could modulate gene expression with differential regulation between promastigote and axenic amastigotes, indicating that this organism may represent an alternative paradigm for eukaryotic differentiation with minimal contributions from changes in mRNA abundance. The transcriptional profiling also revealed differential gene expression in the development of the Leishmania life cycle and the existence of a coordinated response in the absence of arginase activity, providing additional insights into how Leishmania is able to modulate its biological functions to survive during environmental changes. Leishmania (Leishmania) amazonensis (MHOM/BR/1973/M2269), a strain of our laboratory collection at the Institute of Bioscience, and L. amazonensis arginase knockout (La-arg-) [8] promastigotes were grown at 25°C in M199 medium, pH 7.0, supplemented with L-glutamine, 10% heat-inactivated fetal bovine serum, 0.25% hemin, 40 mM NaHCO3, 100 μM adenine, 40 mM HEPES, 100 U/mL penicillin and 100 μg/mL streptomycin. Axenic amastigotes of La-WT and La-arg- were grown in M199 medium supplemented, as described above at 34°C, pH 5.5. For the La-arg- cultures, hygromycin (30 μg/mL), puromycin (30 μg/mL) and putrescine (50 μM) were added. Bone marrow derived-macrophages (BMDM) were collected from the femur of female BALB/c mice (6–8 weeks) from the Animal Center of the Institute of Bioscience of the University of Sao Paulo. The femurs were washed with cold PBS and the cells were collected at 500 x g for 10 min at 4°C. The lysis of erythrocytes was performed with NH4Cl (145 mM) and Tris-base (200 mM) pH 7.0 and incubated on ice for 20 min. After lysis, the cells were washed with cold PBS, collected at 500 x g for 10 min at 4°C and incubated in RPMI 1640 medium supplemented with penicillin (100 U/mL), streptomycin (100 μg/ml), 2-mercaptoethanol (50 μM), L-glutamine (2 mM), sodium pyruvate (1 mM), fetal bovine serum 10% and L929 conditioned medium (15%), as macrophage stimulating factor source. The cells were cultivated for 7 days at 34°C and 5% CO2. After differentiation, cellular viability was evaluated with Trypan blue staining 1:1, and cells were counted in a Neubauer chamber. Approximately 1x106 BMDM were incubated on sterile 13-mm coverslips in 24-well plates overnight at 34°C and 5% CO2 to adhere to the coverslips. Non-adherent cells were removed by PBS washing, and the infection was performed with La-WT or La-arg- axenic amastigotes (MOI 5:1). After 4 h of infection, the cultures were washed with PBS and maintained in culture for 24, 48 and 72 h. Non-infected macrophages maintained in culture at the same conditions were used as control. The infections were evaluated by determining the percentage of infection after counting 200 Panoptic-stained (Laborclin, Parana, Brazil) macrophages. The infection index was determined by multiplying the percentage of infected macrophages by the mean number of parasites per infected cell [18, 19]. Statistical analyses were performed using the t-test. Total RNA from 3 independent biological replicates was isolated from La-WT and La-arg- promastigotes and axenic amastigotes using TRIzol reagent (Life Technologies, Carlsbad, CA, USA), according to the manufacturer’s instructions. RNA samples were treated with DNase I (Thermo Scientific, Lithuania, EU), and the RNA concentration was determined using a spectrophotometer at A260/A280 (Nanodrop ND1000, Thermo Scientific, USA). In addition, the RNA integrity was evaluated using an Agilent 2100 Bioanalyzer and Pico Agilent RNA 6000 kit (Agilent Technologies, Santa Clara, CA, USA), according to the manufacturer’s instructions. rRNA depletion was performed by poly(A) magnetic beads capture protocol, using Strand-specific TrueSeq RNA-seq Library Prep (Illumina), according to manufacturer´s instruction. Library preparations were performed using Strand-specific TrueSeq RNA-seq Library Prep (Illumina), according to the manufacturer’s instructions. Paired-end reads (125 bp) were obtained using the Illumina HiSeq 2000 platform at the Norwegian Sequencing Centre at the University of Oslo. Trimmomatic was used to remove the Illumina adapter sequences [20]. The quality of the produced data was analyzed using FastQC by Phred quality score [21]. Reads with Phred quality scores lower than 20 were discarded. Reads were aligned to the L. mexicana (MHOMGT2001U1103) genomic data obtained from TriTrypDB version 29 (www.tritrypdb.org) using TopHat (-G option) [22, 23]. Maximum 2 mismatches were allowed. Thereafter, the expression level of the assembled transcriptome and abundance estimation were performed using Cufflinks [24]. The abundance of transcripts was calculated as the Fragments Per Kilobase of transcript per Million mapped reads (FPKM), which reflects the abundance of a transcript in the sample by normalization of the RNA length and the total read number [25]. The gene expression level values were calculated from the transcript counts. Differentially expressed gene analysis was performed on four comparisons pairs (La-WT promastigotes vs. La-arg- promastigotes; La-WT axenic amastigotes vs. La-arg- axenic amastigotes, La-WT promastigotes vs. La-WT axenic amastigotes; La-arg- promastigotes vs. La-arg- axenic amastigotes). Genes with zero FPKM were excluded (excepted the arginase gene (LmxM.34.1480)). Transcripts that did not match any previous annotated gene were considered novel, and they were identified using Cufflinks with -g option. Statistical significance of DE genes data was determined using independent t-test and fold change in which the null hypothesis was that no difference exists among groups. False discovery rate (FDR) was controlled by adjusting p value using Benjamini-Hochberg algorithm [26]. Functional annotation was performed using GO (Gene Ontology) and the Kyoto Encyclopedia of Genes and Genomes (KEGG). All analyses were performed by Macrogen (www.macrogen.com). Reverse transcription was performed using 2 μg of total RNA as a template, reverse transcriptase and random primers (Revertaid H minus Reverse Transcriptase kit, Thermo-Scientific, Canada), according to the manufacturer’s instructions. Equal amounts of cDNA were assessed in triplicate in a total volume of 25 μL containing Power SYBR Green qPCR Master Mix (Life Technologies, Warrington, UK) and the following primers (20 μM): GAPDH_F 5´-TCAAGGTCGGTATCAACGGC-3´, GAPDH_R 5´-TGCACCGTGTCGTACTTCAT-3´, arginase F 5´-TCCTGCACGACCTGAACATC-3´, arginase R 5´-CGCCATGGACACCACCTT-3´, glutamate 5-kinase F 5´-AGCTGGTTTTTGGCGACAAC-3´, glutamate 5-kinase R 5´-CGTCGATGTCGCTGAGAATG-3´, pyrroline 5-carboxylase dehydrogenase F 5´-ACGGTGTTTGTGTATGACGACAGT-3´, pyrroline 5-carboxylase dehydrogenase R 5´-ACCGGTCAGGCCGTACTTC-3´, spermidine synthase F 5´-GCAACCAGGGCGAGTCTATCT-3´, spermidine synthase R 5´-TGACCGTGGAAAAGCCAATAT-3´, amastin-like F 5´-GGAGCGCTACTTCAGCTATGGA-3´, amastin-like R 5´-CGGATCATCAATAAGACGATGTTG-3´, amastin-like F 5´-CGGCTGCCTTTTGCTGTACT-3´ and amastin-like R 5´-CAGACAACGCAAGCTGTGACA-3´. The mixture was incubated at 94°C for 5 min, followed by 40 cycles at 94°C for 30 s, 60°C for 30 s and 72°C for 30 s. A negative control in the absence of reverse transcriptase was included in RT-qPCR assays to detect DNA contamination in RNA samples. The copy number of the target and reference genes were quantified in three biological samples, considering the molar mass concentration, according to a standard curve generated from a ten-fold serial dilution of a quantified PCR product. The normalized target/gapdh ratio of the molecules absolute number of each target was used as a parameter of the expression. Reactions were carried out using PikoReal 96 RealTime PCR System (Thermo Scientfic, Finland). Analyses were performed using PikoReal Software 2.2 (Thermo Scientific, Finland). The experimental protocols for the animals were approved by the Animal Care and Use Committee from the Institute of Bioscience of the University of Sao Paulo (CEUA 196/2014). This study was carried out in strict accordance with the recommendations in the guide and policies for the care and use of laboratory animals of the São Paulo State (Lei Estadual 11.977, de 25/08/2005) and Brazil government (Lei Federal 11.794, de 08/10/2008). La-WT and La-arg- promastigotes in the stationary growth phase were submitted to differentiation in medium, pH 5.5 at 34°C, for 48 h. The growth curve of the differentiated amastigotes of both La-WT and La-arg- reached stationary growth phase after 10 days of incubation (S1A Fig). Axenic amastigotes were also induced to differentiate back to promastigotes incubating the parasites at pH 7.0 and 26°C that led to differentiation after 48 h of incubation. Moreover, axenic amastigotes infectivity was evaluated. BMDM from BALB/c mice were infected with La-WT and La-arg- (MOI 5:1), and the infection index was analyzed at 24, 48 and 72 h post infection. According to S1B Fig, both La-WT and La-arg- axenic amastigotes were able to infect and establish the infection. However, the infection index for La-arg- was lower than for La-WT, corroborating the previously determined infection index for La-arg- promastigotes [8]. Once axenic amastigotes infectivity was confirmed, the total RNA from La-WT and La-arg- promastigotes and axenic amastigotes was extracted and submitted to RNA-seq, as described in Methods. Transcriptomic analyses were performed using 3 independent biological replicates from each La-WT and La-arg- promastigotes and axenic amastigotes after Illumina HiSeq2000 sequencing that generated million sequence reads (125 bp) (S1 Table). Sequencing data are available on the NCBI BioProject under accession number PRJNA380128 and Sequence Read Archive (SRA) under accession number SRX2661998 and SRX2661999. RNA-seq data were aligned to the L. mexicana genome [27]. After initial assembling, 8253 transcripts and 443 novel transcripts were identified with genome coverage around 90%. And 60% of the transcripts corresponding to hypothetical proteins as listed in S2 Table. The novel transcripts were those that did not correspond to any previous annotated gene, even in L. mexicana genome, used in the comparisons since L. amazonensis is not completely annotated. According to Fig 1 we demonstrated the arginase transcript (LmxM.34.1480) assembling with genome coverage in La-WT promastigotes and axenic amastigotes. As expected no transcripts were detected in the knockout lines (La-arg- promastigotes and axenic amastigotes). RNA-seq has been described as an accurate method for quantifying transcript levels [10–12, 28, 29]. Our RNA-seq data revealed that 85% of the genes were constitutively expressed, comparing the gene expression profiles of La-WT and La-arg- promastigotes and axenic amastigotes. However, among 15% (1268 genes) DE genes, we identified a vast number of genes differentially expressed. Of the total 378 and 357 DE genes in La-WT promastigotes and axenic amastigotes, 100 and 183 genes were non-common for each line, respectively. Of the total 908 and 62 DE genes in La-arg- promastigotes and axenic amastigotes, 554 genes and 34 genes were non-common for each line, respectively (Fig 2). A direct overlap revealed only 2 transcripts mutually expressed among all samples (Fig 2). The analyses of the DE genes, limited to those presented fold change ≥ 2 and p ˂ 0.05, revealed 195 genes up-regulated and 183 genes down-regulated in the comparison of La-WT promastigote vs. La-arg- promastigote, suggesting a significant amount of DE genes which regulation depends on arginase activity. On the other hand, in the comparison of La-arg- axenic amastigotes vs. La-WT axenic amastigotes, only 37 genes were up-regulated and 25 genes were down-regulated. In addition, in the comparison of La-WT axenic amastigotes vs. La-WT promastigotes we observed 208 up-regulated and 149 down-regulated genes, indicating a significant amount of DE genes during La-WT differentiation. The comparison of La-arg- axenic amastigotes vs. La-arg- promastigotes led to larger number of DE genes (452 up-regulated and 456 down-regulated genes) what could be explained considering the two variables in this comparison, the absence of arginase activity and the life cycle stage (Fig 3). The additional characterization of axenic amastigotes transcripts revealed up-regulation of the following amastins: amastin (LmxM.30.0451 and LmxM.30.0452) and amastins-like (LmxM.08.0750, LmxM.08.0760, LmxM.08.0770, LmxM.08.0800, LmxM.08.0850, LmxM.33.0960, LmxM.33.0961, LmxM.33.1560 and LmxM.33.1920) (S2 Fig). Based on the DE genes analyzed, we generated volcano plots showing the distribution of transcripts by comparing the fold change in the expression (log2) of each group with the corresponding adjusted p value (-log10) (S3 Fig). We further analyzed the volume plot (S4 Fig) identifying the top 5 transcripts that showed higher expression difference compared to the control according to expression volume (Table 1). Comparing the expression profile of La-WT vs. La-arg- promastigotes, we identified the following up-regulated transcripts: a conserved hypothetical protein (LmxM.15.1520) and a non-coding RNA (LmxM.23.ncRNA rfamscan: 218578-218718-1); and the following down-regulated transcripts: a putative nucleolar RNA-binding protein (LmxM.07.0990) and two histone H4 proteins (LmxM.06.0010 and LmxM.15.0010). Comparing the expression of La-WT vs. La-arg- axenic amastigotes, we identified the following up-regulated transcripts: a putative dipeptidyl-peptidase III (LmxM.05.0960), the protein disulfide isomerase (LmxM.06.1050) and a conserved hypothetical protein (LmxM.08.0540); and the following down-regulated transcripts: a conserved hypothetical protein (LmxM.17.0890) and a putative tryparedoxin 1 protein (LmxM.08_29.1160). The comparison of the expression profile of La-WT promastigotes vs. axenic amastigotes led to the identification of the following up-regulated transcripts: the tryparedoxin peroxidase (LmxM.15.1160) and a putative ATP-dependent RNA helicase (LmxM.33.2050); and the following down-regulated transcripts: a putative histone H3 (LmxM.10.0970), the histone H4 (LmxM.15.0010) and an unspecific product (LmxM.13.0290partial). Finally, the comparison of the expression profile of La-arg- promastigotes vs. axenic amastigotes led to the identification of only down-regulated transcripts: an alpha tubulin (LmxM.13.0280), two beta tubulins (LmxM.32.0792, LmxM.32.0794), an unspecific product (LmxM.13.0300) and a non-coding RNA (LmxM.23.ncRNA rfamscan: 218682-218827-1). Furthermore, we performed a KEGG enrichment analysis, which showed a list of the top 20 regulated pathways among all samples. The list includes pathways that can be regulated in the absence of arginase activity, such as the biosynthesis of amino acids, arginine and proline metabolism and arginine biosynthesis (Fig 4 and Table 2). In this work, we focused on the L-arginine pathway and crossed the obtained data with previous metabolome fingerprints, determined by capillary electrophoresis, also focusing on L-arginine metabolism and the modulation of polyamine metabolism comparing La-WT and La-arg- promastigotes. Castilho-Martins et al. (2015) observed that the absence of arginase activity led to an increase of L-arginine and citrulline levels, but a decrease of ornithine, proline and putrescine levels. These results confirmed the importance of L-arginine supplying the polyamine pathway in L. amazonensis and also showed a possible alternative pathway to provide substrates for the pathway in the absence of arginase in the parasite [9]. In fact, to understand why the absence of arginase induced an increase in L-arginine and citrulline levels and a decrease in ornithine, proline and putrescine levels, we analyzed the transcripts levels of specific enzymes involved in these pathways, such as arginase (LmxM.34.1480/EC3.5.3.1), pyrroline 5-carboxylate reductase (LmxM.13.1680/EC1.5.1.2), pyrroline 5-carboxylate dehydrogenase (LmxM.03.0200/EC1.2.1.88), glutamate 5-kinase (LmxM.26.2710/EC2.7.2.11), spermidine synthase (LmxM.04.0580/EC 2.5.1.16), acetylornithine deacetylase (LmxM.07.0270/EC3.5.1.16) and arginosuccinate synthase (LmxM.23.0260/EC6.3.4.5) (Table 3). Arginase transcripts were not detected in the two knockout lines (La-arg- promastigotes and La-arg- axenic amastigotes), as expected. Interestingly, a reduced level of arginase was observed in La-WT axenic amastigotes, compared to promastigotes. The increase of pyrroline 5-carboxylate reductase transcripts in La-arg- compared to La-WT promastigotes showed 1.42-fold change. Additionally, an increase of glutamate 5-kinase transcripts was observed in both La-WT and La-arg- axenic amastigotes, compared to promastigotes. On the other hand, we observed decreased transcripts levels of pyrroline 5-carboxylate dehydrogenase, spermidine synthase, acetylornithine deacetylase and arginosuccinate synthase, with 0.76, 0.31, 0.79 and 0.72-fold change, respectively (Table 3). The crossing of these findings with metabolome data could explain that the increase of L-arginine and citrulline levels could be a consequence of the absence of arginase activity. The increase of L-glutamate levels could be related to the increase of pyrroline 5-carboxylase reductase transcripts. Further, as a consequence of the high consumption of L-glutamate, we observed the decrease of proline that could be related to the decrease of pyrroline 5-carboxylase dehydrogenase transcripts. The decrease of putrescine levels could be related to the decrease of spermidine synthase transcripts. The decrease of ornithine levels could be related to the decrease of acetylornithine deacetylase transcripts. And finally, the decrease of aspartate could be related to the decrease of arginosuccinate synthase transcripts (Fig 5). Additionally, we performed RT-qPCR validation of some enzymes as shown in S5 Fig. Here, we described the DE gene profile comparing the expression in promastigotes and axenic amastigotes in the presence or absence of arginase activity. In addition, we performed a correlation analysis with the KEEG arginine pathway, highlighting the regulated enzymes, as previously described in the metabolome work. L-arginine is an amino acid used as precursor not only for protein synthesis, but also for the synthesis of NO, urea, ornithine, citrulline, creatinine, agmatine, L-glutamate, proline and polyamines [30]. On the other hand, arginase is an enzyme with regulatory roles, modulating L-arginine availability and production of ornithine, a precursor of polyamines, essential for cell replication [30]. Therefore, arginine biosynthesis is an important pathway that not only participates in the regulation of NOS2 parasite killing and arginase-mediated parasite growth, but is also involved in the regulation of the immune system [31, 32]. The infection of murine macrophages with L. amazonensis showed increased levels of arginase I, La-arginase, arginine transporters (CAT2B and LaAAP3) and miRNA modulation [32]. However, infection with La-arg- induces NOS2 expression and the production of NO, causing a lower infection index [8] and blocking miRNA expression [32]. These changes in gene regulation can indicate mechanisms to subvert the defense mechanism developed by the parasite [32, 33]. RNA-seq technology has been used to describe transcriptomic profiles of L. major, L. mexicana and L. braziliensis [10–13, 34]. All of these studies have provided additional knowledge about Leishmania biology and the coordinated response of Leishmania-infected macrophages in relation to gene regulation at the transcriptional level [12, 13, 29]. RNA-seq data are also helping to revise the previous genome annotation of L. mexicana [34] and to reconstruct some genomic regions of the L. major genome that were misassembled [35] in an attempt to improve the current genome and gene annotations. In this work, using the RNA-seq approach, we described the transcriptional profiling of L. amazonensis, compared to the phylogenetically close L. mexicana genome, since L. amazonensis genome is not completely annotated [27, 36, 37]. We obtained RNA-seq data from La-WT and La-arg- promastigotes and axenic amastigotes allowing the comparison of transcript abundance from different life cycle stages in the presence or absence of arginase, an important enzyme of the parasite´s polyamines synthesis. From the 8253 transcripts identified in La-WT and La-arg- promastigotes and axenic amastigotes, 60% of them were identified as hypothetical proteins and 443 were identified as novel transcripts, that did not correspond to any previously annotated genes. Although we obtained less transcripts than previously predicted for L. mexicana [34], we could assure that the transcripts identified fulfill confidence coverage of the RNA-seq data described in this work. Recent studies have been showing the importance of characterizing a hypothetical protein not only by functional genomics, but also according to its general biological features, allowing the acquisition of new knowledge about signaling pathways, metabolism, stress response, drug resistance and in the identification of new therapeutic targets [38]. The identification of novel transcripts has been improving the accuracy of the L. amazonensis genome [34, 35]. The finding of novel transcripts can suggest that the gene content of this organism may be higher than previously determined because the majority of novel transcripts contain open reading frames shorter than coding sequencing of genes in the current genome annotation. Another explanation could point to a novel genome organization and processing [29, 34, 39]. L. major chromosomes are organized as large clusters of genes or open reading frames in the same 5´- 3´ direction on the same DNA strand [40]. Each cluster of genes is processed by a transcription initiation site in a polycistronic transcription [41]. Individual mRNA transcripts are then processed, co-transcriptionally, by 5´RNA splicing and 3´polyadenylation. Therefore, without a gene specific promoter, the entire chromosome is constitutively transcribed [41]. The analysis of the transcriptome revealed 85% of genes were constitutively expressed in promastigotes and axenic amastigotes of L. amazonensis. The Leishmania genome has been described as constitutively expressed, indicating that the parasite is adapted for survival and replication in the sand-fly vector or macrophage host, using an appropriate set of genes/proteins for different environments [40]. Altogether, these results support the previous hypothesis that the Leishmania genome is mostly constitutively expressed. Interestingly, among the 1268 DE genes identified in this work, most was detected in La-arg- promastigotes (908 genes). La-arg- do not use L-arginine to produce ornithine because arginase activity is absent in this parasite line that requires polyamine supplementation for survival and replication [8]. La-arg- also presented a higher concentration of L-arginine in the cytoplasm in relation to the La-WT promastigote, probably because the parasite can sense the L-arginine pool, that leads to the regulation of L-arginine transporter expression and L-arginine uptake [42]. The absence of arginase can induce the parasite to regulate many genes involved in the arginine pathway. Therefore, the null mutant La-arg- was previously characterized and the essentiality of arginase in Leishmania in vitro growth was demonstrated with the requirement of putrescine supplementation [8]. In contrast, the arginase add-back mutant line (La-arg-/+ARG) restored arginase expression, growth and infectivity in vivo [8] and in vitro [32] assays. So, we focused on arginine pathway comparing our RNA-seq data with metabolome fingerprints, previously described by our group [9]. The first transcript analyzed was arginase (EC3.5.3.1) and, as expected, no transcript was observed in both life cycle stages of the arginase knockout line (La-arg- promastigotes and La-arg- axenic amastigotes), reinforcing the efficiency of the knockout methodology used [8] and indicating that the profile is maintained after amastigote differentiation. The metabolomic analysis of La-arg- promastigotes showed that absence of arginase causes an increase in L-arginine and citrulline levels, and the decrease in ornithine, putrescine and proline levels, indicating an alternative pathway to surpass the lack of this enzyme [9, 42]. Citrulline could be metabolized by arginine deiminase (EC3.5.3.6) or oxidoreductases (EC1.14.13.39/EC1.14.13.165). Arginine deiminase acts on carbon-nitrogen bonds [43]. Oxidoreductases act on paired donors, with incorporation or reduction of molecular oxygen on arginine biosynthesis [44]. It is interesting to note that similar transcript levels of this oxidoreductase were observed in La-WT and La-arg- promastigote and axenic amastigotes. Furthermore, the decrease in the levels of aspartate, ornithine, proline and putrescine indicates that this pathway can be used as an alternative pathway due to the differential expression of argininosuccinate synthase (EC6.3.4.5), acetylornithine deacetylase (EC3.5.1.16), pyrroline-5-carboxylate reductase (EC1.5.1.2), pyrroline-5-carboxylase dehydrogenase (EC1.2.1.88) and glutamate 5-kinase (EC2.7.2.11). The decrease in aspartate may be due to its conversion to L-arginine succinate by argininosuccinate synthase. The decrease in ornithine could be due to the absence of L-arginine conversion in ornithine by arginase and/or the acetylornithine deacetylase consumption. Glutamate 5-kinase, which is also involved in glutamate metabolism, was not differentially expressed between La-WT and La-arg- promastigotes. This could be explained by the maintenance from the substrates L-glutamate to glutamyl-P [9]. Interestingly, increased transcript levels of glutamate 5-kinase in both La-WT and La-arg- axenic amastigotes were observed. The glutamate metabolism was previously described to be involved in the differentiation of Trypanosoma cruzi from epimastigotes to metaclyclic trypomastigotes [45, 46]. The L-glutamate levels were increased in L-arginine deprivation, indicating a role of L-glutamate in L-arginine metabolism to supply the absence of L-arginine uptake [9]. In addition, it was described that, as alanine, L-glutamate has a key role in the cell physiology of Leishmania [47]. Other enzymes involved in the arginine pathway were not described in this work since they did not appear to be regulated in the previous metabolome analysis. In the list of top 5 transcripts differentially expressed based on the volume plot, none of the most regulated transcripts was related to the arginine pathway. However, it presented interesting DE genes. Three conserved hypothetical proteins (LmxM.15.1520, LmxM.08.0540 and LmxM.17.0890) were identified and their characterization is important not only for functional genomics but also to improve the knowledge about signaling pathways, metabolism, the stress response, drug resistance, as well as for the identification of new therapeutic targets [38]. In addition, the identification of histone H3 and H4 (LmxM.06.0010, LmxM.15.0010 and LmxM.10.0970) in the comparisons with La-WT (pro La-WT vs pro La-arg-, and pro La-WT vs ama La-WT) is indicative that histones play a central role in transcription regulation, DNA repair, DNA replication and chromosomal stability [40, 48, 49] in Leishmania life cycle, even in the absence of arginase activity. RNA-binding protein (RBP) appeared down-regulated in the comparison pro La-WT vs pro La-arg-. RBPs have been described as regulatory elements controlling the expression of genes, involving changes in mRNA stability and/or translational control. The shift from transcriptional to post-transcriptional control in trypanosomatids appears to be due to a different arrangement of protein coding genes. In addition to this specialized arrangement, the genes lack canonical promoters [40, 50, 51]. Currently, RBPs have been reported in Leishmania [52–54] and can elucidate gene expression regulation. Dipeptidyl-peptidase III (DCP) (LmxM.05.0960) appeared up-regulated in the comparison of La-WT vs. La-arg-axenic amastigotes. DCP belongs to the mono-zinc peptidase family. Peptidases of parasitic protozoa have been suggested as novel virulence factors, potential drug targets and vaccine candidates [55]. Previously, by microarray analyses, DCP was shown to be up-regulated in L. donovani amastigotes compared to promastigotes due to its correlation with an increase in total protease activity [56]. Thus, DCP may have a role in parasite differentiation related to nutrition and pathogenesis [56]. Similar to DCP, the protein disulfide isomerase (PDI) (LmxM.06.1050) appeared up-regulated in La-arg- axenic amastigotes compared to La-WT. PDI, a redox chaperone, has been primarily characterized with virulent and immunogenic potential [57, 58]. Achour et al. (2002) suggested that Leishmania promastigote growth might be due to optimal protein folding as a result of the increased secretion of PDI at the surface of the parasite [58]. Later, Amit et al. (2014) showed that alanine, an inhibitor of PDI activity, caused damage to the parasite mainly in axenic amastigote forms [57]. Tryparedoxin 1 (TXN1) (LmxM.08_29.1160) was down regulated in La-arg- axenic amastigotes compared to La-WT. TXN1 is part of the trypanothione and trypanothione reductase pathway to regulate oxidative stress [59]. The polyamine pathway can be considered metabolically important for survival and infectivity in trypanosomatids [4, 6, 60]. On the other hand, tryparedoxin peroxidase (TXNPx) (LmxM.15.1160) was up-regulated in La-WT axenic amastigotes compared to La-WT promastigotes, corroborating the results of previous studies demonstrating that this increase is necessary for detoxification of peroxides and resistance to NO in L. donovani, which did not show the same profile in the absence of arginase [61–63]. RNA helicases are central players in RNA biology and function. Similar to other eukaryotes, many biological functions have been attributed to trypanosomatid RNA helicases, including RNA degradation, translation regulation and RNA editing [64–66]. We identified that an ATP-dependent RNA helicase (LmxM.33.2050) was up-regulated in La-WT promastigotes compared to La-WT axenic amastigotes indicating a rgene regulation in Leishmania differentiation. Interestingly, the comparison of the expression profile of La-arg-promastigotes vs La-arg- axenic amastigotes showed only down-regulated genes: an α-tubulin (LmxM.13.0280), two β-tubulins (LmxM.32.0792 and LmxM.32.0794), an unspecified product (LmxM.13.0300) and a ncRNA (LmxM.30.ncRNA rfamscan: 218682-218827-1). α-tubulin is a highly conserved protein that interacts with β-tubulin, forming an α/β-tubulin heterodimer, a key to the formation of the eukaryotic cytoskeleton, which is responsible for cell shape and it is involved in many essential processes, including cell division and ciliary and flagellar motility [67, 68]. Altogether, these findings of down-regulated genes could be indicative of cytoskeleton reorganization dependent of the absence of arginase activity. The transcriptional profiling of L. amazonensis reinforces the capacity of the parasite to fine-tune gene expression regulation to adapt to changes in the environment during promastigote and amastigote differentiation. It is interesting to note that although gene expression regulation in Leishmania is considered to occur at post-transcriptional levels, we observed a correlation between the transcriptomic and metabolomics data, focused on the L-arginine pathway. Additionally, the use of the arginase knockout parasite reinforced the importance of this enzyme and provided additional insights into the coordination of gene expression and parasite development and infectivity.

10.1371/journal.pntd.0004010

Epidemiological and Molecular Characterization of Dengue Virus Circulating in Bhutan, 2013-2014

Dengue is one of the most significant public health problems in tropical and subtropical countries, and is increasingly being detected in traditionally non-endemic areas. In Bhutan, dengue virus (DENV) has only recently been detected and limited information is available. In this study, we analyzed the epidemiological and molecular characteristics of DENV in two southern districts in Bhutan from 2013–2014. During this period, 379 patients were clinically diagnosed with suspected dengue, of whom 119 (31.4%) were positive for DENV infection by NS1 ELISA and/or nested RT-PCR. DENV serotypes 1, 2 and 3 were detected with DENV-1 being predominant. Phylogenetic analysis of DENV-1 using envelope gene demonstrated genotype V, closely related to strains from northern India.

We describe the epidemiological and molecular features of DENV currently circulating in the two southwestern districts of Bhutan, demonstrating a shift in serotype dominance from previous DENV-3 (2004–2006) to current DENV-1 (2013–2014). The presence of the dengue virus in Bhutan is a relatively recent one. Unfortunately, dengue epidemiological and molecular data in this country is scarce. A fever outbreak in 2013 and 2014 saw patients seeking care at medical facilities in two district of southwestern Bhutan bordering with India. Analyses of serum specimens collected from these patients indicated that dengue virus was at least a major source of this outbreak. These specimens were analyzed in the Public Health Laboratory in Bhutan and in AFRIMS, Thailand. With a combination of three different assays, we established that 31% of all cases captured were caused by dengue virus, although the proportion was higher in 2013 than in 2014. Three different serotypes of dengue virus were found: DENV-1, -2 and -3. No DENV-4 was found. We successfully isolated DENV-1, from which was sequenced the E gene for further analyses. Our analyses revealed that the current DENV-1 in Bhutan probably originated from India.

Dengue is one of the most common infectious diseases in tropical and sub-tropical regions of the world [1, 2]. The World Health Organization (WHO) estimates 50–100 million infections per year globally; however, other studies have suggested a much higher figure [2]. Southeast Asia and Western Pacific represent about 75% of the global dengue burden [3], causing a substantial economic cost in these regions [1]. Dengue virus (DENV), the etiological agent of dengue, is divided into four genetically and antigenically different serotypes, DENV-1 to 4 [4]. Although infection by a particular serotype is known to confer long-lasting homotypic immunity, circulating heterotypic antibodies are only able to provide transient cross-protective immunity often leading to severe forms of DF, dengue hemorrhagic fever (DHF) and dengue shock syndrome (DSS) [5]. Antibody dependent enhancement (ADE) and cross-reactive T-cell responses have been postulated to explain the possible mechanisms of disease enhancement [5, 6]. Other factors such as host immunity and viral genetics may contribute to severe forms of dengue fever (DF), dengue hemorrhagic fever (DHF) and dengue shock syndrome (DSS) [4, 5]. In recent years, DENV is increasingly being detected in newer geographical areas. Dengue outbreaks are relatively new in Bhutan, a country that shares borders with India to the south and China to the north. The earliest documented dengue outbreak in the country occurred in 2004 andwas caused mainly by DENV-2 and 3 [7]. This was followed by sporadic dengue cases [8]. Although dengue is a reportable disease in Bhutan, it is believed to be inconsistently reported largely because diagnosis is clinically-based with rapid serological assays employed only in a few locations where laboratory diagnostic kits are available [9]. DENV molecular detection, isolation and other advanced testing have not yet been established in Bhutan. In addition, vector control efforts in Bhutan are mostly focused on malarial vector control, which is assumed to cover up for dengue as well. Dengue vector surveillances are in place where vectors were previously detected but no dengue-specific vector control measures have been implemented in the country [10]. Bhutan shares a 700 km border with India, which continues to report co-circulation of all 4 DENV serotypes with increasing frequency [11]. DENVs isolated in 2004–2006 from Bhutan during its first reported dengue outbreak are thought to have originated in India [7]. Similar transmission was reported in Nepal, a country with similar geographical features as Bhutan, gradually leading to endemicity [12]. Due to limited studies done in Bhutan, there is very little information regarding currently circulating DENV serotypes or their molecular and epidemiological characterization. In this study, we undertook laboratory confirmation of clinically suspected dengue patients from the southern part of Bhutan during 2013–2014 and elucidated the molecular epidemiology of DENV-1 in Bhutan. Samples were collected by the Public Health Laboratory in Bhutan as a part of routine diagnosis and surveillance; hence, no written consent was obtained from patients. The Ministry of Health (MOH), Bhutan, provided written permission for use of de-identified specimens and data for further evaluation. Approval of the study was provided by the Institutional Review Boards (IRBs) of Mahidol University, Thailand (COE. No. 2014/020/.1010), and Walter Reed Army Institute of Research (WRAIR), United States (WRAIR No. 2155). Acute blood specimens were collected over a period of two years, 2013–2014, from patients clinically suspected of having dengue. These patients had either visited the outpatient department (OPD) or were admitted to Samtse or Phuntsholing Hospitals in two southwest districts (Samtse and Chukha, respectively) of Bhutan (Fig 1). These districts are located at the foothills of the Himalayas where climate is sub-tropical. Both Samtse and Chukha share porous borders with India, where commerce and tourism are common. After the first outbreak of dengue in Phuntsholing town, Chukha district [7]; this area has continued to report dengue cases. Samtse district, which has similar climatic and ecological factors, was chosen as a site for this study along with Chukha. Other districts were not included in this study since they have had no reported cases of dengue infection. Both Aedes vectors (Ae. Aegypti and Ae. Albopictus) have been found in both Chukha and Samtse districts. These districts also have reasonable access to the Public Health Laboratory in the capital city, Thimphu, where specimens can be shipped for further evaluation Clinically suspected dengue was defined as fever (oral, rectal or axillary temperature ≥38°C), or history of fever lasting 2 to 7 days of unknown origin with two or more of the following: headache, retro-orbital pain, myalgia, arthralgia, rash, hemorrhagic manifestation and leucopenia [13]. Patient demographic and clinical information was collected by attending clinicians. The clinical diagnosis of dengue by attending clinicians was not further categorized as to disease severity. Laboratory confirmation of dengue was carried out by DENV specific NS1 antigen and IgM ELISA, and nested RT-PCR. All acute serum specimens were tested for DENV infection by NS1 antigen detection, IgM ELISA and nested RT-PCR. Both dengue NS1 antigen and dengue IgM detection were carried out at the Public Health Laboratory, Bhutan using DENV Detect NS1 ELISA (InBios, Seattle, Washington) and Dengue Virus IgM ELISA (Calbiotech). Tests were performed according to manufacturer’s instructions. Acute serum specimens were also tested by nested RT-PCR at the Armed Forces Research Institute of Medical Sciences (AFRIMS) in Bangkok, Thailand. Viral RNA was extracted from 140 μl of sera using QIAamp viral RNA mini kit (QIAGEN, Germany) following the manufacturer’s instructions. Nested RT-PCR was performed using a method modified from Lanciotti et al as previously described [14, 15]. One step RT-PCR was carried out using AMV reverse transcriptase (Promega, Madison, WI, USA) and AmpliTaq DNA polymerse (Life Technologies, USA) in the first round PCR. Nested PCR was performed using AmpliTaq DNA polymerase in the second round PCR. All sera positive for DENV by nested RT-PCR were inoculated into freshly prepared mono-layers of C6/36 cells grown in Minimum Essential Medium (MEM, GIBCO) containing 10% heat inactivated fetal bovine serum (HIFBS), 1% Glutamine and 1% Penicillin and streptomycin. These cultures were maintained in maintenance medium (MM) containing RPMI with 5% HIFBS. A mock-infected C6/36 cell flask was included as a negative control. Cells underwent 3 passages and were observed for cytopathic effect (CPE). Identification of DENV serotypes was carried out by antigen capture ELISA as previously described [16, 17]. Molecular confirmation of the isolates was performed by extracting DENV RNA from the cell culture supernatant followed by nested RT-PCR. When cell-based DENV isolation was not possible, nested RT-PCR positive sera were inoculated into Toxorhynchitis splendens mosquitoes (0.3μl/ mosquito) as previously described [18]. Surviving mosquitoes were head squashed on microscopic slides and screened for flavivirus antigen by immunofluorescent antibody (IFA) staining. Virus isolates amplified from cell culture or mosquitoes were used in envelope (E) gene sequencing. E gene of DENV was amplified using one-step RT-PCR amplification [19]. Overlapping fragments were amplified using AccessQuick RT-PCR System (Promega, Madison, WI, USA) with two sets of primers covering the entire E gene. Amplified products were purified prior to sequencing using QIAquick PCR purification kit (QIAGEN) following manufacturer’s instructions. Capillary-based Sanger sequencing was used to obtain E gene sequences (1,485 bp). Base correction for the obtained sequences was performed using Sequencher 5.1. All new sequences were submitted to GenBank (accession numbers KP849860- KP849892). Maximum likelihood (ML) tree was constructed from 33 new Bhutan DENV-1 sequences along with sequences of 56 global DENV-1 and 3 vaccine strains downloaded from GenBank. The tree was constructed using MEGA v.6.0 (www.megasoftware.net) [20]; the Tamura Nei (TN93) model was chosen for nucleotide analysis. A sylvatic strain from Malaysia (accession no. AF425622) was used as the outgroup. Bootstrap value was obtained from 1000 replicates. Selection pressure among the Bhutan DENV-1 sequences was determined using the maximum likelihood approach of codon based test of selection available in Mega v.6.0. Percent identity of nucleotides and amino acids was calculated by Clustal W function available in MegAlign v.5.05 of DNASTAR package. Accession numbers for E gene sequenced in this study: KP849860, KP849861, KP849862, KP849863, KP849864, KP849865, KP849866, KP849867, KP849868, KP849869, KP849870, KP849871, KP849872, KP849873, KP849874, KP849875, KP849876, KP849877, KP849878, KP849879, KP849880, KP849881, KP849882, KP849883, KP849884, KP849885, KP849886, KP849887, KP849888, KP849889, KP849890, KP849891 and KP849892. SPSS version 22 was used for statistical analysis. Independent T-test was used to compare means of various attributes. Frequencies/ percentage of clinical symptoms were compared using Pearson chi-square test. A probability value of p < 0.05 was considered statistically significant. A total of 379 acute sera from suspected dengue cases were collected at the district hospitals in Samtse and Chukha (Fig 1) during 2013 and 2014, of which 119 were laboratory confirmed for DENV infection. In both years, the number of suspected and laboratory confirmed cases peaked during the summer months (June-September), which also corresponds to the monsoon season in Bhutan (Fig 2). Clinically suspected and laboratory confirmed dengue cases during the summer months accounted for 278/379 (73%) and 83/119 (69.7%) respectively, for the entire two years. During the colder months (November to March), only 17/379 (4.5%) of all suspected dengue and 8/119 (6.7%) of laboratory confirmed dengue cases occurred. There seemed to be a remarkable difference in the proportion of laboratory confirmed dengue in 2013 and 2014. In 2013, 100/168 (59.5%) of suspected cases were confirmed to be dengue by laboratory methods, accounting for 84% of all laboratory confirmed dengue. Only 19/211 (9%) of suspected cases were laboratory confirmed in 2014. Age of patients in this study ranged from 2 to 77 years but young adults, 19–35 years, accounted for the largest age group of suspected as well as laboratory confirmed cases; 178/379 (47%) and 68/119 (55.6%), respectively. Both genders were about equally affected (female to male ratio of 1:1.03). Mean age and days of illness (DOI) after onset of symptoms until blood collection were calculated separately for total suspected and laboratory confirmed cases (Table 1). We did not observe any differences among the two groups. Of the 379 suspected dengue patients, 364 visited the outpatient department (OPD) and the remaining 15 were either admitted or visited the emergency room. Clinical data was collected from all 379 patients. All patients had fever and most had features of DF such as headache, myalgia and joint pain (Table 2). Although we were unable to obtain complete information regarding the severity and classification of the disease, hemorrhagic manifestations characteristic of DHF were noted in some laboratory-confirmed dengue cases including petechiae, gastrointestinal bleeding (e.g., haematemesis and melena), and bleeding from the mucosa and/or other sites (Table 2). A total of 97/379 (25.6%) specimens were positive by NS1 ELISA, 29/ 379 (7.6%) specimens were positive by IgM ELISA and 58/379 (15.3%) positive by nested RT-PCR. Combined, 119/ 379 (31.4%) specimens were positive by combination of these methods. The mean DOI was calculated separately for both NS1 and nested RT-PCR positive cases (Fig 3). Specimens positive for only NS1 had a DOI of 3.6 days, which was significantly longer than the DOI from samples positive for only nested RT-PCR (2.6 days, p<0.05). DOI for cases that were both NS1 and RT-PCR positive was 2.9 days. The DOI for IgM positive cases was 4.4 days. Nested RT-PCR was performed on all sera collected. Of 58 positive cases, 53 (91.4%) were DENV-1, 3 (5.2%) DENV-2 and 2 (3.4%) DENV-3; no DENV-4 was detected. Using cell culture and/or mosquito amplification, isolation of DENV was attempted on all 58 sera in order to obtain sufficient sequencing material. Unfortunately, we were unable to isolate any of the DENV-2 and DENV-3 viruses. Nevertheless, 33 viruses (all DENV-1) were successfully isolated and sequenced ML tree was generated using 92 E gene sequences (1,485 bp), including the 33 Bhutan DENV-1 strains reported here, 3 DENV-1 vaccine candidate strains and 56 global DENV-1 strains obtained from GenBank (Fig 4). All 33 Bhutan sequences group to genotype V, using the classification of Weaver et al [21], and were located in the same group as sequences from northern India, categorized as clade IX by Dash et al [22]. Within clade IX, 32 of the Bhutan DENV-1 E gene sequences (GenBank accession no. KP849860-7 and KP849869-92) grouped with the northern Indian sub-clade from 2008–2009, while 1 Bhutan DENV-1 sequence (GenBank accession no. KP849868) grouped with the northern Indian sub-clade from 2010–2011. None of the vaccine candidate strain sequences included in our phylogenetic tree fell within the same group as the Bhutan DENV-1. Considering the diversity observed among the Bhutan DENV-1 E-gene sequences, we calculated the percentage of nucleotide and amino acid identity between the two observed sub-clades of Bhutanese specimens. We found 99.3–99.5% nucleotide identity and 99.6–99.8% amino acid identity. An I461V amino acid substitution was the only mutation found in the KP849868 sequence that differentiated it from the rest of the Bhutan sequences. This mutation was also found in an Indian sequence from 2010 (GenBank accession no. JN415486), but in no other sequence within the same clade. Selection pressure analysis using maximum likelihood approach showed that amino acids of Bhutan DENV-1 E region are under negative (purifying) selection with test statistics (dS-dN) = 2.7 bootstrapped with 1000 replicates. This study demonstrated a higher number of cases during the summer season, especially affecting young adults, 19–35 years of age. Clinical presentation of these patients ranged from classical dengue fever to hemorrhagic manifestations. We established DENV-1 as the dominant of at least three DENV serotypes currently circulating in Bhutan. With a shift in predominant serotype from DENV-3 documented previously to DENV-1, dengue in Bhutan seems to follow a cyclical pattern as seen in other countries [23]. Phylogenetic characterization of DENV-1 revealed that they belong to genotype V [21], and were probably imported from India. A distinct seasonality for DENV infection in the 2 southwest districts of Bhutan was observed. The number of cases was found to peak during the hot and humid monsoon season, probably due to increased habitats for mosquito breeding. Studies have shown a correlation between rainfall, temperature and humidity with serologically confirmed dengue [24, 25]. These climatic and ecological factors in combination with inadequate public services and ineffective vector control are known to contribute to dengue endemicity [24], all of which may be playing a role in Bhutan. The majority of suspected and laboratory confirmed cases occurred in young adults aged 19–35 years. The same was observed in the previous study in Bhutan [7] where the mean age was found to range from 28 to 32 years of age. While the underlying pathogenesis leading to more symptomatic dengue in adults than in children is not clear, dengue as a disease of adults is supported by increasing DF notifications from many ageing countries [26, 27]. Ineffective vector control efforts have been proposed as a reason for causing shift in the age group from children to adults in countries like Singapore and Thailand [28, 29]. However; in the case of Bhutan, the majority of cases being adults suggest that there are still many dengue-naïve or dengue monotypic individuals among the adult population in Bhutan. Whether this mean age pattern changes in the future may depend on ongoing and future intensity of DENV transmission. Considering the difference in sensitivity of NS1 ELISA and RT-PCR methods at various stages of illness [30], using both methods was useful in detecting cases that were positive by just one of the methods. As expected, results by each method varied and seemed to correlate with the number of days of illness after onset of symptoms. Interestingly, the percentage of NS1 ELISA-positive specimens (25.6%) was higher than the percentage of RT-PCR positive specimens (15.3%). It is unlikely that this is due to higher detection sensitivity by the NS1-ELISA method. One possible explanation is the preponderance of patients in this study seeking clinical care after several days of illness, favoring the detection of NS1, which circulates in the serum for longer periods than viral RNA [31]. The mean DOI among patients from whom virus genome could be detected was shorter than patients from whom only NS1 antigen could be detected (Fig 3). It is also possible that the commercial NS1 ELISA test, done in the laboratories in Bhutan and not confirmed at AFRIMS, provided a number of false positive results, accounting for a somewhat inflated positive percentage A large numbers of specimens were negative for DENV by the laboratory methods used in this study, especially in 2014. One possible hypothesis could be that non-specific febrile illness, having fever with rash, can be caused by a number of other infectious diseases that may commonly be mistaken for dengue [32]. For example, an outbreak of chikungunya, which commonly manifests as fever and polyarthralgia, was recently reported in the same geographical location in Bhutan [33, 34]. Clinical manifestations that were observed at significantly higher frequencies in laboratory confirmed dengue as compared to the suspected dengue cases include petechiae and positive tourniquet test, symptoms that are specific to DENV infection. We have yet to confirm the etiology of the dengue-like illnesses in 2014. Other factors such as time of sampling, loss in viral titer as a result of long storage and transportation with several inevitable freeze- thaw cycles (caused by Bhutan’s poor road infrastructure and transportation networks), could have contributed to lesser DENV isolation and further sequencing in general. The report of the first dengue outbreak in Bhutan suggested that DENV was imported from India, probably as early as 2004 [7]. DENV-3 was the predominant serotype in both Delhi (northern India) during 2003–2006 and Bhutan during 2004–2006 [7, 11, 22, 32]. Since 2006, DENV-1 prevalence has increased in India [22, 35, 36]. Similarly, in this study, DENV-1 was predominant in 2013–2014, further supporting the notion that dengue epidemiological patterns in Bhutan and India are closely linked with likely exchange of strains between the two countries. Phylogenetic analysis of the 33 Bhutan DENV-1 isolates showed that all belong to DENV-1 genotype V, also known as the cosmopolitan or the American-African genotype because of its diversity [21, 22]. The Bhutan sequences were found to group with clade IX from northern India, the most recent Indian clade [22]. Inclusion of the Bhutan sequences from our study in the phylogenetic analysis resulted in the divergence of clade IX into two different sub-clades. One sub-clade consisted of sequences from 2008–2009 northern India, and the other of sequences from 2010–2011 northern India. Most of the Bhutan sequences grouped with the 2008–2009 northern Indian sequences, indicating that the 2013–2014 dengue viruses from Bhutan mainly originated from 2008–2009 northern India sub-clade viruses. Only one sequence (GenBank accession no. KP849868) grouped with the 2010–2011 northern India sub-clade. This virus was obtained from a specimen collected in August 2013 in Chukha district and was found to be associated with mild clinical symptoms (no hemorrhagic manifestations). It is difficult to ascertain whether the differences in these two sub-clades may have any effect on the severity of the disease since we had only one virus in the 2010–2011 northern India sub-clade. None of the vaccine candidate strains grouped in the same genotype as the Bhutan DENV-1 sequences. It is unknown whether these discrepancies would have any impact on the efficacy of possible future vaccinations in the region. DENV from Bhutan showed a strong negative (purifying) selection pressure which is usually observed in arboviruses [37]. This has been attributed to evolutionary constraints that make the viruses resistant to change, especially since their lifecycles include both vertebrate and invertebrate hosts. There were several limitations to our study. Since we used pre-existing specimens and data, the available information was not complete, especially regarding disease severity, high risk occupational groups and certain clinical laboratory data such as blood counts. We also did not have access to convalescent sera and, therefore, could not confirm the non-acute dengue cases. Furthermore, this study was limited to only two hospitals in Bhutan so the data presented here may not necessarily represent the entire country. Despite its relative isolation, Bhutan is seeing increased urbanization and travel, factors that have led to DENV of all serotypes co-circulating in many other countries [2, 38]. WHO has at least partially attributed outbreaks in Nepal and Bhutan to increasing global temperatures [10]. Although Bhutan is mostly situated at elevations of greater than a thousand meters above sea level, some areas of Chukha and Samtse districts are located as low as 100 meters above sea level, bearing climates that favor transmission of DENV and other tropical pathogens [39]. More importantly, these two districts share an unrestricted border with India which regularly reports dengue outbreaks of all serotypes [11], contributing to the DENV presence in Bhutan. In this study, we report a shift in predominant DENV serotype in Bhutan in concert with prevailing patterns in neighboring India, along with epidemiological features of dengue in Bhutan. The southern part of Bhutan has likely become dengue endemic, hence enhanced and continuous dengue surveillance is required to generate more robust epidemiological data and to monitor for changes in the characteristics of circulating DENVs.

10.1371/journal.pbio.1002387

Tel1 and Rif2 Regulate MRX Functions in End-Tethering and Repair of DNA Double-Strand Breaks

The cellular response to DNA double-strand breaks (DSBs) is initiated by the MRX/MRN complex (Mre11-Rad50-Xrs2 in yeast; Mre11-Rad50-Nbs1 in mammals), which recruits the checkpoint kinase Tel1/ATM to DSBs. In Saccharomyces cerevisiae, the role of Tel1 at DSBs remains enigmatic, as tel1Δ cells do not show obvious hypersensitivity to DSB-inducing agents. By performing a synthetic phenotype screen, we isolated a rad50-V1269M allele that sensitizes tel1Δ cells to genotoxic agents. The MRV1269MX complex associates poorly to DNA ends, and its retention at DSBs is further reduced by the lack of Tel1. As a consequence, tel1Δ rad50-V1269M cells are severely defective both in keeping the DSB ends tethered to each other and in repairing a DSB by either homologous recombination (HR) or nonhomologous end joining (NHEJ). These data indicate that Tel1 promotes MRX retention to DSBs and this function is important to allow proper MRX-DNA binding that is needed for end-tethering and DSB repair. The role of Tel1 in promoting MRX accumulation to DSBs is counteracted by Rif2, which is recruited to DSBs. We also found that Rif2 enhances ATP hydrolysis by MRX and attenuates MRX function in end-tethering, suggesting that Rif2 can regulate MRX activity at DSBs by modulating ATP-dependent conformational changes of Rad50.

Many tumors contain mutations that confer defects in repairing DNA double-strand breaks (DSBs). In both yeast and mammals, the MRX/MRN complex (Mre11-Rad50-Xrs2 in yeast; Mre11-Rad50-Nbs1 in mammals) plays critical functions in repairing a DSB by either nonhomologous end joining (NHEJ) or homologous recombination (HR). Furthermore, it recruits the checkpoint kinase Tel1/ATM. Although ATM is considered to be a tumor suppressor, up-regulation of ATM signaling promotes chemoresistance, radioresistance and metastasis. For this reason, cancer therapies targeting ATM have been developed to increase the effectiveness of standard genotoxic treatments and/or to set up synthetic lethal approaches in cancers with DNA repair defects. We aimed to identify the precise role of ATM/Tel1 in these processes. By performing a synthetic phenotype screen, we identified a mutation (rad50-V1269M) altering the Rad50 subunit of the MRX complex, which sensitizes cells lacking Tel1 to genotoxic agents. Genetic and biochemical characterization of MRV1269MX protein complex revealed that Tel1 promotes MRX association at DSBs to allow proper MRX-DNA binding that is needed for DSB repair. The role of Tel1 in promoting MRX retention on DSBs is counteracted by Rif2, which can regulate MRX activity at DSBs by modulating ATP-dependent conformational changes in Rad50. Our finding that MRX dysfunctions can be synthetically lethal with Tel1 loss in the presence of genotoxic agents suggests that ATM inhibitors could be beneficial in patients whose tumors have defective MRN functions.

DNA double-strand breaks (DSBs) are among the most cytotoxic DNA lesions, because failure to repair them can lead to genome instability. DSBs can be repaired by either nonhomologous end joining (NHEJ) or homologous recombination (HR). While NHEJ directly ligates the DNA ends, HR requires the 5′ ends of a DSB to be nucleolytically processed (resected) to generate 3′ single-stranded DNA (ssDNA) tails that initiate HR by invading an undamaged homologous DNA template [1]. Generation of DSBs activates a DNA damage response (DDR), which regulates DSB repair and coordinates it with cell cycle progression [2]. In both yeast and mammals, the DDR is initiated by the MRX/MRN complex (Mre11-Rad50-Xrs2 in yeast; Mre11-Rad50-Nbs1 in mammals), which recognizes unprocessed DSBs and activates the checkpoint kinase Tel1/ATM [3]. MRX/MRN recruits Tel1/ATM to the DSB ends through its interaction with the C-terminal domain of Xrs2/Nbs1 and stimulates Tel1/ATM catalytic activity [4–8]. MRX/MRN also plays critical functions in DSB resection and in maintaining the DSB ends tethered to each other [9]. Several studies have shown that the MRX complex consists of a globular head domain from which the long coiled-coil domain of Rad50 protrudes [10–13]. The coiled-coil apex contains a CXXC amino acid motif that can dimerize via tetrahedral coordination of a zinc ion, thereby forming molecular bridges for keeping the DNA ends tethered to each other [14,15]. Mre11 is active as an exo- and endonuclease in vitro [16–19] and initiates DSB resection [20–24]. The functions of MRX in end-tethering and DSB resection are regulated by Rad50, whose ATP binding and hydrolysis activities result in MRX conformational changes [11,25–28]. Mutants that promote the ATP-bound conformation of Rad50 exhibit a higher level of tethering [29], indicating that end-tethering depends on this MRX conformation. In turn, the ATP-bound conformation sterically blocks the Mre11 nuclease activity [29–32], whereas release from this ATP-bound state that occurs with ATP hydrolysis opens Mre11 nuclease active sites so that they can be engaged in DSB resection [13]. Thus, ATP hydrolysis triggers a switch between a closed state, in which Mre11 nuclease domain is occluded, to an open configuration with exposed Mre11 nuclease sites. In addition to its role in DSB repair, MRX works in the same epistasis group of Tel1 to maintain telomere length [33,34]. Interestingly, the lack of Tel1 in Saccharomyces cerevisiae cells causes telomere shortening and a decrease of MRX binding at DNA ends flanked by telomeric DNA repeats [35,36]. On the other hand, telomere length is negatively regulated by Rif2, which is recruited to telomeric DNA ends by Rap1 [37]. Artificial tethering of Rif2 at DNA ends reduces the amount of telomere-bound Tel1, but not that of MRX [35]. This observation, together with the finding that Rif2 appears to compete with Tel1 for binding to the C-terminus of Xrs2 in vitro [35], suggests that Rif2 interferes with MRX-Tel1 interaction to shelter telomeric ends from Tel1 recognition. Although Tel1 is recruited to DSBs and participates in DSB end resection [4,38], its function in DSB repair remains enigmatic because Tel1-deficient S. cerevisiae cells do not show obvious hypersensitivity to DNA damaging agents and are not defective in checkpoint activation in response to a single DSB [38]. To better understand the function of Tel1 in the cellular response to DSBs, we performed a genetic screen aimed at identifying mutants that require Tel1 to survive to genotoxic treatments. We found that the rad50-V1269M allele makes tel1Δ cells hypersensitive to DNA damaging agents. The MRV1269MX complex associates poorly to a DSB and the lack of Tel1 further reduces its retention at DSB ends. As a consequence, rad50-V1269M tel1Δ cells are severely defective in maintaining the DSB ends tethered to each other. These findings indicate that Tel1 promotes proper MRX association to DNA ends, and this function is required to support the end-tethering activity of MRX. The Tel1 function in promoting MRX retention to DSBs is counteracted by Rif2, which is recruited to DSB ends. Rif2 also enhances MRX ATPase activity and attenuates MRX function in end-tethering, suggesting that it modulates MRX function not only by inhibiting MRX association to DSBs but also by regulating ATP-dependent Rad50 conformational changes. To gain insights into the role of Tel1 at DSBs, we searched for mutations that caused hypersensitivity to DNA damaging agents only in the absence of Tel1. For this purpose, tel1Δ clones were screened for decreased viability in the presence of camptothecin (CPT) and/or phleomycin. Hypersensitive tel1Δ clones that lost the DNA damage hypersensitivity after transformation with a plasmid containing wild-type TEL1 were crossed to a wild-type strain followed by sporulation and tetrad analysis to verify that the DNA damage hypersensitivity was due to the combination of tel1Δ with a mutation in an unknown single gene. This procedure allowed us to identify five single-gene mutations belonging to three distinct allelism groups. Genome sequencing of the clone that showed the most severe synthetic phenotype and subsequent genetic analyses established that the mutation responsible for the DNA damage hypersensitivity of tel1Δ cells was a single nucleotide change in the RAD50 gene, resulting in substitution of valine 1269 with methionine in the C-terminal ATPase domain (Fig 1A). Both rad50-V1269M and tel1Δ single mutant cells were as sensitive as wild type to phleomycin, methyl methanesulfonate (MMS), and low CPT doses, while the sensitivity to the same drugs was greatly increased in tel1Δ rad50-V1269M double mutant cells (Fig 1B), indicating that the Rad50-V1269M variant requires Tel1 to support cell viability in the presence of genotoxic stress. As Tel1 is a protein kinase, we asked whether the rad50-V1269M allele also exacerbated the sensitivity to DNA damaging agents of cells expressing a Tel1 mutant variant (Tel1-kd) carrying G2611D, D2612A, N2616K, and D2631E amino acid substitutions that abolished Tel1 kinase activity in vitro [39]. Telomeres in tel1-kd cells are shorter than in wild-type cells and indistinguishable from those of tel1Δ cells [39], indicating that these mutations abolish Tel1 function at telomeres. Surprisingly, the viability of tel1-kd rad50-V1269M double mutant cells in the presence of DNA damaging agents was similar to wild-type cells (Fig 1C), suggesting that Rad50-V1269M mutant variant requires the presence of Tel1 but not its kinase activity to support cell viability in the presence of genotoxic stress. Rad50 binds DNA and has ATPase activity [27]. These functions reside in the globular domain formed by the N- and C-termini of the protein, which are separated by an antiparallel coiled-coil domain [9]. The V1269M mutation is very closed to the H-loop (Fig 1A), whose histidine residue has been proposed to promote ATP hydrolysis by positioning the first water molecule needed for the reaction and/or by forming a catalytic dyad with the Walker B glutamate [11,40]. Thus, we asked whether and how the rad50-V1269M mutation affects MRX ATPase and/or DNA-binding activities. The Rad50 and the Rad50-V1269M proteins were purified to near homogeneity by following our published procedure (Fig 2A) [18]. Purified Rad50 and Rad50-V1269M were then individually incubated with Mre11 and Xrs2, and the fully assembled complexes were separated from free proteins by gel filtration. Rad50-V1269M could be expressed to the same level as the wild-type protein, behaved well chromatographically, and yielded the same amount of trimeric complex with Mre11 and Xrs2. As shown in Fig 2A, the stoichiometry of the three components in the MRV1269MX mutant complex was very similar to that of the wild-type MRX complex. As we reported previously [41], Rad50 hydrolyzed ATP only within the context of the MRX complex. The MRV1269MX mutant complex exhibited a reduced ATPase activity compared to wild-type MRX (Fig 2B), indicating that the rad50-V1269M mutation affects ATP hydrolysis. Aside from Xrs2 and Mre11, Rad50 also binds DNA [41], and we found that Rad50-V1269M is compromised for DNA binding in vitro (Fig 2C). Subsequently, we examined DNA binding by MRX wild-type and MRV1269MX mutant complexes and noticed insignificant difference between the two complexes (Fig 2D). We note that the DNA binding deficiency of the Rad50-V1269M mutant may be masked by the DNA binding attribute of Mre11 and Xrs2 within the MRX complex [19,41]. We next analyzed DSB association of Rad50-V1269M and MRV1269MX in vivo by chromatin immunoprecipitation (ChIP) followed by quantitative real-time PCR (qPCR). To generate a single DSB at a specific chromosomal locus, we used a strain expressing a galactose-inducible HO endonuclease. In this strain, induction of HO by galactose addition leads to the generation at the MAT locus of a single DSB that cannot be repaired by HR because the strain carries the deletion of the homologous donor loci HMLα and HMRa [42]. Consistent with the finding that the Rad50-V1269M mutant variant is compromised in DNA binding (Fig 2C), the amount of Rad50-V1269M bound at the HO-induced DSB was lower than that of wild-type Rad50 (Fig 3A). Furthermore, although binding to DNA of the MRV1269MX mutant complex was not affected (Fig 2D), the amount of Mre11 associated to the HO-induced DSB was significantly lower in rad50-V1269M than in wild-type cells (Fig 3B). This decreased Rad50-V1269M and MRV1269MX association to the DSB is not due to reduced protein levels or altered MRV1269MX complex formation. In fact, protein extracts from wild-type and rad50-V1269M cells contained very similar amounts of Rad50, Rad50-V1269M, and Mre11 proteins (Fig 3C). Furthermore, equal amount of Rad50-V1269M and Rad50 could be immunoprecipitated with the Mre11 protein (Fig 3D). As MRV1269MX binding to DNA was not significantly affected (Fig 2D), these data indicate that the rad50-V1269M mutation impairs MRX retention to DSBs. As expected from the previous finding that MRX is required to load Tel1 at the DSB ends [4], the amount of Tel1 bound at the HO-induced DSB was lower in rad50-V1269M cells than in wild type (Fig 3E). This attenuated Tel1 association to the DSB is not due to either reduced Tel1 level in rad50-V1269M cells or impaired MRV1269MX-Tel1 interaction. In fact, similar Tel1 amounts could be detected in protein extracts prepared from wild-type and rad50-V1269M cells (Fig 3F). Furthermore, wild-type and mutant MRX complexes could be coimmunoprecipitated equally well with Tel1 (Fig 3G). The MRX complex plays multiple functions in DSB repair: it promotes DSB resection and checkpoint activation [20,21,38] and it keeps the DSB ends tethered to each other [43–46]. The severe DNA damage hypersensitivity of tel1Δ rad50-V1269M cells is not due to defect in DNA damage-induced checkpoint activation, as wild-type and tel1Δ rad50-V1269M cells phosphorylated the checkpoint kinase Rad53 with similar kinetics in response to phleomycin (Fig 4A) or MMS treatment (Fig 4B). Furthermore, tel1Δ rad50-V1269M cells phosphorylated Rad53 with wild-type kinetics in response to an irreparable HO-induced DSB (Fig 4C). Recent data indicate that MRX in the ATP-bound state promotes end-tethering, whereas ATP hydrolysis opens the MRX conformation to promote Mre11 nuclease activity and DSB resection [13,29]. As MRV1269MX exhibits reduced ATPase activity and MRX variants impaired in ATP hydrolysis are endonuclease-defective [27,47], we asked whether rad50-V1269M cells were defective in DSB resection. To monitor directly the generation of ssDNA at the DSB ends, we used strains expressing a galactose-inducible HO endonuclease, which generates at the MAT locus a single irreparable DSB [42]. Resection of the HO-induced DSB renders the DNA sequence flanking the HO break resistant to cleavage by restriction enzymes, resulting in the appearance of resection intermediates that can be detected by Southern blot analysis with a probe that anneals to the 3′ end at one side of the break. Because resection was much slower in cells arrested in G2/M than in replicating cells [48], HO was induced by galactose addition to cell cultures that were arrested in G2 with nocodazole and kept blocked in G2 by nocodazole treatment to detect even subtle differences in resection efficiency. Consistent with MRV1269MX deficiencies in ATP hydrolysis and DNA binding, rad50-V1269M mutant cells showed a slight defect in DSB resection compared to wild-type cells (Fig 4D). Importantly, the lack of TEL1, which caused per se a slight delay in DSB resection (Fig 4D) [38], did not reduce further the resection efficiency of rad50-V1269M cells, as rad50-V1269M and tel1Δ rad50-V1269M cells resected the DSB with similar kinetics (Fig 4D). These findings indicate that the severe DNA damage sensitivity of tel1Δ rad50-V1269M double mutant cells is not due to a resection defect. Further supporting this conclusion, EXO1 overexpression, which suppresses the hypersensitivity to DNA damaging agents and the DSB resection defect of mre11Δ cells [38,49], did not suppress the hypersensitivity to DNA damaging agents of tel1Δ rad50-V1269M double mutant cells (Fig 4E). MRX function in end-tethering is largely dependent on Rad50 coiled-coil domains [12,14]. Nonetheless, structural studies suggest that also DNA binding of the globular domain is important for end-tethering, possibly because it increases intercomplex hook-hook dimer formation by causing the Rad50 coils to become more rigid and parallel to one another [12]. As the MRV1269MX mutant complex was poorly recruited to the DSB (Fig 3A and 3B), we asked whether rad50-V1269M and tel1Δ rad50-V1269M cells were defective in keeping the DSB ends tethered to each other. To detect the association of broken DNA ends, we used a yeast strain where the DNA proximal to an irreparable HO-induced DSB can be visualized by binding of a LacI-GFP fusion protein to multiple repeats of the LacI repressor binding site (LacO) that are integrated on both sides of the HO cleavage site on chromosome VII at a distance of 50 kb [43]. HO was induced by galactose addition to cell cultures that were arrested in G2 with nocodazole and kept blocked in G2 by nocodazole treatment in order to ensure that all cells would arrest in metaphase. The majority of wild-type cells showed a single LacI-GFP focus both before and after HO induction, indicating their ability to hold the broken DNA ends together (Fig 4F). Consistent with previous results [45], tel1Δ cells showed a slight increase of two LacI-GFP spots at 1–3 h after HO induction (Fig 4F). An increase of two LacI-GFP spots compared to the uninduced condition could be detected also in rad50-V1269M cells (Fig 4F). Strikingly, the number of cells showing two LacI-GFP spots after HO induction was greatly increased in tel1Δ rad50-V1269M double mutant cells compared to each single mutant, reaching a percentage similar to that observed in mre11Δ cells (Fig 4F). The MRX complex has been implicated in sister chromatid cohesion [50], prompting us to evaluate whether the increase frequency of two LacI-GFP foci after HO induction in tel1Δ rad50-V1269M cells was due to end-tethering and/or cohesion defects. We therefore induced HO expression in α-factor-arrested cells that were kept arrested in G1 by α-factor in the presence of galactose. About 50% of G1-arrested tel1Δ rad50-V1269M cells showed two LacI-GFP foci 1 h after HO induction similarly to mre11Δ cells (Fig 4G), indicating that the appearance of two LacI-GFP foci in these cells is primarily due to defective end-tethering. We also monitored the ability of tel1Δ, rad50-V1269M, and tel1Δ rad50-V1269M cells to maintain cohesion between sister chromatids by determining formation of LacI-GFP foci in nocodazole-arrested cells in the absence of HO induction. Under these conditions, the amount of tel1Δ and rad50-V1269M cells showing two LacI-GFP foci was similar to that found in wild-type cells (Fig 4H), indicating that cohesion is not affected by either the lack of Tel1 or the presence of Rad50-V129M variant. By contrast, a slight cohesion defect was detectable in nocodazole-arrested tel1Δ rad50-V1269M cells, which showed a ~5% increase of two LacI-GFP foci compared to wild-type cells (Fig 4H). Because the ability of the above strains to held together the DSB ends was determined by using target sequences integrated at a distance of 50 kb from the DSB, we also monitored end-tethering by using a strain expressing LacI-YFP and TetR-RFP fusion proteins, which bind LacO and TetO tandem arrays, respectively [51]. These arrays are integrated at a distance of 7 kb from the DSB that is generated by the endonuclease HO on chromosome III, with each kind of array marking one specific side of the break. The frequency of cells showing separated LacI-YFP and TetR-RFP foci dramatically increased after HO induction in G2-arrested tel1Δ rad50-V1269M cells compared to wild-type cells (Fig 4I), confirming that tel1Δ rad50-V1269M cells are defective in end-tethering. Thus, the absence of Tel1 severely reduces the end-tethering activity of MRV1269MX, indicating a role for Tel1 in supporting this MRX function. The maintenance of the DSB ends tethered to each other is a relevant event in the repair of a DSB by both NHEJ and HR [43,44,46,52]. Thus, we asked whether tel1Δ rad50-V1269M cells were defective in HR and/or NHEJ. Among the HR pathways, single-strand annealing (SSA) is devoted to repair a DSB that is flanked by direct repeats and requires resection of the DSB ends followed by Rad52-dependent annealing of the resulting complementary ssDNA sequences [53]. To investigate possible HR defects, we first monitored the ability of tel1Δ rad50-V1269M cells to repair a DSB by SSA. To this end, we used a strain carrying a galactose-inducible GAL-HO construct, as well as tandem repeats of the LEU2 gene, with a recognition site for the HO endonuclease adjacent to one of the repeats (S1A Fig) [54]. Galactose was added to G2-arrested cells to induce HO production and it was maintained in the medium so that continuously produced HO could re-cleave the HO sites eventually reconstituted by NHEJ. When kinetics of DSB repair was monitored by Southern blot analysis with a LEU2 probe, accumulation of the 8 kb SSA repair product was slightly delayed in both tel1Δ and rad50-V1269M single mutants, whereas it was severely defective in tel1Δ rad50-V1269M double mutant compared to wild-type cells (S1B and S1C Fig). This finding indicates that Tel1 is important to support MRX function in DSB repair by SSA. The observation that tel1Δ, rad50-V1269M and tel1Δ rad50-V1269M cells all delay resection to the same extent (Fig 4D) indicates that the SSA defect of tel1Δ rad50-V1269M cells cannot be explained by a resection defect. We noticed that all the galactose-induced cell cultures exhibited a DNA band that migrated slower than the uncut band and appeared concomitantly with the SSA products (S1C Fig). This band was not detectable in rad51Δ cells, indicating that it was generated by Rad51-mediated recombination events (S1D Fig). Because SSA repair pathway does not involve strand invasion and therefore does not require the recombination protein Rad51 (S1D Fig) [55], we also monitored the HR events that depend on the Rad51-dependent invasion and pairing of broken DNA ends with intact homologous sequences present on a sister chromatid or at an ectopic location in the genome. In the major HR pathway, the 3′-ended ssDNA tail invades an intact duplex homologous, creating a loop structure (D-loop) consisting of a region of heteroduplex DNA and displaced ssDNA. If this ssDNA anneals with the complementary sequence on the other side of the DSB (second end capture), subsequent extension and ligation result in the formation of a double Holliday junction intermediate, whose random cleavage yield an equal number of noncrossover (NCO) and crossover (CO) products [53]. Alternatively, if the newly synthesized strand is displaced, it can anneal with the 3′ ssDNA end at the other end of the DSB. This event leads to the generation of NCO products in a process called synthesis-dependent strand-annealing (SDSA) [56–58]. To monitor CO and NCO formation, we used haploid strains that bear two copies of the MATa sequence. A MATa gene introduced in chromosome V can be cleaved by a galactose-inducible HO endonuclease and repaired by Rad51-dependent HR using a uncleavable MATa donor on chromosome III that contains a single base pair substitution preventing HO cleavage (MATa-inc) (Fig 5A) [59]. This repair event can lead to NCO and CO outcomes, with the proportion of COs being ~5% among the overall repair events [59]. Galactose was added to induce HO production and then it was maintained in the medium to cleave the HO sites that were eventually reconstituted by NHEJ-mediated DSB repair. The 3 kb MATa band resulting from NCO recombination events re-accumulated less efficiently in tel1Δ rad50-V1269M double mutant cells compared to both tel1Δ and rad50-V1269M single mutant cells, which generated NCO products similar to wild-type cells (Fig 5B and 5C). Interestingly, while tel1Δ rad50-V1269M double mutant cells showed decreased amount of NCOs compared to wild-type cells, the percentage of COs in the same cells was similar to that observed in wild-type cells (Fig 5B and 5C). As most of the NCO products are generated by the SDSA mechanism, this finding suggests that tel1Δ rad50-V1269M cells are specifically defective in SDSA. Because SDSA is thought to be the main mechanism responsible for mating type switching [60], we investigated the ability of tel1Δ rad50-V1269M cells to switch the mating type. HO expression was induced for 30 min by galactose addition to MATa cells and was then rapidly shut off by the addition of glucose to allow repair of the HO-induced break by gene conversion. Since there is a strong mating type-dependent preference for the choice of the two silent donor loci HMLα and HMRa [60], the MATa sequence will be replaced preferentially with the HMLα donor sequence to generate the MATα product. Strikingly, tel1Δ rad50-V1269M cells accumulated the MATα repair product less efficiently than rad50-V1269M cells, which generated this product with almost wild-type kinetics (Fig 5D and 5E). These findings indicate that tel1Δ rad50-V1269M double mutant cells are defective in mating type switching, supporting the hypothesis that they are specifically impaired in SDSA-based recombination mechanisms. Next, we investigated whether tel1Δ rad50-V1269M cells were defective in NHEJ. To this purpose, we used the strains previously used to monitor DSB repair by SSA. HO expression was induced for 30 min by galactose addition and was then rapidly shut off by the addition of glucose to allow NHEJ-mediated repair of the DSB. To ensure that repair of the HO-induced DSB occurred mainly by NHEJ, HO was induced in G1-arrested cells that were kept arrested in G1 with α-factor (Fig 6A). In fact, the low Cdk1 activity in G1 cells prevents resection of the HO-induced DSB and therefore its repair by SSA [61,62]. NHEJ-mediated DSB repair was severely affected in tel1Δ rad50-V1269M cells. In fact, the 14.5 kb uncut band resulting from NHEJ-mediated ligation of the DSB ends failed to re-accumulate in tel1Δ rad50-V1269M cells compared to wild type (Fig 6B and 6C), which also showed the expected decrease of both the 2.5 and 12 kb HO-cut band signals due to DSB repair by NHEJ (Fig 6B). Some defective re-accumulation of the 14.5 kb uncut band could be detected also in both rad50-V1269M and tel1Δ single mutant cells, although this defect was much less severe compared to that observed in tel1Δ rad50-V1269M cells (Fig 6B and 6C). To confirm the NHEJ defect, we used the GAL-HO strain, where HO induction by galactose addition generates a DSB at the MATa locus. This strain lacks the homologous donor sequences HMLα and HMRa and therefore can repair the HO-induced DSB only by NHEJ. HO expression was induced for 30 min by galactose addition to G1-arrested cells and then shut off by glucose addition to allow DSB repair by NHEJ (Fig 6D). The MATa sequence resulting from NHEJ repair events re-accumulated in wild-type cells but not in tel1Δ rad50-V1269M double mutant cells (Fig 6E), thus confirming a severe NHEJ defect. Finally, we measured NHEJ efficiency also as the ability of the above cells to re-ligate a plasmid that was linearized before being transformed into the cells [63]. Both rad50-V1269M and tel1Δ mutants showed only a slight reduction in the efficiency of plasmid re-ligation compared to wild-type cells (Fig 6F). By contrast, the re-ligation efficiency in tel1Δ rad50-V1269M cells dramatically decreased to a level similar to that found in dnl4Δ cells that lack the NHEJ enzyme responsible for DSB end ligation (Fig 6F). Thus, DSB repair by NHEJ, which is only slightly affected by the lack of Tel1 or the presence of the Rad50-V1269M mutant variant, is lost in tel1Δ rad50-V1269M double mutant cells, indicating that Tel1 supports MRV1269MX function also in NHEJ. A structural study has proposed that DNA tethering requires proper binding to DNA of the MRX globular domain to induce intercomplex hook-hook dimer formation [12]. The finding that the lack of Tel1 exacerbates the end-tethering defects of rad50-V1269M cells raises the possibility that Tel1 supports MRX function in DNA tethering by promoting proper MRX association to damaged DNA. It has been previously shown that the lack of Tel1 decreases Mre11 binding at DNA ends flanked by telomeric DNA repeats [35,36], whereas only a slight reduction of Mre11 association, if any, can be detected in tel1Δ cells at 1 kb from an HO-induced DSB [35]. However, since only a limited amount of MRX is bound at this distance from the DSB (Fig 3A and 3B) [35], we analyzed Rad50 and Mre11 association at 0.6 kb from the DSB, where Mre11 is strongly enriched (Fig 3B). The amount of Rad50 (Fig 7A) and Mre11 (Fig 7B) bound at the HO-induced DSB ends was lower in tel1Δ cells than in wild-type cells, indicating that Tel1 promotes MRX association to DSBs. Consistent with the finding that the lack of Tel1 kinase activity did not exacerbate the DNA damage sensitivity of rad50-V1269M cells (Fig 1C), the association of Rad50 and Mre11 to DSBs was not affected in tel1-kd cells (Fig 7A and 7B). The lack of Tel1 also reduced the association to the DSB of Rad50-V1269M (Fig 7C) and Mre11 in rad50-V1269M cells (Fig 7D). The decreased DSB association of Rad50, Rad50-V1269M and Mre11 in tel1Δ compared to wild type was not due to reduced protein levels, as similar amounts of Rad50 and Mre11 could be detected in protein extracts prepared from wild-type, tel1Δ, rad50-V1269M and tel1Δ rad50-V1269M cells (Fig 7E). As MRX is required to load Tel1 on the DSB ends, these findings indicate that Tel1, once loaded onto the DSB by MRX, promotes MRX association/persistence in a feedback loop and this function is necessary to support the end-tethering activity of MRX. Rif2 was shown to counteract MRX association at telomeres by inhibiting the recruitment of Tel1, which in turn promotes MRX accumulation at telomere ends [35]. We asked whether Rif2 can modulate MRX association also at intrachromosomal DSBs by investigating the effect of RIF2 deletion on MRX binding at the HO-induced DSB in tel1Δ, rad50-V1269M, and tel1Δ rad50-V1269M cells. ChIP and qPCR analysis showed that the amount of MRX (Fig 8A) and MRV1269MX (Fig 8B) bound at the HO-induced DSB ends was slightly higher in the absence than in the presence of Rif2, although similar amount of Mre11 could be detected in protein extracts prepared from wild-type, rif2Δ, and rif2Δ rad50-V1269M cells (S2 Fig). This finding indicates that Rif2 counteracts MRX and MRV1269MX association to DSBs. This Rif2 function was completely dependent on Tel1, as the amount of both MRX (Fig 8A) and MRV1269MX (Fig 8B) bound at the DSB decreased to similar levels in both tel1Δ and tel1Δ rif2Δ cells. Consistent with a previous finding that Rif2 competes with Tel1 for the binding to MRX [35], the interaction between Tel1 and MRV1269MX was strongly attenuated by Rif2 (Fig 8C). All together, these data suggest that Rif2 inhibits MRX association at DSBs by counteracting MRX-Tel1 interaction. Consistent with a direct role of Rif2 in DSB metabolism, purified Rif2 turned out to bind dsDNA in vitro (Fig 8D). Furthermore, following HO induction by galactose addition, a fully functional Myc-tagged Rif2 variant was efficiently recruited close to the HO-induced DSB and its binding increased over 3 h, spreading to 2 kb from the HO cleavage site (Fig 8E). As Rif2 physically interacts with MRX [35], we asked whether MRX may contribute to Rif2 binding at the DSB. Indeed, the amount of Rif2 bound at the DSB decreased in both rad50-V1269M and mre11Δ cells (Fig 8E), indicating that Rif2 association to the DSB is partially dependent on MRX. The finding that the lack of Rif2 increases MRX and MRV1269MX association to DSBs prompted us to ask whether RIF2 deletion could restore end-tethering in rad50-V1269M mutant cells. Indeed, RIF2 deletion suppressed the end-tethering defects of G1- and G2-arrested rad50-V1269M cells (Fig 9A and 9B) and this suppression is specific for rad50-V1269M, as rif2Δ did not restore end-tethering in G1- and G2-arrested tel1Δ cells (Fig 9A and 9B). As a consequence, the lack of Rif2 also rescued the NHEJ defects of rad50-V1269M cells, as rif2Δ rad50-V1269M cells re-ligated the BamHI-cut plasmid more efficiently than rad50-V1269M cells (Fig 9C). Interestingly, both rad50-V1269M and tel1Δ single mutant cells did not lose viability in the presence of phleomycin, MMS or low CPT doses (Fig 1B), but they exhibited hypersensitivity to high doses of CPT (Fig 9D and 9E). Consistent with a role of Rif2 in limiting MRX functions, rif2Δ also suppressed the CPT hypersensitivity of rad50-V1269M (Fig 9D), but not that of tel1Δ mutant cells (Fig 9E). The lack of Rif2 increased the association of MRX and MRV1269MX at DSBs in a Tel1-dependent manner (Fig 8A and 8B), indicating that Rif2 counteracts the ability of Tel1 to enhance MRX association to DSB. If the lack of Rif2 suppressed the CPT hypersensitivity and the end-tethering defect of rad50-V1269M cells by increasing the amount of MRV1269MX bound to the DSB, this rif2Δ-mediated suppression should require Tel1 and therefore rif2Δ should not be able to suppress the same defects in tel1Δ rad50-V1269M cells. However, we found that rif2Δ restored both end-tethering (Fig 9A and 9B) and NHEJ (Fig 9C), as well as DNA damage resistance (Fig 9F) also of tel1Δ rad50-V1269M double mutant cells. This finding indicates that the restored DNA damage resistance and end-tethering in tel1Δ rif2Δ rad50-V1269M cells do not simply depend on increased amount of MRX bound at the DSB. The finding that the lack of Rif2 restores DNA damage resistance and end-tethering in rad50-V1269M cells even in the absence of Tel1 suggests that Rif2 has other functions in regulating MRX activity besides limiting MRX recruitment to DNA ends. It has been proposed that ATP hydrolysis induces the change from a closed MRX complex, required for end-tethering, to an open configuration that promotes Mre11 nuclease activity and DSB resection [29]. Thus, we asked whether Rif2 attenuates MRX function in end-tethering by influencing its ATPase activity. The ATPase assay was performed in the presence of 200 nM of 100-bp double-stranded DNA (dsDNA) to fully activate MRX ATPase activity. Since efficiently shifting 10 nM of the same DNA requires at least 300 nM Rif2 (Fig 8D), 2 μM Rif2 was used to investigate the potential effect of Rif2-mediated DNA binding on ATP hydrolysis by MRX in reactions that contained 200 nM DNA. We found that the addition of purified Rif2 increased the ATP hydrolysis activity by both wild-type MRX (Fig 10A) and MRV1269MX complexes (Fig 10B). Since end-tethering requires the MRX ATP-bound state [29], this finding suggests that the lack of Rif2 suppresses the hypersensitivity to DNA damaging agents and the end-tethering defect of rad50-V1269M and rad50-V1269M tel1Δ cells by increasing the time spent by MRX in an ATP-bound closed conformation. As an earlier study has shown that Rif2 binds to the C-terminus of Xrs2 [35], we asked whether this Rif2-mediated regulation of MRX activity requires Rif2-MRX interaction. Interestingly, the lack of Rif2 cannot suppress the hypersensitivity to DNA damaging agents of rad50-V1269M cells carrying the xrs2-11 mutation (Fig 10D), which causes the lack of the Xrs2 C-terminal part and therefore of the MRX-Rif2 interaction. This finding suggests that regulation of MRX function by Rif2 requires its interaction with Xrs2. This requirement can be bypassed in vitro where these proteins are already in close proximity, as Rif2 can also enhance the ATPase activity of the MR complex (Fig 10C). Mutants that promote the ATP-bound closed conformation of Rad50 exhibit a higher degree of tethering activity [29], and Rif2 enhances ATPase activity not only of MRV1269MX, but also of wild-type MRX (Fig 10A and 10B). If this Rif2 function is physiologically relevant also in a wild-type context, then rif2Δ cells should show improved efficiency of DNA tethering. We found that the percentage of rif2Δ cells showing two LacI-GFP spots was reproducibly decreased compared to wild-type cells, indicating that the tethering efficiency is higher in rif2Δ cells than in wild type (Fig 9A and 9B). Furthermore, Rif2 limits DSB repair by NHEJ, as rif2Δ mutant cells re-ligated the BamHI-cut plasmid more efficiently than wild-type cells (Fig 9C). These observations, together with the finding that Rif2 enhances ATP hydrolysis by MRX, suggest that Rif2 regulates MRX function by promoting ATP-driven Rad50 conformational changes. We provide evidence that Tel1, once loaded to a DSB by the MRX complex, promotes/stabilizes MRX association to the DSB in a positive feedback loop. Tel1 exerts this function independenly of its kinase activity, suggesting that it plays a structural role in promoting/stabilizing MRX retention to DSBs. This Tel1-mediated control of MRX association can be important to ensure that MRX binding to DNA is end-specific and becomes crucial for cell viability after genotoxic treatment when MRX accumulation at DSBs is suboptimal, such as in rad50-V1269M mutant cells. The rad50-V1269M mutation impairs MRX association at DSBs and the lack of Tel1 reduces further the amount of MRV1269MX bound at DSBs. As a consequence, tel1Δ rad50-V1269M double mutant cells are much more sensitive to genotoxic agents compared to each single mutant. This DNA damage hypersensitivity is not due to defective DSB resection. Instead, tel1Δ rad50-V1269M cells are severely defective in keeping the DSB ends tethered to each other and in repairing a DSB by HR and NHEJ. Since the mainteinance of the DSB ends in close proximity is a relevant event in the repair of DSBs by both NHEJ and HR [43,44,46,52], the low degree of end-tethering in tel1Δ rad50-V1269M cells can explain the poor ability of the same cells to repair a DSB by both repair mechanisms. During HR, the second DSB end can be captured by the D-loop to form an intermediate with double Holliday junctions, whose random cleavage results in equal number of NCO or CO products. However, if the newly synthesized strand is displaced by the D-loop and anneals to the other DSB end, this results in NCO products by the SDSA mechanism [53]. Consistent with the finding that only some MRX functions are lost in tel1Δ rad50-V1269M cells, COs are generated with wild-type kinetics in tel1Δ rad50-V1269M cells, whereas formation of these repair products are impaired by the lack of any MRX subunit [64]. Interestingly, tel1Δ rad50-V1269M cells are specifically defective in the generation of NCO products, suggesting that these cells can be specifically impaired in SDSA. This observation raises the possibility that the function of MRX in keeping the DSB ends in close proximity can be particularly important to facilitate the annealing of the displaced strand to the other DSB end. By contrast, this function can be escaped when the second DSB end is already captured by the D-loop and the DNA intermediate is stabilized by the formation of a double Holliday junction. MRX association to DNA has been shown to induce parallel orientation of the Rad50 coiled-coils that favours intercomplex association needed for DNA tethering [12]. Our results support a model wherein Tel1, once loaded at DSBs by MRX, exerts positive feedback by promoting an end-specific association of MRX with DNA (Fig 10E). This Tel1-mediated regulation of DNA-MRX retention is important for proper MRX conformation needed for the tethering of broken DNA ends. Previous data have shown that the lack of Sae2 impairs end-tethering and increases MRX association to DSB ends [65,66]. These and our findings suggest that it is not the amount of MRX bound at DNA ends per se that simply dictates the integrity of end-tethering. Instead, a proper MRX-DNA interaction is required to allow the establishment of a productive MRX intercomplex association that is needed to maintain DNA strands in close proximity. The amount of MRX and Tel1 bound at telomeres is lower than that found at DSBs [36]. This difference is due to a Rif2-mediated inhibition of Tel1 accumulation at telomeric ends, which has been proposed to protect telomeric DNA ends from over-elongation and checkpoint activation [35,36,67]. This Rif2 function in modulating MRX activity is not restricted to telomeric DNA ends. In fact, although the amount of Rif2 bound at an HO-induced DSB flanked by telomeric repeats is higher than that found at HO-induced DSB containing no telomeric sequences [35], we show that the lack of Rif2 increases the association of MRX in a Tel1-dependent manner also to intrachromosomal DNA ends. As Rif2 competes in vitro with Tel1 for the binding to MRX, Rif2 can limit MRX association to DSBs by reducing MRX-Tel1 interaction. Consistent with a direct role of Rif2 at DSBs, Rif2 can bind DNA both in vitro and in vivo and its binding at DSBs is partially dependent on MRX. We also found that the lack of Rif2 suppresses the DNA damage sensitivity and the end-tethering defects of tel1Δ rad50-V1269M double mutant cells. As rif2Δ increases MRX association to DSBs only in the presence of Tel1, the finding that Tel1 is not required for rif2Δ-mediated suppression of the rad50-V1269M phenotypes suggests that Rif2 has other functions in regulating MRX activity besides limiting its association to DNA ends. Based on the characterization of Rad50 variants that either promote or destabilize the ATP-bound state, it has been proposed that the ATP-bound conformation of MRX promotes end-tethering, whereas release from this ATP-bound state by ATP hydrolysis is necessary to allow access to DNA of the Mre11 nuclease active site and subsequent DSB resection [29]. Our data show that Rif2 enhances the ATP hydrolysis activity of the MRX complex, suggesting that the lack of Rif2 might restore end-tethering and DNA damage resistance in tel1Δ rad50-V1269M cells by increasing the time spent by MRX in the ATP-bound closed conformation. Consistent with this hypothesis, rif2Δ cells show an increased efficiency of both end-tethering and NHEJ compared to wild-type cells. The finding that the lack of Rif2 suppresses the end-tethering defect and the DNA damage hypersensitivity of tel1Δ rad50-V1269M cells without increasing MRX association at DSBs suggests that the transition between closed and open MRX conformations does not necessarily result in different amounts of MRX bound at DSBs. Thus, we propose that Rif2 has a dual function in regulating MRX functions at DSBs: (i) it counteracts MRX association at DSBs by inhibiting MRX-Tel1 interaction; (ii) it enhances MRX ATPase activity, promoting the transition of the complex from a closed state, required for tethering, to an open state that unmasks Mre11 nuclease active sites and thus is competent for DSB resection (Fig 10E). Interestingly, Rif2 is known to counteract NHEJ at telomeres [68]. Whether this Rif2 function depends on a Rif2-mediated regulation of MRX conformational changes is an interesting question that remains to be addressed. Cancer therapies targeting ATM/Tel1 have been developed to increase the effectiveness of standard genotoxic treatments and/or to set up synthetic lethal approaches in cancers with DNA repair defects [69]. Interestingly, analysis of the mutational landscape in 7,494 sequenced tumors across 28 tumor types revealed that approximately 4% of all human tumors harbor mutations in the MRN/MRX complex [70]. Our finding that MRX dysfunctions can be rendered synthetically lethal with tel1Δ in the presence of genotoxic agents suggests that ATM inhibitors in combination with DNA-damaging chemotherapy could be beneficial in patients whose tumors are defective in MRN function. Strain genotypes are listed in S1 Table. Strains JKM139 and YMV45 were kindly provided by J. Haber (Brandeis University, Waltham, United States). Strains JYK40.6 and W4441-11C, used to detect end-tethering, were kindly provided by D. P. Toczyski (University of California, San Francisco, US) and M. Lisby (University of Copenhagen, Denmark), respectively. Strain tGI354, used to detect ectopic recombination, was kindly provided by J. Haber. Cells were grown in YEP medium (1% yeast extract, 2% bactopeptone) supplemented with 2% glucose (YEPD), 2% raffinose (YEPR) or 2% raffinose and 3% galactose (YEPRG). Gene disruptions were generated by one-step PCR disruption method. All the synchronization experiments have been performed at 26°C. To search for mutations that sensitize tel1Δ cells to DNA damaging agents, tel1Δ cells were mutagenized with ethyl methanesulfonate and plated on YEPD plates. Approximately 100,000 survival colonies were replica plated on YEPD plates with or without phleomycin or CPT. Phleomycin and/or CPT sensitive clones were recovered and transformed with plasmid containing wild-type TEL1 to identify those that lost DNA damage hypersensitivity. The corresponding original clones were then crossed to wild-type cells to identify by tetrad analysis the clones in which the increased sensitivity to DNA damaging agents was due to the simultaneous presence of tel1Δ and a single-gene mutation. Subsequent genetic analyses of the positive clones allowed grouping them in three allelic classes. In one class, the mutation responsible for the tel1Δ hypersensitivity to CPT and phleomycin was identified by genome sequencing and genetic analyses. To confirm that the rad50-V1269M mutation was responsible for the hypersensitivity to DNA damaging agents of tel1Δ cells, a KANMX gene was integrated downstream of the rad50-V1269M stop codon and the resulting strain was crossed to tel1Δ cells to verify by tetrad dissection that the increased sensitivity to CPT and phleomycin of tel1Δ cells co-segregated with the KANMX allele. DSB end resection at the MAT locus in JKM139 derivative strains was analyzed on alkaline agarose gels, by using a single-stranded probe complementary to the unresected DSB strand, as previously described [71]. This probe was obtained by in vitro transcription using Promega Riboprobe System-T7 and plasmid pML514 as a template. Plasmid pML514 was constructed by inserting in the pGEM7Zf EcoRI site a 900-bp fragment containing part of the MATα locus (coordinates 200870 to 201587 on chromosome III). DSB repair by NHEJ and SSA in YMV45 strains were detected by Southern blot analysis using an Asp718-SalI fragment containing part of the LEU2 gene as a probe, as previously described [71,72]. To determine the efficiency of NHEJ, the intensity of the uncut band at 30 min after HO induction (maximum efficiency of DSB formation), normalized respect to a loading control, was subtracted to the normalized values of the same band at the subsequent time points after glucose addition. The obtained values were divided by the normalized intensity of the uncut band in raffinose (100%). To determine the efficiency of DSB repair by SSA, the normalized intensity of the SSA product band at different time points after HO induction was divided by the normalized intensity of the uncut band at time zero before HO induction (100%). The loading control was obtained by hybridization of the filters with a probe that anneals to the RAD52 gene. DSB repair by NHEJ in JKM139 strains was detected by Southern blot analysis of SspI-digested genomic DNA with a MATa probe (200870–201587 coordinates of chromosome III). Strain W303 was transformed with a plasmid carrying HO under the control of a galactose inducible promoter. Mating type switching was detected by Southern blot analysis using a MATa probe (201082–201588 coordinates of chromosome III). This probe is complementary also to the HMLα locus (13826–13918 coordinates of chromosome III). To determine the efficiency of mating type switching, the intensity of the MATα band at different time points after glucose addition, normalized respect to a loading control, was divided by the normalized intensity of the uncut MATa band in raffinose (100%). DSB repair by ectopic recombination was detected by using the tGI354 strain as described in [72]. To determine the repair efficiency, the intensity of the uncut band at 2 h after HO induction (maximum efficiency of DSB formation), normalized respect to a loading control, was subtracted to the normalized values of NCO and CO bands at the subsequent time points after galactose addition. The obtained values were divided by the normalized intensity of the uncut MATa band at time zero before HO induction (100%). The centromeric pRS316 plasmid was digested with the BamHI restriction enzyme before being transformed into the cells. Parallel transformation with undigested pRS316 DNA was used to determine the transformation efficiency. Efficiency of re-ligation was determined by counting the number of colonies that were able to grow on medium selective for the plasmid marker and was normalized respect to the transformation efficiency for each sample. The re-ligation efficiency in mutant cells was compared to that of wild-type cells that was set up to 100%. ChIP analysis was performed as previously described [73]. Input and immunoprecipitated DNA were purified and analyzed by qPCR using a Biorad MiniOpticon. Data are expressed as fold enrichment at the HO-induced DSB over that at the non-cleaved ARO1 locus, after normalization of each ChIP signals to the corresponding input for each time point. Fold enrichment was then normalized to the efficiency of DSB induction. All the protein purification steps were carried out at 0–4°C. Rad50, Mre11 and Xrs2 were overexpressed in yeast and purified as described previously [18,41]. The Rad50-V1269M mutant was expressed and purified with a similar yield using the procedure devised for the wild-type protein. To assemble the MRX complex [74], Rad50, Mre11, and Xrs2 were incubated together for 5 h on ice. The resulting MRX complex was separated from unassembled proteins in a Sephacryl S-400 gel filtration column. Tel1 was overexpressed in the protease deficient yeast strain BJ5464 (MATa ura3-52 trp1 leu2Δ1 his3Δ200 pep4::HIS3 prb1Δ1.6R can1 GAL) using pGAL-FLAG-Tel1 (a kind gift from K. Sugimoto). An overnight yeast culture was diluted 1:100 into 8 L of omission medium with 2% raffinose. Cells were further cultured at 30°C until the OD660 reached 0.8, when 2% galactose was added to induce Tel1 expression. Cells were cultured for another 16 h before harvest. The pellet (~40 g), after being agitated with dry ice in a coffee grinder, was resuspended in 40 ml of ice-cold lysis buffer (40 mM KH2PO4, pH 7.4, 20% glycerol, 1 mM EDTA, 0.1% NP-40, 2 mM DTT, 600 mM KCl, and a cocktail of protease inhibitors consisting of aprotinin, chymostatin, leupeptin, and pepstatin A at 5 μg/ml each, and also 1 mM phenyl-methylsulfonyl fluoride). The lysate was clarified by centrifugation (20,000 xg, 30 min) and the supernatant was incubated with 0.5 ml of anti-FLAG-M2 agarose resin for 2 h. After washing the matrix with 30 ml of K buffer (20 mM KH2PO4, pH 7.4, 10% glycerol, 0.5 mM EDTA, 0.01% NP-40, 1 mM DTT) with 500 mM KCl, Tel1 was eluted with 1 ml of K buffer with 500 mM KCl and 200 μg/ml FLAG peptide for 1 h. The eluate containing purified Tel1 was concentrated and filter dialyzed against K buffer with 500 mM KCl in an Ultracel-30K micro-concentrator (Amicon). Purified Tel1 was stored at -80°C in small aliquots. To express (His)6-tagged Rif2, the pET-Rif2 plasmid (a kind gift from K. Sugimoto) was introduced into BL21 Rosetta cells. Early log phase culture was treated with 0.3 mM IPTG to induce Rif2 expression. After 4 h of incubation at 37°C, cells were harvested and Rif2 was purified using the following procedure. Briefly, the clarified cell lysate from 38 g pellet prepared by sonication in 100 ml T buffer (25 mM Tris-HCl, pH 7.5, 10% glycerol, 0.5 mM EDTA, 0.01% Igepal, and 1 mM DTT) containing 300 mM KCl and the protease inhibitor cocktail was mixed gently with 6 ml Ni-NTA resin (Qiagen) to capture the (His)6-tagged Rif2. After washing extensively with 100 ml T buffer containing 1M KCl and 20 mM imidazole, bound proteins were eluted with 8 ml T buffer containing 150 mM KCl and 200 mM imidazole. The eluate was applied onto a SP Sepharose column (6 ml), which was developed with a 90 ml gradient of 50–350 mM KCl in T buffer. The peak fractions containing (His)6-Rif2 were pooled and then applied onto a Mono S column (1 ml), which was developed with a 30 ml gradient of 75–450 mM KCl in T buffer. After concentrating the pooled peak fractions to 1.5 ml in an Ultracel-30K concentrator (Amicon), the preparation (0.5 ml protein) was further fractionated in a Superdex 200 column (24 ml) in T buffer with 300 mM KCl. The highly purified (His)6-Rif2 protein (1 mg) was concentrated and stored in small portions at -80°C. The ATPase assay was performed as described previously with the level of ATP hydrolysis being measured by thin layer chromatography and phosphorimaging analysis [27]. Briefly, wild-type MRX or MRV1269MX (100 nM) and Rif2 (2 μM) were used in the presence of 100-bp dsDNA (200 nM). The effect of Rif2 on the ATPase activity of the MR complex (100 nM) was examined as described above. The DNA binding assay for Rad50 and MRX was slightly modified from that published previously [27]. Briefly, wild-type Rad50 and Rad50-V1269M (50, 100, and 200 nM), wild-type MRX and MRV1269MX (10, 20, and 40 nM) were incubated with radiolabeled 70-bp dsDNA substrate (10 nM) in the reaction buffer (25 mM Tris-HCl, pH 7.5, 150 mM KCl, 2 mM MgCl2, 1 mM DTT, 100 μg/ml BSA, 2 mM ATP) at 30°C for 10 min. The reaction mixtures were resolved in a 0.3% agarose gel in SB buffer (10 mM NaOH, 40 mM boric acid, pH 8.0) on ice. The gel was dried and then subject to phosphorimaging analysis. To examine Rif2 for DNA binding activity, the indicated amount of purified Rif2 was incubated with radiolabeled 100-bp dsDNA substrate (10 nM) in the reaction buffer at 30°C for 10 min, followed by the same analytical procedure described above. Flag-tagged Tel1 was incubated with MRX or MRV1269MX in 30 μl T buffer containing 100 mM KCl for 2 h on ice. The reaction was mixed gently with anti-FLAG-M2 agarose resin (10 μl) for 2 h to capture Flag-tagged Tel1 and associated MRX or MRV1269MX. After washing the resin three times with 200 μl T buffer, bound proteins were eluted with SDS-PAGE loading buffer and then subject to western blot analysis. To determine the effect of Rif2 on the interaction between Tel1 and MRV1269MX, Flag-tagged Tel1 was incubated with MRV1269MX in the absence or presence of Rif2, and then subjected to affinity pull-down as described above.

10.1371/journal.pntd.0000202

Microsatellites Reveal a High Population Structure in Triatoma infestans from Chuquisaca, Bolivia

For Chagas disease, the most serious infectious disease in the Americas, effective disease control depends on elimination of vectors through spraying with insecticides. Molecular genetic research can help vector control programs by identifying and characterizing vector populations and then developing effective intervention strategies. The population genetic structure of Triatoma infestans (Hemiptera: Reduviidae), the main vector of Chagas disease in Bolivia, was investigated using a hierarchical sampling strategy. A total of 230 adults and nymphs from 23 localities throughout the department of Chuquisaca in Southern Bolivia were analyzed at ten microsatellite loci. Population structure, estimated using analysis of molecular variance (AMOVA) to estimate FST (infinite alleles model) and RST (stepwise mutation model), was significant between western and eastern regions within Chuquisaca and between insects collected in domestic and peri-domestic habitats. Genetic differentiation at three different hierarchical geographic levels was significant, even in the case of adjacent households within a single locality (RST = 0.14, FST = 0.07). On the largest geographic scale, among five communities up to 100 km apart, RST = 0.12 and FST = 0.06. Cluster analysis combined with assignment tests identified five clusters within the five communities. Some houses are colonized by insects from several genetic clusters after spraying, whereas other households are colonized predominately by insects from a single cluster. Significant population structure, measured by both RST and FST, supports the hypothesis of poor dispersal ability and/or reduced migration of T. infestans. The high degree of genetic structure at small geographic scales, inferences from cluster analysis and assignment tests, and demographic data suggest reinfesting vectors are coming from nearby and from recrudescence (hatching of eggs that were laid before insecticide spraying). Suggestions for using these results in vector control strategies are made.

Chagas disease is a protozoan infection caused by the parasite Trypanosoma cruzi. Chagas is prevalent throughout Central and South America, and it remains a chief concern in Bolivia. A movement that began in 1991 called the Southern Cone Initiative has been successful in reducing the incidence of Chagas disease in the Southern Cone countries of Argentina, Brazil, Chile, and Uruguay; but due to socio-economic and other factors, incidence remains high in Bolivia. The most important mode of transmission of T. cruzi to humans and other mammals is through feces of triatomine bugs. Thus, disease control and transmission prevention focus on elimination of triatomine vectors, and more specifically in Bolivia, it focuses on the elimination of Triatoma infestans. This study focuses on T. infestans in the Department of Chuquisaca, Bolivia. Ten highly variable microsatellite markers were used to analyze the population structure of insects collected in different towns. Statistical analyses show that T. infestans are highly structured, which means that they colonize on a small geographic scale. The results also suggest little active dispersal. These findings should be implemented during control efforts so that insecticide spraying focuses on geographic areas of colonization and re-colonization.

Chagas disease is a parasitic disease in which the pathogenic agent, Trypanosoma cruzi is transmitted by hematophagous insects of the sub-family Triatominae. Triatoma infestans is the major vector in the Andean highlands where the disease is endemic and has infected humans for over 9000 years [1]. Chagas disease is the most important parasitic disease in the Americas in terms of mortality and economic impact [2]. In Bolivia the endemic area covers 55% of the country and, in 1985, more than one million people were infected [3]. In 1991 a public health program, the Southern Cone Initiative was launched by the World Health Organization to eliminate vector populations [4], through spraying of houses and surrounding areas with pyrethroid insecticides [5]. In Argentina, Brazil, Chile, and Uruguay, T. infestans is exclusively domestic or peri-domestic, thus eradication of the vector in these regions, followed by vigilance against re-infestation, has proven largely successful in reducing transmission of T. cruzi and thus the prevalence of Chagas disease [6]. In contrast, in Bolivia the vectors occur in domestic, peri-domestic, and sylvatic environments [7]; thus, control of T. infestans in towns and homesteads is confounded by the possible re-infestation from surrounding sylvatic areas. Molecular genetic research can help vector control programs by identifying and characterizing genetically distinct vector populations and then developing effective intervention strategies [8]. Several genetic markers including isozymes and the mitochondrial cytochrome b gene have proved useful in studying the genetic diversity of T. infestans [9],[10]; however, markers with more resolution would aid vector control efforts. DNA based microsatellite markers have been widely used in population studies because of their large polymorphism information content, widespread distribution in the eukaryotic genome and robust methodology. To reduce transmission of Chagas disease, estimates of population differentiation are crucial to understand vector dispersal, sources of reinfestation, and gene flow; this genetic information is an important tool for effective management of insect control programs. Here we aimed to investigate the population genetic structure and inferred the source of colonization of vectors in the department of Chuquisaca, Bolivia using ten highly polymorphic microsatellite markers. The geographic region has high levels of human infection and house infestation and is located in a region thought to be the evolutionary origin of T. infestans. Insects were collected from 23 localities including both peri-urban (inhabited areas in the immediate vicinity of a city) and rural sites (less than 2000 inhabitants) in the provinces of Oropeza, Zudañez, Azurduy, Yamparaez, Tomina, Belisario Boeto and Hernando Siles within the Department of Chuquisaca, in the Bolivian highlands ranging from 1079 to 3020 meters above the sea level (Table 1, Figure 1). This area presents a broken topography with numerous valleys and small plateaus characterized by very diverse climates. In the Andean highlands, wheat is grown predominantly in small-scale, subsistence farming systems. In higher precipitation areas, potato is the preferred crop. Rainfall in these areas ranges from approximately 300 to 600 mm per year. In the Andean Plateau the average temperature is less than 10°C and there is less than 500 mm of annual precipitation. The Andean valleys present moderate climates, with average temperatures of 18°C and approximately 500 and 600 mm of rain every year. The relative humidity varies throughout the year, showing a similar pattern to the other climatic parameters. The majority of the vegetation in the plateau is grassy plain with a rich variety of grasses and dichotomous herbs, but also shrubs and some trees. The valleys contain fertile soils where vegetables, cereals and fruits are grown. Specimens of T. infestans included in the present study were a mixture of nymphs and adults, collected from inside as well as the immediate vicinity of homes. Collections were made in the months of the Southern hemisphere summer 2002, spring 2005 and fall 2005. Forty-four insects came from a single corral in the community of Jackota in the province of Zudañez, 78 insects were collected in the community of Zurima in the province of Oropeza, and 37 were collected in Sucre the capital and main city of Chuquisaca located in the province of Oropeza. The remaining 71 insects came from collections in 20 localities throughout Chuquisaca. All insects included in the study were identified as T. infestans using taxonomic keys [11]. Insects from the first collection were frozen live. Those from subsequent collections were placed in 95% ethanol while alive. Specimens then were sent to Vermont, USA for molecular analysis. DNA was extracted from three legs or 25 mg of tissue obtained from the posterior part of the abdomen of a given specimen using the Qiagen DNeasy DNA extraction kit (Qiagen, Inc., Valencia, CA). Care was taken to avoid sampling from the mid-abdomen as the stomach may inhibit the PCR reaction [12]. We investigated population genetic structure at both ecological and geographic levels (Table 1 a–e). Ecological grouping included: Eastern, low altitude (97 individuals) vs. Western, high altitude (133 individuals) regions (Table 1 a) and domestic (36 individuals) vs. peri-domestic habitats (42 individuals) within Zurima (Table 1 e). The geographic groupings included: among 5 communities within a 100 Km diameter with a total of 193 individuals (Table 1 b), among 7 households within a 750 m diameter (defined as a house and the associated peri-domestic buildings and corrals, with 4, 7, 14, 7, 6, 11 and 3 insects respectively) within Zurima (Table 1 c), and 36 nymphs from a single corral in Jackota (Table 1 d). Four insects from a household in Zurima were collected in 2002 before spraying, all other specimens were sampled in 2005, up to 6 months after spraying and were re-infesting insects. There was significant genetic differentiation among populations based on RST and FST estimates for all hierarchical levels analyzed (Table 2). Between low altitude East and high altitude West, RST and FST are statistically significant (RST = 0.08, FST = 0.02); both measures are also significant among the five communities <100 Km apart (RST = 0.12, FST = 0.06) and among houses in Zurima (RST = 0.14, FST = 0.07). We also observed significant differentiation between domestic and peri-domestic populations within the community of Zurima (RST = 0.05, FST = 0.03). Although East and West were genetically differentiated, we did not observe a trend towards higher diversity at higher altitude when we compared the Western populations with a mean altitude of 2600 m, which comprises the provinces of Oropeza and Yamparaez, with the Eastern populations having a mean altitude of 2300 m which includes the provinces of Zudañez, Belisario Boeto, Azurduy, Tomina and Hernando Siles. The mean number of alleles per locus was 15.3±2.23 and 13.6±2.31 at the high and low altitudes respectively (t-test, P>0.05). The dendogram based on Nei's genetic distances showed a cluster comprising populations from Zurima, El Chaco and Sucre differentiated from a sister cluster with the Jackota population (Figure 2). These two clusters were well differentiated from a cluster containing populations from the more distant Serrano (Table 3). Pairwise estimates of RST and FST among communities (Table 4) support the conclusion that El Chaco, Zurima and Sucre are genetically similar to each other and that these communities differ from Jackota and Serrano. Within the town of Zurima, the estimates of RST and FST among the 7 households are shown in Table 5. With respect to RST, households 4 and 5 are the most different from other households. These households represent peri-domestic samples and their difference from the other households is also shown by the significant difference among habitats (Table 2 e). Five clusters were identified among the 5 communities (Table 3). When assigning individuals to genetic populations based on these communities, 78–86% of the individuals were assigned. The clusters represent insects with similar genotypes. Assignment tests can be viewed in terms of the number and evenness of communities in a single cluster and with respect to the number and evenness of clusters represented in a single community. Cluster 1 was a mixture of insects from the three close localities, Sucre, El Chaco and Zurima. The other four clusters contained insects from primarily one locality: clusters 2 and 3 were primarily from Zurima (24/29 = 83% and 24/28 = 86% respectively); cluster 4 from Jackota (32/33 = 97%) and cluster 5 from Serrano (18/28 = 64%). About 15–20% of the insects from each community were not assigned. From the community perspective, most of the insects from four of the communities are from a single genetic group: Jackota (73% from cluster 4), Sucre (67% from cluster 1), El Chaco (56% from cluster 1) and Serrano (72% from cluster 5). Zurima contains a mixture of groups, 13% group 1, 31% from group 2 and another 31% from group 3. At the household level, five genetic clusters were identified from the seven households (Table 6). Insects from households 1, 2, 5 and half of those from household 7 were collected in peri-domestic settings, all the others came from domestic structures. The assignment test was quite successful for some households (100% assigned), yet for other households none of the insects were assigned. There does not seem to be any tendency for insects collected from domestic vs. peri-domestic sites to be assigned. With respect to the life stage and household of origin for the insects in each cluster, clusters 3 and 5 were mostly from a single household (86% and 100% respectively) with cluster 5 being composed only of the most geographically isolated insects and cluster 3 containing 5 nymphs and one adult from household 3 along with one adult male from household 6. Cluster 2 contains insects from 5 of the 7 households and cluster 1 contains insects coming from 4 households. Cluster 4 contains only nymphs, five from household 3 and four from household 2. The fifth cluster was a mix of adults and nymphs coming exclusively from Z-6. All four insects from the pre-spraying collection were not assigned to any cluster (Z-1) (Table 6). Relatedness of insects in nine out of seventeen houses was not significantly different from 0 (Table 7). From these nine households, in six cases at least one adult was collected and in three cases only nymphs were collected. For one household (S-1), r<0 (P<0.05) indicating significant outcrossing. For seven houses r>0 (P<0.05). A value of r≈0.25 (half sibs) was obtained for four households, and although the relatedness was similar in these households, the composition of the insect collection varied (1 site only adults 1 site only nymphs and 2 sites a mix of adults and nymphs). For the sites with the highest relatedness values (r≈0.33, 0.44 and 0.48), in 2 houses a single adult and 2–4 nymphs were collected and for one household only nymphs were collected. The estimates of the effective number of migrants per generation, Nm, among towns <40 Km apart was higher (2.03) compared with those among more distant communities (1.42) and among houses within the town of Zurima (0.99). The Mantel test of isolation by distance revealed a non-significant correlation between Slatkin's linearized FST and Nm vs. the natural log of geographic distance (R2 = 0.001, P = 0.294; R2 = −0.184, P = 0.725 respectively). Non-significant results were also observed when applying the Mantel test for a correlation between Nei's genetic distances and geographic distances among populations (R2 = 0.00056, P = 0.135), and altitude (R2 = −0.000012, P = 0.548). The Mantel tests had low power because of the small samples within many of the communities. Our study region is an ecologically diverse but geographically small valley–mountain environment in the department of Chuquisaca in Southern Bolivia. This region has high levels of house infestation and vector and human T. cruzi infection [22]–[24]. The use of microsatellite loci, now routine in many insect population genetic studies because they are inherently more polymorphic than allozyme loci and generally not targets of selection, allows us to detect population structure with more statistical power [25]. Previous studies on population genetics and morphometry of T. infestans from Bolivia have found geographical variation in patterns of population structure in this vector; therefore we examined distinct ecological and geographic hierarchical groups ranging from a single goat corral to comparing western and eastern regions of Chuquisaca. Genetic analysis over twenty-three localities throughout the department of Chuquisaca have revealed moderate but highly significant levels of genetic variation among populations. Both FST and RST showed differentiation even within a community. Previous study in the same area using a mitochondrial cyt b gene [10] failed to verify significant genetic diversity comparing distant rural and peri-urban settings. However, significant differentiation was revealed when populations from Chuquisaca (Andean) were compared with non-Andean populations from Brazil, Argentina and the Bolivian Chaco. Cytogenetic [26] and allozyme [9] studies have also confirmed genetic differences between T. infestans from highlands (>1800 m) and lowlands (<500 m). We examined insects from eastern and western Chuquisaca that significantly differ in altitude, both groups are >2000 m, and we detected significant differentiation at this ecological level. In our study, RST values were larger than FST, suggesting polymorphism is high and rates of migration are low [27]. The IAM-based estimates (FST) indicate lower differentiation because they do not distinguish among shared alleles in different populations that are not identical by descent. Similar values of RST and FST are only to be expected when mutation rates are negligible in comparison to migration and drift. When the SMM contributes to population differentiation, RST values should be larger than FST values [28]. When comparing the 5 communities (Table 4), in general, pairwise RST>FST suggesting that mutation contributes to differences at this geographic level. However, there is no such pattern for pairwaise RST and FST among households suggesting that mutation does not contributes much to differentiation at this level. As suggested by RST>FST, T. infestans has a low capacity for active dispersal [29] but can passively disperse over long distances when associated with human migration. It seems that this has been the structuring pattern of T. infestans in Chuquisaca. In our study, the results of the assignment of individuals to genetic clusters (Table 3) shows the assignment of insects to genetic populations located >100 Km apart. Several studies using isozymes have examined population structure in T. infestans and report variation among regions in the spatial scale of population differentiation. Variation in population structure among regions was established using twelve isozymes [9],[30]. There was significant differentiation of T. infestans populations between villages located 50 Km apart in Vallegrande, Santa Cruz yet in the Yungas of La Paz, populations only a few Km apart showed significant differences. Using 19 isozyme loci, significant differences in allele frequencies between populations separated by 20 Km were found in central Bolivia [31], but this study failed to detect differentiation between sylvatic and domestic populations of T. infestans. By contrast, incipient differentiation between sylvatic and domestic populations was revealed using morphometry of the head capsule [32]. Other studies [33] have indicated that the panmictic unit may be no larger than a single household, based on the finding of significant differentiation within households in Yungas, Bolivia. Differences have also been detected between geographically close populations based on wing geometric morphometry [34]. The results of our study show significant population structure among communities. These results are supported by cluster analysis, which identified the geographically isolated communities as separate clusters (Jackota and Serrano, Table 3); however the closer communities are not as genetically distinct (Sucre, El Chaco and Zurima, Table 3). If migration depends on habitat quality, when insects find favorable conditions at the microhabitat level it can reduce their dispersal tendency and consequently reduce gene flow. Within the community of Zurima we sampled 7 houses and statistical analysis estimated 5 clusters within an area of 750 m diameter. These results suggest the single household is not the panmictic unit in this area of Chuquisaca and is in accordance with a study on dispersal capacity in the towns of Trinidad and Mercedes, Argentina, that clustered the source of re-infestation at ∼500 meters [35]. The isolation-by-distance tests based on allozyme markers in populations from several areas in Bolivia and Peru found a positive correlation between genetic and geographic distances [9]. We found no evidence of isolation by distance within this area of Chuquisaca. Differences between the two studies may result because our study had low statistical power due to sampling a relatively small number of communities, few samples per community and microsatellite data, because of the high number of alleles, require large sample sizes. However, the non-significant results may also be because our study covers a small geographic area of Chuquisaca characterized by a high human migration rate in the last 40 years [36]. Previous studies [37] identified unique local characteristics in landscape and vegetation, distances between houses, the abundance of bugs and hosts, and presence of many peri-domiciliary structures in conjunction with the existence of sylvatic populations as contributing to spatial patterns of re-infestation. Identification of the source of re-colonizers can direct control programs in the surveillance phase. We have found significant differentiation at the household level in populations from Chuquisaca, Bolivia. Cluster analysis, relatedness estimates and life stage data can be combined to understand pre-spraying population dynamics and infer patterns of re-colonization. Within Zurima, individuals collected in the most geographically isolated household (Z-6) were assigned to one cluster. The relatedness of insects in Z-6 was significantly greater than 0 (Z-6, r>0.17, c.i. = 0.15, Table 7). Eight of the nine adults and the two nymphs in Z-6 were assigned to a single cluster, but this house also had insects from two other clusters. The reinfestation patterns for individual houses are quite variable including repeated colonization from several sources (Z-2, seven peri-domestic adults, r≈0.10, c.i. 0.13, Table 7), a single multiply mated female (S-3, 1 adult 5 nymphs, r≈0.26, c.i. = 0.21, Table 7), multiple colonization from a single source (Z-5, 3 males and 3 females, r≈0.23, c.i. = 0.21, Table 7), recrudescence of full sibs (Z-10, 3 nymphs, r≈0.48, c.i. = 0.45, Table 7) and recrudescence of unrelated eggs (Z-3, 14 insects mostly nymphs, r≈0.05, c.i. = 0.07, Table 7). Of course there are multiple possibilities for each household and these inferences are to show the range of possibilities, not to infer a given scenario for a specific household. The presence of adults in many households less than 6 months after spraying suggests that for many cases, structures around human habitations may be playing a key role as the source of insects invading houses. The presence of nymphs in houses where no adults were found suggests recrudescence. Hence, recrudescence from a residual population and colonists from peri-domicile structures, rather than reinvasion from surrounding localities, seems to be a probable explanation of the source of re-colonists found during surveillance activities in this area. The variety of results suggest that continuous surveillance consisting of analyzing relatedness among reinfesting insects at the household level is critical to maintain insect free houses and optimize insecticide spraying in this region.

10.1371/journal.pcbi.1005792

The structured ‘low temperature’ phase of the retinal population code

Recent advances in experimental techniques have allowed the simultaneous recordings of populations of hundreds of neurons, fostering a debate about the nature of the collective structure of population neural activity. Much of this debate has focused on the empirical findings of a phase transition in the parameter space of maximum entropy models describing the measured neural probability distributions, interpreting this phase transition to indicate a critical tuning of the neural code. Here, we instead focus on the possibility that this is a first-order phase transition which provides evidence that the real neural population is in a ‘structured’, collective state. We show that this collective state is robust to changes in stimulus ensemble and adaptive state. We find that the pattern of pairwise correlations between neurons has a strength that is well within the strongly correlated regime and does not require fine tuning, suggesting that this state is generic for populations of 100+ neurons. We find a clear correspondence between the emergence of a phase transition, and the emergence of attractor-like structure in the inferred energy landscape. A collective state in the neural population, in which neural activity patterns naturally form clusters, provides a consistent interpretation for our results.

Neurons encoding the natural world are correlated in their activities. The structure of this correlation fundamentally changes the population code, and these effects increase in larger neural populations. We experimentally recorded from populations of 100+ retinal ganglion cells and probed the structure of their joint probability distribution with a series of analytical tools inspired by statistical physics. We found a robust ‘collective state’ in the neural population that resembles the low temperature state of a disordered magnet. This state generically emerges at sufficient correlation strength, where the energy landscape develops an attractor-like structure that naturally clusters neural activity.

The past decade has witnessed a rapid development of new techniques for recording simultaneously from large populations of neurons [1–4]. As the experimentally accessible populations increase in size, a natural question arises: how can we model and understand the activity of large populations of neurons? In statistical physics, the interactions of large numbers of particles create new, statistical, laws that are qualitatively different from the original mechanical laws describing individual particle interactions. Studies of these statistical laws have allowed the prediction of macroscopic properties of a physical system from knowledge of the microscopic properties of its individual particles [5]. By exploiting analogies to statistical physics, one might hope to arrive at new insights about the collective properties of neural populations that are also qualitatively different from our understanding of single neurons. The correlated nature of retinal ganglion cell spike trains can profoundly influence the structure of the neural code. Information can be either reduced or enhanced by correlations depending on the nature of the distribution of firing rates [6], the tuning properties of individual neurons [7], stimulus correlations and neuronal reliability [8, 9], the patterns of correlations [10], and interaction among all these factors [11]. In addition, the structure of the decoding rule needed to read out the information represented by a neural population can be strongly influenced by the pattern of correlation regardless of whether it reduces or enhances the total information [12, 13]. One approach to understanding the properties of measured neural activity is to study the nature of minimally structured (‘maximum entropy’) models of the probability distribution that reproduce the measured correlational structure [14–16]. These models have been shown to be highly accurate in reproducing the full statistics of the activity patterns of small numbers of neurons [15, 17, 18]. The hope is that even if these models underestimate the real structure of larger neural populations, the properties of the distribution which arise in these simplified models are general and of consequence to the true distribution. Maximum entropy models that constrain only the pairwise correlations between neurons are generalized versions of the Ising model, one of the simplest models in statistical physics where collective effects can become significant. The macroscopic behavior of these models varies substantially depending on the parameter regime. By fitting these models to measured neural acitivity, we can begin to explore (by analogy) the ‘macroscopic’ properties of the retinal population code. In particular, we can gain insight into these macroscopic properties by introducing a fictitious temperature variable into the maximum entropy model. By changing this temperature, we can continuously tune the population from a ‘high temperature’ regime where correlation has a minimal effect on the probability distribution to a ‘low temperature’ regime where correlation dominates. Furthermore, we expect to see signatures of a phase transition at the boundary between these regimes. Thus, by observing a phase transition at a particular value of the temperature variable, we can determine if the real state of the neural population more closely resembles the high or low temperature state. One macroscopic property of interest is the specific heat. Discontinuity or divergence of this property is an indicator of a second-order phase transition, which implies a qualitative change in the properties of the system. Previous studies have shown that the specific heat has a peak that grows and sharpens as more neurons are simultaneously analyzed [19–22]. Most of the literature on this topic can be divided into two camps: the ‘proponents’ who argue that this is a signature of criticality, i.e. the system is poised in between high and low temperature phases, in a manner that might optimize the capacity of the neural code [19, 21, 23], and the ‘sceptics’ who argue that this is merely a consequence of ignored latent variables [24, 25], ignored higher order correlation structure in the data [26], or even the presence of any correlation at all [27]. An alternative interpretation is that system is in a ‘low temperature’ state, and that the peak in the heat capacity is a signature of a first-order phase transition. The difference between the two types of transitions rests on which macroscopic properties of the system are discontinuous at the transition: first-order phase transitions have discontinuities in the entropy (hence having an infinite heat capacity), while second-order phase transitions will have discontinuities, or integrable divergences, in the specific heat. The observed sharpening of the specific heat is influenced by finite-size effects which could be consistent with either a delta function (first-order) or a divergence (second-order) in the specific heat. In principle, one can use finite-size scaling arguments to argue that the sharpening is more or less consistent with one of the two possibilities. In practice however we do not think that we can convincingly distinguish between these two possibilities with our analysis of the specific heat, and so we cite other forms of evidence in favor of our interpretation. Empirically, the peak of the specific heat is found at a higher temperature than the operating point of the real system (T = 1), suggesting that the system is on the low temperature side of the phase transition. Low temperature phases in statistical physics are usually associated with structure in the distribution of states, in which the system can ‘freeze’ into ordered states. High temperature phases, in contrast, are associated with weakly correlated, nearly independent structure in the population of neurons. From this perspective, the phase transition that many studies have observed as a function of an imposed temperature variable, T, serves as an indicator of structure in the probability landscape of the neural population at the real operating point (T = 1). Maximum entropy models fit particular statistics of the distributions of experimentally measured neural neural activity [16]. Because retinal responses are specific to both the adaptational state of the retina and the ensemble of stimuli chosen to probe them, the measured pattern of neural correlation—and hence the detailed properties of the maximum entropy model—will also vary. Therefore, it is yet unclear how robust is the presence of a low temperature state to different experimental conditions. The phase transition itself arises as a consequence of the correlations between neurons. This pattern of correlation in turn has contributions from correlations in the stimulus and from retinal processing. The distribution of correlations also has a particular shape, with many weak but statistically non-zero terms. It is unclear how these different properties contribute to the nature of the structured collective state of the neural population. Here, we show that while the detailed statistics of the retinal population code differ across experimental conditions, the observed phase transition persists. We find that retinal processing provides substantial contributions to the pattern of correlations among ganglion cells and thus to the specific heat, as do the many weak but statistically non-zero correlations in the neural population. We also find that the spatio-temporal processing of the classical receptive field is not sufficient to understand the collective properties of ganglion cell populations. To address the nature of the collective state of the retinal population code, we explored how a particle representing the state of neural activity moves over the system’s energy landscape under the influence of finite temperature. We find that the energy landscape has regions that “trap” particle motion, in analogy to basins of attraction in a dynamical system. By varying the overall correlation strength, we show that the emergence of this structure is closely connected to the emergence of the measured phase transition. This emergence occurs at surprisingly low overall correlation strength, indicating that the real population is robustly within the structured regime. One of the main goals of our study is to test whether the collective state of a neural population is robust to different experimental conditions. Adaptation to the statistics of the visual input is a central feature of retinal processing, and any robust property of the retinal population code should be present in different adaptational states. We focused first on adaptation, and in particular we chose an experiment probing adaptation to ambient illumination. To process the visual world, the retina has to adapt to daily variations of the ambient light level on the order of a factor of a hundred billion [28]. Prominent examples of known sites in the retinal circuit with adaptive mechanisms include the voltage-intensity curves of photoreceptors [29–31], the nonlinear output of bipolar cells [32], and the surround structure of ganglion cells [33]. A significant contribution to these effects arises from the light-dependent global dopamine signal [34, 35], which regulates retinomotor changes in the shapes of photoreceptors [36], and gap junction transmission in horizontal [37] and AII amacrine cells [35]. The global nature of the dopamine signal suggests that any cells or synapses in the retina that possess dopaminergic receptors (almost all retinal cell types studied, [38]), will experience adaptive effects of changes in the mean light level. Though the literature on single cell adaptation to ambient illumination is extensive, little is known about the changes in correlational structure across the population of retinal ganglion cells. We recorded from a population of tiger salamander (Ambystoma Tigrinum) retinal ganglion cells responding to the same natural movie at two different ambient light levels. In Experiment #1, we recorded with and then without an absorptive neutral density filter of optical density 3 (which attenuates the intensity of light by a factor of 103) in the light path of the stimulus projecting onto the retina. The stimuli consisted of a chromatic checkerboard and a repeated natural movie (details in Methods). Thus, the contrast and statistics of each visual stimulus were the same under both luminance conditions, with the only difference being that in the dark condition, the mean light level was 1000 times lower than in the light condition. The responses of individual cells to the checkerboard allowed us to measure the spatiotemporal receptive field of each ganglion cell using reverse correlation [39, 40] in each light condition. For most cells these three linear filters were scaled versions of each other, suggesting a single filter with different sensitivities to the red, green and blue monitor guns (Fig 1A). The vast majority of rods in the tiger salamander retina are classified as ‘red rods’ (98%); similarly, most of the cones are ‘green cones’ (80%) [41, 42]. We estimated the relative sensitivities of these two photopigments to our monitor guns from the reported spectra of these photopigments [42–44] and a measurement of the spectral output of the monitor guns (see Supplement). We found that for many ganglion cells the relative amplitudes of these three sensitivities were closely consistent with the estimated sensitivity of the red rod photopigment in the dark recording, and the green cone photopigment in the light recording (Fig 1B and 1C). These results suggested to us that in our experiments, retinal circuitry was in the scoptopic, rod-dominated limit in our dark condition and in the photopic, cone-dominated limit in our light condition. How much does the feature selectivity of ganglion cells change across the two light-adapted conditions? We can see from the temporal kernel of the spike-triggered average that there are some changes in temporal processing along with a big change in response latency (Fig 1A). Another way to compare feature selectivity is to look at what times a ganglion cell fires spikes to repeated presentations of the same natural movie (Fig 1D). The spike timing of individual ganglion cells was reproducible across repeats of a natural movie within a particular luminance condition. However, across conditions there were significant changes in which times during the stimulus elicited a spike from the same cell (Fig 1D). To quantify this effect over our entire recording, we estimated the shuffled autocorrelation function of each cell’s spike train (i.e. the correlation function between spikes on one trial and those on another trial [45]). The width of this shuffled autocorrelation function is one measure of timing precision for the spike train [45]. The narrower width of the autocorrelation curve in the light adapted condition indicated greater spike timing precision in that condition. The 200ms offset in the peak of the cross correlation curve was characteristic of the longer delays in signal processing that arise in the rod versus cone circuitry, which can also be seen in the different latencies of the reverse correlation (Fig 1A). The area under the curve, but above the random level set by the firing rate, is a measure of reproducibility across trials. Normalizing this measure across luminance conditions yielded an estimate of spiking reproducibility across light conditions. A value of unity for this measure indicates that spiking reproducibility across light conditions is as high as the reproducibilities within each condition. We found a wide range of reproducibility across our populations of neurons (0 to 0.75, Fig 1E), suggesting that significant changes in feature selectivity occured for most neurons. Our results are qualitatively consistent with a recently published study [46], where retinal ganglion cells gain or lose specific firing events at different ambient light levels. Given that our observed changes in chromatic sensitivites and feature selectivities of ganglion cells were consistent with distinct adaptational states for the retina, it seemed likely that we would find statistically significant changes in the average firing rates, pairwise correlations, and the distribution of simultaneously active cells. Because the details of the maximum entropy model are a function of this set of constraints, statistically significant changes in these moments across light conditions were particularly important. Otherwise, any observed ‘robustness’ at the collective level would be trivial. We found that firing rates mostly increased at the higher light level, an effect that was highly statistically significant (Fig 2A). The correlation coefficients between pairs of cells showed some increases and some decreases between the light and dark adaptational states (Fig 2B). Overall, the distribution of correlation coefficients was roughly the same. However, the detailed pattern of correlation appeared to change. To evaluate the significance of changes in correlation coefficients, we estimated the difference between the correlation coefficients in the two light-adapted conditions normalized by the error bar—a measure also known as the z-score. The distribution of z-scores across all pairs of ganglion cells (Fig 2D, black curve) had significant density in the range of 5 − 10 standard deviations, a result that is not consistent with controls (random halves of the dark dataset compared again each other, gray curve), or with the curve expected for the null hypothesis (a Gaussian with standard deviation of one, red curve). Thus correlation coefficients between individual pairs of cells change far more than expected by chance. The final ingredient in the maximum entropy model, the probability of K simultaneously active cells, (P(K)), displayed a statistically significant overall shift towards sparseness in the dark adapted condition (Fig 2C). Following previously published results [16, 19], we modeled the distributions of neural activity with the k-pairwise maximum entropy model, which approximates the probability of all patterns of activity in the ganglion cell population. Here, we binned each ganglion cell’s spike train in 20 ms time windows, assigning 0 for no spikes and 1 for one or more spikes, ri = [0, 1]. We denote a particular population activity pattern as R = {ri}. The probability of state R in this model is given by: P ( R ) = 1 Z exp ( - E k - pairwise ( R ) ) (1) where Z is the normalizing factor, and we’ve introduced a unitless ‘energy’: E k - pairwise ( R ) = ∑ i N h i r i + ∑ i , j ≠ i N J i j r i r j + ∑ k = 1 K δ ( ∑ i r i , k ) λ k (2) with δ the Kronecker delta. The shapes of the probability and energy landscapes have a one-to-one relationship to each other, with energy minima corresponding to probability maxima. Because of the extensive intuition surrounding the concept of an energy landscape in physics, we will often use this term. This model is constrained to match the expectation values 〈ri〉, 〈rirj〉 and P(K) measured in the data. We inferred these models with a modified form of sequential coordinate descent [16, 47, 48] [see also Methods, S2, S3, S4 and S5 Figs]. A first step towards the study of collective phenomena in neural populations is to understand what is the qualitative nature or ‘phase’ of the neural population. Phases of matter occur everywhere in nature where there is some collective structure in the population. In the theory describing phase transitions in statistical physics, first-order phase transitions can occur when a particular system can decrease its free energy by transitioning to a new phase. While such a transition in our work here will occur in a region of parameter space that is not real—i.e., is not visited by the retina experimentally—its occurence provides evidence for structure in the real experimentally measured distribution. Thus, an explanation of previously observed phase transitions is that the pattern of correlation among ganglion cells induces a highly structured phase which is qualitatively different from the phase found in the high temperature limit. From this perspective then, we ask whether this phase is robust to different adaptational states, and what are the properties of the retinal population code that give rise to that phase? To study the emergence of a phase transition with increasing system size, we subsampled groups of N neurons and inferred models for these subsets of the full neural data. For all of these networks, we then introduced a fictitious temperature parameter, T, into the distribution, P(R) = (1/Z(T))exp(−E(R)/T). This parameter allows us to visit parameter regimes of our model where the qualitative nature of the system changes. If the shape of the specific heat as a function of temperature exhibits a sharp peak, this indicates a phase transition—a macroscopic restructuring of the properties of the system across parameter regimes. Thus, an analysis where we vary the effective temperature allows us to gain insight into the state of the real neural population at T = 1. As previously described [19], we found a peak in the specific heat that sharpened and moved closer to T = 1 with increasing system size, N (Fig 3A–3C). The systematic changes that we observe as a function of the system size N (Fig 3A–3C) indicate that correlation plays a more dominant role as the population size increases. To further understand the role of correlations, we performed a shuffle test, where we broke correlations of all orders in the data by shifting each cell’s spike train by a random time shift (including periodic wrap-around) that was different for each cell. Following this grand shuffle, we repeated the full analysis procedure described above (fitting a maximum entropy model to the shuffled data and estimating the specific heat). We found that the heat capacity had a much lower and broader peak that did not change as a function of N. In addition, this heat capacity curve agreed closely with the analytical form of the specific heat for an independent neural population (S6 Fig). This analysis demonstrates that the sharpening of the specific heat that we observed is a direct consequence of the measured pattern of correlation among neurons [Fig 3E]. The shuffled curves were noticeably different across light adapted conditions (Fig 3D). This is not surprising as the analytical form for the specific heat of a network of independent neurons depends only on the average firing rate of each neuron, and these are substantially different between the two luminance conditions (Fig 2A). However, the heat capacity peaks for both the dark and the light conditions became more similar with increasing N. Clearly some macroscopic properties of the network were conserved across luminance conditions for the real, correlated, data (Fig 3). The correlation structure of natural movies can in principle trigger a broad set of observed retinal adaptation mechanisms, such as adaptation to spatial contrast [49, 50], temporal contrast [51], and relative motion of objects [52]. To generalize our results to these higher-order adaptive mechanisms, we ran another experiment comparing the distributions of responses of the same retina to two different natural stimuli ensembles, without a neutral density filter (Experiment #2, see Methods). These two natural movies were of grass stalks swaying in the wind (M1, the same movie as in the previous experiment), and ripples on the surface of water near a dam (M2). The first movie (M1) had faster movements, larger contrasts, and fewer periods of inactivity. Likely as a consequence, we found higher firing rates in ganglion cells during M1 (Fig 4A). We found statistically significant differences in the correlation coeffecients, Cij, and P(K) across the two stimulus conditions (Fig 4B and 4C). However, the specific heats of the full networks in the two movies sharpened similarly across conditions (Fig 4D), indicating that this macroscopic property of the retinal population code was also robust to different choices of naturalistic stimuli. So far, our results have demonstrated that the peak in the specific heat is due to the pattern of correlation among neurons. However, these correlations have contributions both from retinal processing, such as the high spatial overlap between ganglion cells of different functional type [53, 54], and from the correlation structure in the stimulus itself. In order to compare the relative importance of these two different sources of correlation among ganglion cells, we measured neural activity during stimulation with a randomly flickering checkerboard. By construction, our checkerboard stimulus had minimal spatial and temporal correlation: outside of 66 μm squares and 33 ms frames, all light intensities were randomly chosen. Returning to Experiment #1 in the light-adapted condition, we compared the response of the retina to the natural movie and the checkerboard stimuli (Note that here we are working with the N = 111 ganglion cells that were identifiable across both conditions, a subset of the N = 128 ganglion cells we worked with in Fig 3). The distribution of pairwise correlation coefficients was tighter around zero when the population of ganglion cells was responding to the checkerboard stimulus (Fig 5A). The specific heat in the checkerboard was smeared out relative to the natural movie, but was still very distinct from the independent population (Fig 5B). This suggested to us that most, but not all, of the contributions to the shape of the specific heat were shared across the two stimulation conditions, and therefore arose from retinal processing. A simple and popular view of retinal processing is that each ganglion cell spike train is described by the spatio-temporal processing of the cell’s classical receptive field. In this picture, correlation between ganglion cells arises largely from common input to a given pair of ganglion cells which can be described by the overlap of their receptive fields. To explore the properties of this simple model, we estimated linear-nonlinear (LN) models for each of the N = 111 ganglion cells in the checkerboard recording (Methods). We then generated spike trains from these model neurons responding to a new pseudorandom checkerboard sequence, and binarized them into 20ms bins in the same manner as for the measured neural data. As expected, the receptive fields had a large degree of spatial overlap [53, 55], which gives rise to significant stimulus-dependent correlations. We found that these networks did not reproduce the distributions of correlations found in the data, instead having lower values of correlation and fewer outliers (Fig 6A). The specific heat of the network of LN neurons was reduced relative to the neural data that the LN models were based upon (Fig 6B). Thus, the peak in the specific heat is enhanced by the nonlinear spatial and temporal computations in the retina that are not captured by models of the classical receptive field. Because the detailed properties of the maximum entropy model depend strictly on correlations measured within the neural population, we wanted to develop a more general understanding of what aspects of the pattern of correlation were essential. To do this, we altered particular properties of the measured matrix of correlations, keeping the firing rates constant. We then inferred the maximum entropy model parameters for these new sets of constraints, and estimated the specific heat. For these manipulations, we worked with the simpler pairwise maximum entropy model. We made this choice for several reasons. First, manipulating only the pairwise correlation matrix made our analysis simpler and more elegant than also having to perturb the distribution of spike counts, P(K). There is a large literature reporting values of pairwise correlation coefficients, helping us to make intuitive choices of how to manipulate the correlation matrix, while very little such literature exists for P(K). Additionally, any perturbation of the correlation matrix consequently changes P(K), so that attempting to change the correlation matrix while keeping P(K) fixed is a nontrivial manipulation. Second, in the pairwise model all effects of correlational structure are confined to the interaction matrix. This interaction matrix has been studied extensively in physics [56], and hence there is some intuition as to how to interpret systematic changes in the parameters. Conversely, we have little intuition currently for the nature of the k-potential. Our final and most important reason was that the qualitative behavior in the heat capacity (sharpening with system size, convergence across light and dark datasets) is the same for both pairwise and k-pairwise models across all conditions tested (S7 Fig). The correlation matrix in the retinal population responding to a natural stimulus has many weak but statistically non-zero correlations [54, 55], a result also found elsewhere in the brain [57, 58]. To test their contribution to the specific heat, we kept only the largest L correlations per cell, replacing the other terms in the correlation matrix with estimates from the shuffled (independent) covariance matrix. If our results are based on a “small world network” of a few, strong connections [59], then the specific heat for small values of L should begin to approximate our results for the real data. Clearly (Fig 7A), even keeping the top L = 10N (out of a total of L = 63.5N values) strongest correlations did not reproduce the observed behavior. Therefore, the full “web” of weak correlations contributed substantially to the shape of the specific heat of the retinal population code. We were next interested in understanding the qualitative nature of how networks transition between the independent and fully correlated regimes. Our approach was to scale all the pairwise correlations down by a constant (α), to subselect groups of neurons as previously in Fig 3, and to follow the inference procedure described above. Specifically, we formed a new correlation matrix C i j mixed ( α ) = α C i j true + ( 1 - α ) C i j shuff. We found that the specific heat of the neural population exhibited a transition between independent and fully correlated behavior, as the correlation strength, α, ranged from 0 to 1 (Fig 7B). Peaks in the heat capacity similar to the one observed in the full model emerge when α is greater than a critical value α*, which we estimated to be between 0.225 to 0.25 (Fig 7B and 7C). Substantially similar behavior was observed in the dark condition as well (Fig 7D). These data suggest that the low temperature phase emerges near α*. In fact this behavior constitutes another phase transition which can be observed without the introduction of a fictitious temperature parameter, in the curve of the specific heat of the real system (T = 1) as a function of correlation strength α (Fig 7C). There is a clear emergence of a contribution to the specific heat that depends on the system size, N. This additional contribution gives rise to a discontinuity in either the specific heat or its derivative. Similar behavior is observed in a classic model of spin glasses, the Sherrington Kirkpatrick (SK) model [60], as we describe in the Discussion. Importantly, the critical value of alpha at which we see a transition to structure, α*, was substantially smaller than the measured correlation strength, α = 1. This indicates that the population of retinal ganglion cells had a overall strength that was “safely” within the strongly correlated regime. Thus, the low temperature state is robust to changes in adaptational state or stimulus statistics that might shift the overall strength of correlations among neurons. Our hypothesis was that the emergence of a phase transition was correlated with the emergence of structure in the energy landscape. Previously, the structure of the energy landscape has been studied with zero temperature Monte Carlo (MC) mapping of local minima [16], where one changes the activity state of single neurons such that the energy of the population activity state always decreases. States from the data were thereby assigned to local minima in the energy landscape, which can be thought of as a method of clustering a set of neural activity patterns into population “codewords” [16]. If each cluster encodes a specific visual stimulus or class of stimuli, then this clustering operation provides a method of correcting for errors introduced by noise in the neural response. There are two reasons why we chose to study the structure of the energy landscape at the operating point of the system (T = 1). First, when we performed zero temperature descent with our models, our primary finding was that the overwhelming majority of states descended into the silent state (only 503 out of 1.75 ⋅ 105 did not descend into silence on a sample run). This indicated that the energy landscape had very few local minima. Thus we needed a different approach to explore the structure of the energy landscape. Second, we were interested in properties of the system (such as the specific heat) that were themselves temperature dependent, so it made sense to stick with the real operating point of the neural population (T = 1). When analyzing sufficiently large neural populations (typically, “large” means N > 20 cells), there are too many states to simply ennumerate them all. As a consequence, the energy landscape was accessed indirectly, through a Markov Chain Monte Carlo (MC) sampler [61], which simulates an exploration of phase space by defining the state-dependent transition probabilities between successive states. Provided that these transition probabilities are properly defined, the distribution of samples drawn should approach the desired (true) distribution with sufficient sampling. The set of these transition probabilities across all the neurons defines a ‘direction of motion’ in neural response space. We will study these directions of motion as a way to gain more insight into the properties of the energy landscape of our measured neural populations. Note, however, that MC sampling dynamics is used here as a tool to explore the geometric properties of the probability landscape of the retinal population; these are not claims of how the real dynamics of activity states change in the retina under the influence of the visual stimulus. To study the relationship between the directions of motion given by the MC sampling process and the observed phase transition, we returned to the manipulation with scaled covariances. For a given state R, the MC sampler in each model (inferred for a particular value of the correlation strength α) will return a vector of conditional probabilities X(R, α) = {xi}, where the conditional probability for each cell i’s activity is given by x i ( R , α ) = exp ( h i ( eff , α ) ( R ) ) / ( exp ( h i ( eff , α ) ( R ) ) + 1 ) (3) with an effective field, given by h i ( eff , α ) ( R ) = h i ( α ) + 2 ∑ j ≠ i J i j ( α ) r j (4) We can now ask how the shape of the energy landscape evolved with respect to α. Specifically, we first compared the similarity in direction of these ‘Monte Carlo flow’ vectors with the vectors defined for the fully correlated model (at α = 1), by calculating the average overlap between states, p, at a correlation strength α with the same states at α = 1, X ^ ( R p , α ) · X ^ ( R p , α = 1 ). We found that in the independent limit, α = 0, the flow vectors pointed in substantially different directions than in the fully correlated population (Fig 8A). This indicates that there is very different structure in the energy landscape in these two limits. Furthermore, as alpha increased from zero, we found a steep increase in the similarity of flow vector directions until a little past our estimate of the critical value, α ≈ 0.3, at which point the slope became more and more shallow. The probability landscape roughly settled into its fully correlated ‘shape’ by α ≈ 0.5, which was comparable with the point by which the specific heat had stabilized near its fully correlated value as well (Fig 7C). This is consistent with a tight connection between the emergence of a phase transition and the development of structure in the probability landscape. We carried out a similar analysis to compare the amplitude of Monte Carlo flow vectors as a function of the correlation strength, α (Fig 8B). At α = 0, we found that flow vector magnitudes were very different from the fully correlated population for states with high spike count, K. The similarity in amplitude increased gradually up to α*, increased sharply from α* up to α ≈ 0.5, and then changed slowly at higher values of α. Again, these results are consistent with the interpretation that the shape of the energy landscape emerged at a correlation strength near α* and that further increases in α served to ‘deepen’ the existing contours in the energy landscape. While our energy landscape does not have many true local minima, we can gain insight into the nature of the energy landscape induced by correlations by considering how long the system remains in the vicinity of a given state under T = 1 MC dynamics. Since the experimentally measured neural activity is sparse, the directions of motion are heavily biased towards silence. Regardless of initial state, the sampler will eventually revisit the silent state. To demonstrate this, we returned to the data from Experiment #1 (M1, light), and the corresponding k-pairwise model fits. Due to the addition of the k-potential in this analysis, the effective field was now h i ( eff ) ( R ) = h i + 2 ∑ j ≠ i J i j r j + λ K + 1 - λ K (5) with K = ∑j≠i rj.The effective field is derived as the difference in energy of the system when cell changes from silent to spiking (Eq 3), and so it has three contributions: one from the local field, hi, another from the sum of pairwise interactions, and a final term from the change in the k-potential, λK+1 − λK. We worked with all the states observed in the light condition that had K = 12 spiking cells (total number of states = 3187). We selected a set of initial activity states all having the same value of K, as these analyses depended strongly on K. We wanted a value of K large enough that effective fields were large, and hence collective effects of the population code were significant. At the same time, we needed K to be small enough that we could observe many such states in our sampled experimental data. Balancing these two concerns, we chose K = 12. In order to characterize the ‘dwell time’ of a particular state, we initiated many MC sampling runs from that given state. On each of these runs, we defined the dwell time as the number of MC samples (where all N cells were updated) required to change 9 of the 12 originally spiking cells to silent. For each given initial state, our analysis produced a distribution of dwell times, due to the randomized order of cell choice during sampling, as well as the stochasticity inherent in sampling at T = 1. We found significant differences in the distributions of dwell times across different initial states (Fig 9A). This demonstrated that Monte Carlo flow was trapped for a longer amount of time in the vicinity of particular states, consistent with subtle attractor-like properties in the geometric structure of the probability landscape. For the same initial states, average dwell times measured on the energy landscape for independent models were almost an order of magnitude shorter, indicating that these effects were due to the measured correlations (Fig 9B). We searched for a measure that could capture the variability in dwell times across initial states in the full model, based on our intuition that a state that started near a local minimum would have a long persistence time before finite temperature MC sampling would move it far away. This led us to define a persistence index (PI) that captured the tendency of a state to remain near its starting point under MC sampling dynamics. Specifically, for a given state R, we define P I = X ^ ( R ) · R ^, namely the cosine of the angle between the initial state and the average next state. If the PI is close to 1, then the direction that the state R evolves towards under MC sampling is the state itself, and hence the state will remain the same. Over all the initial states studied, we found a large positive correlation between the average dwell times and the persistence indices for the fully correlated population (correlation coefficient of 0.75, Fig 9C). This significant correlation justifies the use of the PI as a simpler proxy for the dwell time. In contrast, the correlation was only 0.38 when measured on the energy landscape of the independent model and dwell times were systematically smaller by orders of magnitude (Fig 9C blue). The persistence index also allowed us to characterize the similarity of the neural code for the same natural movie between light and dark adapted conditions. If the structure in the energy landscape changed between the dark and light conditions, then one would expect the dwell times, and hence the persistence indices, to change as well. Instead, we found a strong correspondence across the light and dark experimental conditions, that was absent in the independent model (Fig 9D, S8 Fig correlation coefficient of 0.90 vs 0.37). To estimate the variability in this measure, we compared the PI across models inferred for two separate random halves of the light condition (S8 Fig, correlation coefficient of 0.97). Thus, the pattern of correlation measured in the two luminance conditions created similar structure in the system’s energy landscape, even though the detailed statistics of neural activity were quite different. This structure endows the population code with a form of invariance to light level that is not present at the level of individual ganglion cells. Across different naturalistic stimulus conditions, we observed a similar sharpening of the specific heat as we increased the number of retinal ganglion cells, N, that were analyzed together. While this phase transition occurs at a temperature that does not correspond to any measured neural population, it nonetheless gives us insight into the collective properties of that real population. We’ve shown that aspects of this structure are conserved across large changes in average luminance, and that this structure is largely insensitive to minor perturbations of the correlation matrix. Our work ties together two seemingly disparate ideas: first, the observations of phase transitions in models of the real neural population [19, 21], and second, that the probability landscape might be organized into a discrete set of clusters [8, 62, 63]. We suggest that the former is a direct consequence of the latter. Representing visual information with specific multineuronal activity states is an error-prone process, in the sense that responses of the retina to repeated presentations of identical stimuli evoke a set of activity states with different probabilities. A many-onto-one mapping of individual activity states to cluster identities naturally reduces this variability, thus endowing the population code with a form of error correction. In fact, this appealing and intuitive idea has been recently demonstrated for the retinal population code [62, 63]. Clustering of activity states manifested itself in the geometric structure of the probability distribution, which we characterized by several analyses, including Monte Carlo sampling dynamics, Monte Carlo flow vectors, and persistence indices. This structure was found to be preserved across variations in ambient luminance, and was robust to minor perturbations of the correlation matrix. This robustness was not obviously evident in the lower order statistics of the distribution (firing rates, correlations, spike count distribution), which were measured directly. For downstream readout mechanisms which access only the incoming retinal population code, such a robustness in the clustered organization constitutes a form of invariance in the retinal representation. The variability in the interaction matrix gives rise to the variability that we observed in measures of persistence across states with the same number of spiking cells K (see Fig 9). In our picture in which the probability distribution over all neural activity states is organized into a set of clusters, some states are ‘attractor-like’. These states have a higher density of nearby states, which corresponds to lower energies, and thus traps states in their vicinity under MC sampling dynamics. Other states do not have this property at all, and hence the dwell time around these states under MC sampling dynamics approaches that of a network of independent neurons. This property depends crucially on the detailed structure of the pairwise interactions Jij. If for instance all the interaction matrix terms were set equal to a positive constant as in a ferromagnet, i.e. Jij = J0, then the effective fields for all cells would have almost the same contribution from interactions, namely a quantity proportional to KJ0. The only variability in the conditional probabilities, X(R), that cells would experience would be due to the local fields and whether or not the cells were active in the state (which reduces the effective K by one); this variability would be overwhelmed with increasing K. As a result, the effective field would eventually tend to a constant for all cells at large enough K. However, we observe in our data that this is not true: the persistence index of states with the same K varied substantially (Fig 9). The effect of variability in the distribution of interaction matrix terms has been studied extensively in spin glass models [56]. In order to convey some of our intuition about the low temperature regime, we will discuss a particular example of a glassy model and its relationship to our work. But keep in mind that while we believe there is a useful analogy between some of the properties of the glassy limit of the Sherrington-Kirkpatrick (SK) model [60] and our measured neural distribution, we are not claiming that our distributions are fit by the SK model. The SK model is a model of all-to-all connectivity in a population, similar to the high degree of connectivity we observe in our inferred models of a correlated patch of ganglion cells. For example, cells have an average of 46 non-zero interactions per cell, out of N = 128 possible in the k-pairwise model in the light condition. The SK model itself has two regimes, characterized by the relationship between the variance (σ J 2) and mean (μJ) of the interaction parameters, Jij. When the variability is large relative to the mean, the low temperature phase of the SK model is a spin glass phase, where the probability distribution over all activity states is characterized by an abundance of local maxima. The glassy SK model also undergoes a phase transition from a weakly-correlated paramagnetic phase to a structured spin glass phase as temperature is lowered. This transition is an ergodicity-breaking transition that is third-order (see below) [56, 64]. In other words, this transition is characterized by a significant reduction in the number of states available to the system, as the glassy state confines the system to particular valleys in the energy landscape. A note is in order: ergodicity-breaking phase transitions can only be formally defined when there is no possibility that the system can escape a particular valley. This is only true mathematically in the thermodynamic limit (N → ∞), as a finite-sized system will always have a small but nonzero probability with which it can escape the phase space it is confined to. So we should always keep in mind that as we compare our data to the SK model, we need to consider the finite size limit of the SK model. By comparison, when we scaled up the correlation strength α from values below the critical correlation strength, α*, to values above, we observed a phase transition where the smooth independent distribution wrinkled to form a set of attractor-like states in the energy landscape (Figs 7, 8 and 9), perhaps with the geometry of ‘ridges’ [65]. While this wrinkling transition does not strictly break ergodicity, like in the thermodynamic limit of the SK model, it does confine the system near attractor-like states. This confinement may in fact be quite similar to the ergodicity-breaking transition in the SK model. Indeed, our analysis of this transition suggests that it is consistent with a third-order transition, where there is a cusp in the specific heat at the critical temperature (i.e. a discontinuity in the derivative of the specific heat, not in its value; see S8 Fig). This behavior is reminiscent of the fact that the ergodicity-breaking transition for a spin glass in the SK model is also a third-order phase transition (see Figure 3 in ref. [60]). These similarities suggest that the geometric properties of the probability landscape of the neural population are somewhat akin to the properties of the SK model in the spin glass phase, which is appealing for the connection between phase transitions and clustering of neural activity. But again, we are not arguing that our probability distributions over neural activity are exactly reproduced by the SK model. For instance, there are no local fields in the SK model, and these play a significant role in the properties of our maximum entropy models of neural data. Fundamentally, error-correcting structure is not present in populations of independent neurons: if one neuron in an independent population misfires, that neuron’s information is lost [66]. The qualitative separability between the regimes of error-correction and independence is our proposed origin of the observed phase transition with respect to our temperature variable [19, 21]. To summarize these ideas, we present a suggested picture of the phase transitions studied in our work, by comparison with the Ising ferromagnet and the SK spin glass models (Fig 10). In the Ising ferromagnet at low temperature (Fig 10A), the constant and equal interactions cause all the cells to tend to be active or quiet simultaneously. At high temperature, fluctuations wash out the interactions and the system is in a weakly correlated (paramagnetic) phase. As temperature is decreased at zero applied field, the system reaches a critical point where it chooses the all-active or all-quiet half of phase space (grey dot in Fig 10A). Below the critical temperature the two halves of phase space are separated by a line of first-order phase transitions that separates a ferromagnetic phase with all the spin aligned in one direction from a similar ferromagnetic phase with all the spins aligned in the opposite direction (black line in Fig 10A). As described above, the SK model has both a spin glass and a ferromagnetic limit (Fig 10B, sketched following [67]). In the SK model in the spin glass limit (μJ ≪ σJ), decreasing the temperature at zero applied field causes the model to freeze into a spin glass phase. This ergodicity-breaking phase transition is third-order (Fig 10B, magenta arrows). In our picture of the retinal population code, a third-order phase transition occurs when the correlation strength α is increased from 0 at T = 1 (gray dot and magenta arrow in Fig 10C). In this phase transition, the distribution of neural activity wrinkles to form a set of ridges in the probability landscape. Because firing rates are constrained during this manipulation, this manipulation is analogous to varying temperature at zero applied field in the nearest neighbor Ising ferromagnet, and to varying temperature in the spin glass limit of the SK model. Increasing the temperature in our models of the retinal population when the correlation α is greater than α* causes the distribution to melt to a weakly correlated state in a transition that is first-order (see Fig 7B). Because of the local fields in our models, variations in temperature are accompanied by changes in firing rate, and this is most similar to varying the applied field, h, in the Ising ferromagnet. The analogy between the axes in the Ising ferromagnet and our model only extends to the horizontal (temperature) axis of the SK model: here the phase space far from the dotted line is shown for clarity. So in summary, the main phase transition that we see when we change temperature is first-order, as is the transition between ferromagnet states in the Ising model when the applied field is changed. In both cases, there is a substantial change in the state of individual elements—the firing rate of neurons in the retinal population and the magnetization of spins in the Ising model. Furthermore, the 3rd-order phase transition that we observe when α increases above α* is reminiscent of the 3rd-order phase transition in the SK model as a function of temperature in the limit of highly variable interactions. In our case, increasing α increases the strength of correlations, while in the SK model, lowering temperature increases the impact of interactions on the network state. This analogy provides support for the conclusion that the low temperature phase of the retinal population resembles a spin glass. In statistical physics, the peak in the specific heat that we observed in our models (Fig 3) could be consistent with two types of phase transition, which are classified by the order of the derivative of the free energy which exhibits a discontinuity. In statistical mechanics, every physical system can be described on a macroscopic level by a free energy function. When the system transitions between phases, some order of the derivative of this free energy function will have a discontinuity. The first-order derivative of free energy versus temperature is the entropy, so when there is a discontinuity in the entropy, the system is said to exhibit a ‘first-order’ phase transition. The second-order derivative of free energy versus temperature is proportional to the specific heat, so a system with a discontinuity in the heat capacity exhibits a ‘second-order’ phase transition. The ambiguity we suffer in interpreting our data arises due to the fact that phases and transitions are rigorously defined in the thermodynamic limit which is when the system size N is taken to infinity. At finite sizes, it may be difficult to tell these two possibilities apart. The first possibility is that the observed phase transitions exhibit a divergence in the specific heat, making them second-order. This would indicate criticality in the retinal population code [19, 21, 27]. The second possibility is that the entropy is discontinous and that the specific heat exhibits an infinite value only at the critical temperature (i.e. the specific heat has a delta-function form). This would be a first-order transition, and it does not indicate criticality. Both of these hypotheses would be consistent with a sharpening of the specific heat as the system size increased (Fig 3). How could we distinguish definitively between these two types of transition and what are the consequences of the difference? This issue would be resolved if we could convincingly relate the properties of our model to some well known example in physics that does have a critical point. There are two examples that we have in mind here. The first is the symmetry-breaking mechanism for criticality (see for example chapter 142 onwards in [5]). For concreteness we’ll take the example of the structure in the nearest neighbor Ising ferromagnet, where all the non zero interactions Jij are a positive constant and there are no other parameters. In the high temperature state, the system has the ability to visit any one of the 2N possible states, and the average firing probability for each neuron is 0.5 in a single time bin. In the low temperature state, the interactions cause all the cells to behave similarly, either mostly silent or mostly active, with some fluctuations allowed by the temperature. Importantly, however, the two phase spaces centered on all-silent and all-active are separated by an energy barrier that increases with system size. Because of this property, at large enough system sizes the available phase space in the low temperature state is reduced by a factor of two. The low and high temperature phases are known as low and high symmetry states, respectively, and it is the change in symmetry at the phase transition which leads to discontinuities in the specific heat. The critical point here is thus characterized by a reduction in symmetry where the symmetry described in this example is the deviation of the average firing rate from one half. In our models, the presence of local fields means that all cells have some bias towards silence or activity. As a consequence, the symmetry in firing rate is absent regardless of the temperature, and no symmetry breaking can occur with respect to changes in firing rates. There might be some other symmetry that is broken at the transition between phases, but no one has identified it yet [19, 21]. Failing to identify a symmetry breaking mechanism in our analysis does not prove that the peak in the heat capacity is a first-order phase transition, but it is consistent with this interpretation. The second type of critical point that we’ve considered is the type that occurs in a spin glass, as in the SK model. These critical points are also characterized by a reduction in the size of the available state space, as in the symmetry breaking example described above. Because these transitions occur between paramagnetic and glassy regimes, they are also candidates for describing the transition that we observe in Fig 3. However, these critical points in glassy systems are not, to our knowledge, characterized by a divergence in the specific heat. The second-order transition in the SK model is discontinuous in the specific heat, but not divergent (see Figure 3 in ref. [60]). If such a critical transition did occur, it would not lead to a sharpening of the specific heat as we observe in Fig 3. Instead, it would resemble the transition we observe as a function of increasing correlation strength, α, where the specific heat is either discontinuous or has a cusp at α* (Fig 7 and see our discussion in the supplement, S8 Fig). Additionally, such critical points are typically marked by a divergence in some higher order statistic, such as the nonlinear susceptibility. Our measurements of the nonlinear susceptibility have not shown such a result (S11 Fig). To summarize, we have considered two natural mechanisms that could connect our observed peak in the heat capacity versus temperature to a well known critical point in statistical physics, and we find that these mechanisms are simply not consistent with the observed properties of the data. It is possible that some other analogy provides a connection between criticality and our observations, but we are not aware of it. So, taken together, we believe that these observations argue against the hypothesis that the retinal ganglion cell population is poised at a critical point. Criticality implies that there is something special in the distribution of neural responses: for example, that the specific heat is maximized with respect to some properties of the retinal circuit, and hence that “the distribution of equivalent fluctuating fields must be tuned, rather than merely having sufficiently large fluctuations” [19]. Our analysis in which we scaled the strength of all pairwise correlations by a constant factor is not consistent with the notion of fine tuning. Specifically, we could decrease pairwise correlations by a factor of more than 2 without significant changes in the specific heat (Fig 7C) or the geometric structure of the probability distribution (Fig 8). Additionally, we actually observed higher peaks in the specific heat at T ≠ 1 in the partially correlated networks than in the fully correlated networks (Fig 7). This fact also argues against the idea that the heat capacity of the system is strictly maximized by some principle requiring fine tuning. The interpretation of the peak in the heat capacity as a first-order phase transition also provides an explanation for the proximity of the system to the transition. Our argument is that correlations in the distribution create a phase which is qualitatively different than the high temperature phase. This qualitative difference then requires that there be a a transition between these two regimes. Furthermore, the sparseness of neural activity implies that the zero-temperature limit of the model is the all-silent state. Consequently, the average firing rates must follow a sigmoidal function of temperature, starting at zero for T = 0, rising for T > 0, and then saturating at 0.5 firing probability for T → ∞. This sigmoid sharpens into a step with increasing system size, with the width of the step corresponding to the area over which finite-size effects smear out the phase transition (see Fig 11). Such a step-like change in the firing rate is most consistent with a first-order phase transition (where first-order derivatives of the free energy, such as the firing rates, change discontinuously). Given this context, constraining the average firing rate to be some small but nonzero value forces the system to be poised in the vicinity of the phase transition. Thus, we believe that the proximity of the system to the transition point simply follows as a consequence of constraining the structured phase to have neurons with low firing rates. In addition, it is difficult to reconcile the notion that the retinal code is poised at a very special, fine-tuned operating point with the observation that similar peaks in the specific heat arise in many circumstances, including very simple models [27]. In this sense, we agree with the conclusion that Nonenmacher and colleagues reached that this notion of fine tuning is not well supported by these wider considerations. In contrast, a first-order phase transition does not imply that something is “special” or “optimal” about the retinal population code. Instead, the hypothesis that the population is in a low temperature state is appealing, because this state can be robustly present without fine tuning. To sum up, our heat capacity analyses do not clearly disprove the hypothesis that the system is poised at a critical point. However, first-order phase transitions are far more common in nature than second-order transitions. They do not require any special tuning of the parameters of the system. Thus, the interpretation that the neural population is in a low temperature state serves as a simple hypothesis that is consistent with all of our data. This interpretation has additional value, because it suggests a connection between the phase transition and the emergence of structure in the probability landscape. And in fact, our analyses have directly confirmed this connection (Figs 8 and 9). While we believe there are no inconsistencies in our empirical findings and those of other studies that focused on criticality in retinal population codes, our interpretations differ substantially [24, 25, 27]. A recent study tested the generality of phase transitions in simulations of retinal ganglion cells [27], simulating networks of neurons in the retina, and then estimating the specific heat following procedures similar to those presented here. Their results are consistent with ours in that phase transitions were robustly present in different networks of neurons, and the presence of these phase transitions was largely invariant to experimental details. Similar to us, they also found that the sharpness of the peak in the specific heat was systematically enhanced by stronger pairwise correlations. Because Nonnenmacher et al. also found this behavior in very simple models, such as homogeneous neural populations, they concluded that the sharpening peak in the heat capacity does not necessarily provide insight, by itself, into the structure of the population code. We agree. However, because we went on to analyze the structure of the probability landscape for our real, measured neural populations, we could show that the emergence of the peak was related to the clustering of neural activity patterns (Figs 8 and 9). The relationship between the low temperature phase and clustering is not understood in general. Studies of homogeneous models of neural populations demonstrate that the low temperature phase is not sufficient for clustering [27], while the current study suggests that the low temperature phase might be necessary. Other factors, such a sufficient heterogeneity of single neuron properties, are presumably also required. In any case, we interpret the robustness of a phase transition in correlated neural populations not as an argument that this property is trivial, but instead as evidence for the generality of clustering in neural population codes. A separate line of work has investigated the presence of Zipf-like relationships in the probability distribution of neural codewords [23–25]. A true Zipf Law is intimately related to a peak in the heat capacity at T = 1, and Schwab et al. found that a Zipf Law was present under fairly generic circumstances, in which neural activity was determined by a latent variable (e.g., an external stimulus) with a wide range of values [24]. Again, this result is broadly consistent with our finding of great robustness of the low temperature state, and we interpret this a positive evidence for these properties being generically present in correlated neural populations. However, we can’t make more detailed comparisons to [24, 25], because we have not chosen to analyze Zipf-like relationships in our experimental data. There are several reasons for this choice: 1) we can only sample ∼2 orders of magnitude in rank (S9 Fig), making it difficult to estimate power law exponents; 2) we typically observe small deviations from the power law trend, and we are uncertain about how to interpret the importance of these “bumps”. In our study, we have characterized the neural response in a single time bin, ignoring the role of temporal correlations across time bins. One can extend our approach to include temporal correlations by concatenating multiple time bins into each neural codeword. When the number of total time bins was systematically increased in such a manner, the peak in the specific heat sharpened substantially [21]. The authors interpreted these results as further evidence in favor of the critical properties of neural population codes. However, in all cases in both our study and of [21], the peak was above T = 1, consistent with our interpretation that neural populations are in a low temperature state. Since increasing the number of time bins this way drastically increases the complexity of the distribution, this treatment of temporal correlations increases the structure of the low temperature state in a manner similar to an increase in the number of neurons analyzed together, N. Using several different analysis methods, neural activity evoked by repeated presentations of the same stimulus has been shown to form clusters in the space of all possible activity patterns. Zero temperature descent in the energy landscape defined by the maximum entropy model mapped a large fraction of all neural activity patterns to non-silent energy basins, which were robustly activated by the same visual stimulus [16, 68]. Mapping neural activity patterns to latent variables inferred for a hidden Markov model revealed similar robust activation by the stimulus [63, 65]. Huang et al. found a form of first-order phase transition as a function of the strength of an applied external field, from which they concluded that the energy landscape formed natural clusters of neural activity with no applied field [69]. Ganmor et al. recently uncovered a striking block diagonal organization in the matrix of semantic similarities between neural population codewords [62], arguing for a clustered organization of neural codewords. All these analyses are likely to be different ways to view the same underlying phenomenon, although a detailed exploration of the correspondences among these methods is a subject for future work. Are our results specific to the retina? In our approach, the collective state of the retinal population code is entirely determined by the pattern and strength of measured correlation. There is nothing about this pattern of correlation that makes specific reference to the retina. This means that any neural population having similar firing rates and pairwise correlations would also be in a similar collective state. Additionally, the strength of pairwise correlations we report here is smaller than or comparable to those reported in higher order brain areas, such as V1 and MT [57, 58, 70, 71]. This suggests that the collective state of neural activity, which arises due to a clustering of neural activity patterns, could occur throughout higher-order brain regions in population recordings of a suitable size (N >100). This study was performed in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the Institutional Animal Care and Use Committee (IACUC) of Princeton University (Protocol 1828). We recorded from larval tiger salamander (Ambystoma tigrinum) retina using the dense (30 μm spacing) 252-electrode array described in [1]. In Experiment #1, which probed the adaptational state of the retina at normal and low ambient illumintation levels, the salamander was kept in dim light and the retina was dissected out with the pigment epithelium intact, to help the retina recover post dissection and adjust to the low ambient light levels in the dark condition. The rest of the procedure in Experiment #1, and the full procedure for Experiment #2, followed [1]. The chromatic checkerboard stimulus (CC) consisted of a random binary sequence per color (R,G,B) per checker, allowing 8 unique values for any given checker. Checkers were 66 μm in size, and refreshed at 30 Hz. There were two (gray scale, 8 bit depth) natural movies used: grass stalks swaying in a breeze (M1, 410 seconds) and ripples on the water surface near a dam (M2, 360 seconds). Both were gamma corrected for the CRT, and displayed at 400 by 400 pixel (5.5 μm per pixel) resolution, at 60 Hz. In Experiment #1, after adapting the retina to the absolute dark for 20 minutes, we recorded in the dark condition first (by placing an absorptive neutral density filter of optical density 3 [Edmund Optics] in the light path), stimulating with (CC) for 60 minutes, and with (M1) for 90 minutes. The filter was then switched out for the light condition, in which we recorded for an additional 60 minutes of (M1) and another 60 minutes of (CC). To avoid transient light adaptation effects we removed the first 5 minutes of each recording (10 minutes from the first checkerboard) from our analysis. During stimulation with (M1) we sampled 340 sec long segments from (M1) with start times drawn from a uniform distribution in the [0 60] second interval of (M1). The spike sorting algorithm [1] was run independently on the recordings in response to (M1) at the two light levels, generating separate sets of cell templates at the two light levels, which were then matched across the two conditions, yielding N = 128 ganglion cells. The spike trains were then binned in 20 ms time bins and binarized, giving 2.5 ⋅ 105 states in the dark and 1.75 ⋅ 105 states in the light. For the recordings from the checkerboard (in both light conditions), the templates from the light recording were used to fit the electrode activity. Across all four stimulus conditions this left us with N = 111 cells for the comparisons of natural movies to checkerboard. The checkerboard in the light condition was binned into N = 145623 states. In Experiment #2, we alternated stimulation between (M1) and (M2) every 30 seconds, sampling 10 sec segments from both movies. For our analysis here we worked with the statistics of the last 9.5 seconds of each 30 second bout, yielding 8 ⋅ 104 states per stimulus condition, for N = 140 cells. Our maximum entropy model inference process implements a modified form of sequential coordinate gradient descent, described in [16, 47, 48], which uses an L1 regularization cost on parameters. For the k-pairwise model, we inferred without a regularization cost on the local fields. Further details are given in the Supplement (S1 Text) To measure the heat capacity we simulated an annealing process. Initializing at high temperatures, we monte carlo sampled half a million states per temperature level (in 100 parallel runs of 5 ⋅ 103 samples each), initializing subsequent lower temperature runs with the final states of preceeding higher temperature runs. The heat capacity at a particular temperature was then evaluated as C = (〈E2〉 − 〈E〉2)/T2. Our model LN neurons were estimated over the chromatic checkerboard recording in the light condition. For each cell, the three color-dependent linear filters (the full STA) were weighted equally before convolution with the stimulus for an estimate of the linear response q. The non-linearity was estimated over the same data by Bayes rule, P(spike|q) = P(q|spike)P(spike)/P(q). Spike trains were simulated from a novel pseudorandom sequence put through the model’s filters and non-linearity, with the non-linearity shifted horizontally to constrain the firing rates of the neurons to be the same as in the experimental recordings. The result was binned and binarized to yield N = 145623 states, for which we inferred the k-pairwise model.

10.1371/journal.pntd.0007269

Low population Japanese encephalitis virus (JEV) seroprevalence in Udayapur district, Nepal, three years after a JE vaccination programme: A case for further catch up campaigns?

The live attenuated Japanese encephalitis (JE) vaccine SA14-14-2 has been used in Nepal for catch-up campaigns and is now included in the routine immunisation schedule. Previous studies have shown good vaccine efficacy after one dose in districts with a high incidence of JE. The first well-documented dengue outbreak occurred in Nepal in 2006 with ongoing cases now thought to be secondary to migration from India. Previous infection with dengue virus (DENV) partially protects against JE and might also influence serum neutralising antibody titres against JEV. This study aimed to determine whether serum anti-JEV neutralisation titres are: 1. maintained over time since vaccination, 2. vary with historic local JE incidence, and 3. are associated with DENV neutralising antibody levels. We conducted a cross-sectional study in three districts of Nepal: Banke, Rupandehi and Udayapur. Udayapur district had been vaccinated against JE most recently (2009), but had been the focus of only one campaign, compared with two in Banke and three in Rupandehi. Participants answered a short questionnaire and serum was assayed for anti-JEV and anti-DENV IgM and IgG (by ELISA) and 50% plaque reduction neutralisation titres (PRNT50) against JEV and DENV serotypes 1–4. A titre of ≥1:10 was considered seropositive to the respective virus. JEV neutralising antibody seroprevalence (PRNT50 ≥ 1:10) was 81% in Banke and Rupandehi, but only 41% in Udayapur, despite this district being vaccinated more recently. Sensitivity of ELISA for both anti-JEV and anti-DENV antibodies was low compared with PRNT50. DENV neutralising antibody correlated with the JEV PRNT50 ≥1:10, though the effect was modest. IgM (indicating recent infection) against both viruses was detected in a small number of participants. We also show that DENV IgM is present in Nepali subjects who have not travelled to India, suggesting that DENV may have become established in Nepal. We therefore propose that further JE vaccine campaigns should be considered in Udayapur district, and similar areas that have had fewer vaccination campaigns.

In Nepal, immunisation using a live attenuated vaccine is given against Japanese encephalitis (JE), caused by the mosquito-transmitted JE virus (JEV). JE immunisation has taken place via catch-up campaigns and is now part of the routine immunisation programme. Although previous studies have shown good vaccine efficacy in areas where there is a lot of natural exposure to the virus (high endemicity), it is suggested that the efficacy may wane in areas where transmission is lower. Dengue virus (DENV) belongs to the same family and genus as JEV. Previous infection with DENV may also influence the immune response to JEV. Therefore, we conducted a cross-sectional study in Nepal to measure immunity to JE, in districts of differing historic JE incidence, and time from JE vaccination. This showed that neutralising antibody to JEV was found more frequently in districts which had been the subject of more vaccination campaigns, rather than in the most recently vaccinated district. In addition, we cannot rule out a role for natural exposure to JEV in maintaining higher antibody levels. Additionally, the study showed that previous exposure to DENV was positively associated with an immune response to JEV, though this effect was modest. We conclude that there is a need to consider further JE vaccine catch up campaigns in some areas especially given that we could detect JEV IgM, indicating ongoing transmission. We show that ELISA yielded many false negative results for exposure to JEV or vaccination, when compared with neutralising antibody. We also identified some individuals during the course of the study with DENV IgM in their blood, but with no history of travel to India. This suggests that DENV may have become established in some areas of Nepal.

Japanese encephalitis virus (JEV), an arthropod-borne virus belonging to the genus Flavivirus of the family Flaviviridae, is transmitted by Culex mosquitoes, with birds and pigs acting as natural reservoirs and amplifying hosts. Humans are an accidental dead-end host [1]. JEV is enzootic in many parts of rural South and Southeast Asia where it is a common cause of viral encephalitis, particularly in children [1–4]. Less than 1% of those infected develop symptoms [1]. Disease usually occurs in childhood or in adults who live in areas of low enzootic circulation of JEV and may be non-immune [2]. JE begins as an undifferentiated febrile illness followed in some by the development of seizures, altered sensorium or other neurological signs [5, 6]. Occasionally, acute flaccid paralysis and Parkinsonism occur [7, 8]. The case fatality rate is 20%-30% [2]; and neuropsychiatric disability is seen in up to 50% of survivors [9–12]. JE vaccines were first developed in the 1950s, and several types have been used [13]. JEV SA14-14-2 is an attenuated vaccine strain derived from JEV SA14, which was isolated from mosquitoes in China in the late 1940s. JEV SA14-14-2 was developed in China and licensed in 1988. JE was first seen in Nepal in 1978 [14] but vaccination against JE using SA14-14-2 was not introduced in the country until July 1999 when 224,000 doses of the vaccine were administered to children aged 1–15 years (as 0.5 ml intramuscular administration of single dose) in the Bardiya, Banke and Kailali districts [15]; however, reported vaccination coverage between the districts was variable (Bardiya 83%, Banke 41% and Kailali 22%)[16]. Between 2005 and 2010, 62 of the 75 districts in Nepal reported cases of JE, mostly those at lower altitude, thought to be because at higher altitudes there is less mosquito activity [17]. From 2006 to 2011, 31 high and moderate JE risk districts were included in a phased catch-up immunisation campaign; some districts vaccinated everyone in the population older than 1 year of age whereas others vaccinated only those aged 1–15 years [16, 18]. In addition, three doses of inactivated JE vaccine were given between 2000 and 2001 to children aged 6 months to 10 years living in the districts of Banke, Rupandehi, Kailali, Dang, and Kanchanpur [19]. In 2000, a single dose of SA14-14-2 vaccine was given to 98 children aged 1–15 years in Chitwan as part of a vaccine efficacy study [20]. Currently, JEV SA14-14-2 is given to children aged 12–23 months living in 31 high JE risk districts in Nepal as part of the national immunisation schedule introduced in a step-wise manner starting with 22 districts in 2009 and then expanded to 31 districts by 2012 [16]. Given the combination of catch-up campaigns and routine immunisation it is possible that a proportion of the population were vaccinated more than once. Despite cases of JE in Nepal still occurring particularly in the summer months [21], vaccination has resulted in a 78% reduction in the number of JE cases nationally [22]. Anti-JEV antibodies that develop after SA14-14-2 vaccination wane over time [23]. Whether further doses are required, however remains unclear, as clinical protection against JE may still occur despite low neutralising antibody titres [23, 24]. The efficacy of JE vaccine SA14-14-2 in a study in 2000 in Bardiya and Banke districts (higher JE incidence) was 98.5% after one year [25] falling very slightly to 96.2% after five years [26]. However, In Chitwan, an area of lower JE endemicity, neutralising antibodies against JEV persisted in only 63.8% of subjects 5 years after vaccination in 2000 [20] and a study from India showed that vaccine efficacy reduced from 87% in those vaccinated less than one year ago to 66% in those vaccinated a year or longer ago [27]. It has been shown that in older populations, after the antibodies wane following vaccination, they may rise again after natural infection [23]. Given that JE is seen more frequently in children in areas where the virus is not newly established [1, 28] it is therefore possible that natural infection may play a role in protection and a single dose of the vaccine may be sufficient in areas where maintenance of the antibody response, or of immunity, occurs due to high rates of natural infection. Dengue virus (DENV) is a flavivirus closely related to JEV that is transmitted by the Aedes aegypti and Aedes albopictus mosquitoes [29]. Dengue has only recently been documented in Nepal with the earliest case seen in 2006. Initially dengue cases were thought to be secondary to the movement of people from India, but more lately indigenous cases have occurred. In 2010 there was an outbreak in Chitwan and Rupandehi districts resulting in 917 cases. In 2012–2013 there were 77 laboratory confirmed cases [30]. Spatially, there have been cases of both JE and dengue documented in a number of districts [17]. There is immunological cross-reactivity between flaviviruses but the clinical effect of this is variable. Previous infection with DENV reduces the severity of subsequent JE [31, 32]. Vaccination with an inactivated JE vaccine might reduce the severity of dengue fever modestly [33]. Conversely, in a more recent study in Thailand, anti-JEV antibodies predisposed to symptomatic dengue fever, as opposed to clinically silent infection [34]. Despite the variance in clinical findings, given that JEV and DENV have been found in many of the same districts in Nepal [17], induction or maintenance of JE immunity by DENV infection is a theoretical possibility, leading to higher JEV seroprevalence, but less disease. The Nepali districts of Banke, Rupandehi and Udayapur are all JE enzootic. Banke reported 16.7 JE cases per 100,000 in 2004–6, tenfold higher than Rupandehi and Udayapur (1.6/100,000 in 2004–6) [18]. All three districts have been targeted by JE vaccination campaigns: Banke in 1999; Banke and Rupandehi in 2000–1 and 2006; and Udayapur in 2009. These JE vaccination campaigns were undertaken in all ages, with immunisation coverage rates reported to be 86% and over in all age groups [16]. Subsequently, the vaccination of children was introduced into the national programme for immunisation in all three districts in 2009 with vaccination coverage of 97% in Rupandehi, 76% in Banke and 71% in Udayapur [16]. Banke saw a reduction in the number of cases of JE, from ninety-four in 2005 to three in 2012. This compared with a reduction from nineteen cases in 2005 to zero in 2012 in Rupandehi, and from two cases to zero in Udayapur over the same period [16]. Udayapur is at higher altitude than Banke and Rupandehi, although the incidence of JE in 2004–6 in Udayapur and Rupandehi was the same, despite the difference in altitude, indicating an environment that can sustain JEV transmission in Udayapur. The existence of three districts of Nepal with varying JE incidence, which have been the target of single dose JE vaccination campaigns at different times in all ages regardless of JEV transmission intensity, provided an opportunity to compare the seroprevalence of JEV according to time from vaccination and level of historical JEV transmission (Fig 1). Moreover, there may be benefit from further JE vaccination catch-up campaigns in low JE incidence districts. This study was designed to answer the following questions: A sero-prevalence survey was conducted across the three districts of Banke, Rupandehi and Udayapur in Nepal (Fig 1). Banke and Rupandehi were sampled during October–November, and Udayapur in December 2012. Fieldwork was performed by research assistants employed by the Center for Molecular Dynamics, Nepal (CMDN). A representative random sample was selected in each of the three study districts using community-based stratified multi-stage cluster methodology as follows: Based on the average number of households visited during vaccination and surveillance campaigns, it was estimated that eight households per day could be surveyed. Sampling of 600 households was therefore deemed adequate within the time frame allocated for recruitment which was up to a maximum of six months. Accounting for 10% refusal rate, this would leave 540 households. It was estimated that on average, two adults per household would be recruited giving a total number of participants of 1080. The number of households per district was calculated based on the population proportional to sample size. Assuming that 50% of the total population sampled were JEV seropositive, using the R package ‘epiR’ and the function epi.clustersize, nine clusters (VDCs) would give a 95% certainty of being within 13% of the true population. Residents aged 19 years and over and who were present within the household at the time of recruitment were invited to participate in the study. This age group was selected for reasons of practicality, as drawing blood samples from well children in this setting can be challenging and a high rate of non-consent was assumed. In addition, as the vaccination campaigns included adults as well as children, sampling adults would give adequate insight into population level immunity, and would exclude sampling people too young to have been included in the campaigns. Written consent was obtained from participants after they had read or, in the case of those unable to read and write, had listened to a researcher reading out, the detailed description of the survey procedures contained in the participant information sheet. Those who were unable or unwilling to complete this consent process were excluded. Participants who consented to take part in the survey answered questions relating to their demographics, travel history (including family travel history), and recent history of fever within the past year or symptoms suggestive of JE or dengue fever using a simple questionnaire (Supp. File 1). A 5ml blood sample was collected for serum separation. All samples were collected in the period September–December 2012. No cases of dengue virus were found in any of the three study districts in 2012–2013 [30]. Anti-JEV IgM antibodies in serum were measured using the InBios JE detectTM IgM capture ELISA (sensitivity 99.2%, specificity 56.1% compared with the AFRIMS in house ELISA [36]; sensitivity 57%, specificity 95.4% compared with the CDC diagnostic algorithm including neutralising antibody [37]). Anti-DENV IgM antibodies in serum were measured using the InBios Dengue detectTM IgM capture ELISA (sensitivity 88.7%, specificity 93.1% [38]). IgG antibodies were measured using the In Bios JE and DENV detectTM IgG ELISAs, at the National Public Health Laboratory, Teku, Nepal. ELISAs were carried out according to the manufacturer’s instructions. For IgM ELISAs, test samples were incubated in wells coated with anti-human IgM antibody, followed by incubation with recombinant viral antigen or a control normal cell antigen (NCA), and anti-viral detection antibody. For IgG ELISAs, test samples were incubated with recombinant viral antigen (captured onto the plate via a monoclonal antibody), and NCA. The assay results are then expressed as a ratio of viral antigen to NCA, referred to as the immune status ratio (ISR). An ISR of >5 was considered positive, and an ISR of 2–5 equivocal. The DENV In Bios ELISA uses recombinant antigens of all four serotypes of DENV. Fifty percent plaque reduction neutralisation titres (PRNT50) against JEV (strain P3) and DENV serotypes 1–4 (Dengue 1: 16007, Dengue 2: 16681, Dengue 3: 16562, Dengue 4: C036/6) were measured at Mahidol University, Bangkok, Thailand using the method of Russell et al. (1967) [39]. Three serum ten-fold dilutions were used starting at 1:10 except in the case of two samples where there was minimal serum (<0.4ml) so the first dilution was 1:20. Ethical approval was obtained from Liverpool School of Tropical Medicine (LSTM) (12.19RS) and the Nepal Health Research Council (Reg no. 5/2012). The study was conducted according to the Declaration of Helsinki. All participants gave written informed consent. Data were analysed using STATA version 13 and R statistical software (www.r-project.org). For the dichotomous outcome measures, differences between the three study districts were evaluated using Fisher’s exact test and the z test to calculate the difference between proportions; adjustment was then made for important covariates using negative binomial regression with robust standard error estimates to take into account clustering effects at the VDC level. For PRNT50 < 1:10, values recorded as zero were imputed as exact figures were not available. Antibody levels were transformed to geometric mean titres to remove positive skewness and differences were compared using the Wilcoxon signed rank test, differences between ISR were also compared using the Wilcoxon signed rank test. A total of 1077 participants were recruited, of whom 943 provided fully completed consent forms and serum samples for antibody measurement (Fig 2). The demographic characteristics of these participants are summarised in Table 1. More females (59.8%; 95%CI 56.6–62.9%) than males (40.2%; 95%CI 37.1–43.4%) were recruited into the study with this difference being consistent across all districts. Overall, 23.2% (95%CI 20.6–26.1%) and 26.1% of participants (95%CI 23.3–29.0%) had visited India within the last year or had a family member who had. A higher proportion of respondents or their family members from Banke (26.1%, 95%CI 20.5–32.3% and 33.0%, 95%CI 27.0–39.5% respectively) and Rupandehi (32%, 95%CI 27.7–36.4% and 26.8%, 95%CI 22.8–31.1% respectively) had visited India within the last year compared with Udayapur (4.4%, 95%CI 2.2–7.7% and 18.4%, 95%CI 13.8–23.8% respectively). Sixty-one percent (95%CI 58.3–64.6%) of all respondents were farmers. Of those from Rupandehi, a higher proportion were farmers compared with the other districts. Twenty-five percent (95%CI 22.5–28.1%) of respondents had anti-DENV PRNT50 of ≥1:10; a higher proportion of those living in Rupandehi had an anti-DENV PRNT50 of ≥1:10 compared to the other districts (Table 1). The median age of respondents was 35 years and duration of residency, 27 years. More young than older participants were enrolled in all three districts, with a similar age profile in each district (S1 Fig). All 943 serum samples were assessed for the presence of JEV and DENV antibody by neutralisation assay (PRNT50), and IgM and IgG using InBios JE and DENV detect diagnostic ELISAs. Twenty-eight samples had anti-JEV IgM detected by ELISA (3.0%; 95% CI 2.1–4.3%), of which nine were JEV IgG negative, 17 equivocal and 2 were IgG positive, indicating recent infection and ongoing enzootic circulation of JEV. Six IgM positive subjects were from Banke, 17 from Rupandehi and five from Udayapur; the proportion of samples tested that were JEV IgM positive was similar across all three districts (Fig 3A). Ninety-six samples were anti-JEV IgG positive (10.2%; 95% CI 8.4–12.3%), 20 from Banke, 58 from Rupandehi and 18 from Udayapur. IgM ELISA results for each district are shown in Fig 3, expressed as the ratio of OD obtained for JEV antigen over normal cell antigen for the same sample (Immune Status Ratio, ISR). The median anti-JEV IgM log10 ISRs for Banke and Rupandhi were 0.33 and 0.34 respectively, corresponding to ISR values of 2.14 and 2.19. The value for Udayapur was 0.29 (ISR 1.95), significantly lower (p = 0.007). Seven samples had anti-DENV IgM (0.7%; 95% CI 0.4–1.5%) of whom one had been to India within the last year. The proportion of participants that had anti-JEV neutralising antibody (PRNT50 ≥1:10) differed across the study districts: seroprevalence was 81% in Banke (95% CI 75.5–86%) and Rupandehi (95% CI 77.1–84.4%), whereas it was significantly lower, 41%, in Udayapur (95% CI 35.4–48.0), the district with low historical JE incidence but the most recent vaccine coverage (2009) [Fisher’s exact test for comparison between Udayapur, and Banke and Rupandehi combined, p<0.001] (Fig 4). In this population, therefore, ELISA was insensitive for the detection of both anti-JEV and anti-DENV antibody compared with PRNT50, with sensitivity 18% (95% CI 15–21%) for JEV and 54% (95% CI 48–61%) for DENV (Table 2). Combining equivocal and positive results improved the sensitivity (59 (95% CI 55–63%) for JEV, 73% (95% CI 67–79%) for DENV), but still left a significant proportion of samples testing negative. For the JEV IgG ELISA alone, more suited to detection of immune memory, the performance of the assay was influenced by the presence of DENV neutralising antibody. For example, the sensitivity of the assay in DENV NAb negative subjects was only 47% (95% CI 42–52%) but the sensitivity in DENV NAb positive subjects was 71% (95% CI 65–77%), with a corresponding drop in specificity (Table 2). This indicates that the JEV IgG ELISA cross-reacts to some extent with DENV antibody. Seroprevalence of anti-DENV antibody measured by PRNT50 ≥1:10 was 25.2% (95% CI 22.5–28.1), or 238 individuals sero-positive (Fig 4). As was the case for JEV, this proportion differed significantly across the study districts: DENV seroprevalence in Banke was 20% (95% CI15.0–25.8), 32.2% in Rupandehi (95% CI 27.9–36.6) and 17.2% in Udayapur (95%12.7–22.5) [Fisher’s exact test between Rupandehi, and Banke and Udayapur combined, p<0.001]. Evidence of past infection by DENV serotype 2 was detected with DENV serotype 2 antibody found most commonly in 18.6% of participants (95% CI 16.1–21.2) and at highest titre (Fig 5). Most subjects with DENV neutralising antibody had antibody to two or more flaviviruses (220 of 238), and although most had antibody to DENV multiple serotypes, a number of participants had neutralising antibody only to DENV2, or to DENV2 and JEV. Similar, though less frequent, examples of monotypic DENV neutralising antibody could be identified for serotypes 1 and 4 (S2 Fig). No convincing monotypic antibody responses were identified for DENV serotype 3. In keeping with the likely circulation of DENV2, antibody titres to this serotype were higher, though less than anti-JEV titres (Fig 5A). The magnitude of JEV and DENV neutralising antibody titres were also correlated, though modestly (Spearman’s R 0.11–0.18, Fig 5B). The proportion of respondents with an anti-JEV PRNT50 ≥1:10 remained significantly lower in Udayapur than in both Banke and Rupandehi after adjustment (multivariate logistic regression) for age, gender, travel to India in previous year of self or a family member, working as a farmer and a level of anti-DENV PRNT50 ≥1:10 (Table 3). There was no significant difference between the proportion of respondents with an anti-JEV PRNT50 ≥1:10 in Banke and Rupandehi. Increasing age and a level of anti-DENV PRNT50 ≥1:10 were also identified as being (independently) associated with a level of JEV PRNT50 ≥1:10. JEV and DENV antibody titres were also correlated with each other, though the effect was modest (Spearman’s r 0.11–0.18, p < 0.0005 for all DENV serotypes, Fig 5). After a similar adjustment (multivariate logistic regression) for age, gender, travel to India in previous year of self or a family member and working as a farmer, the proportion of respondents with an anti-DENV PRNT50 ≥1:10 remained significantly lower in both Udayapur and Banke than in Rupandehi, the district with lower historical JE incidence and vaccinated earlier (2006). In addition, women were observed to be significantly less likely than men to have anti-DENV PRNT50 ≥1:10 (Table 4) however, significantly more males than females had travelled to India within the past year (31.7% (95% CI 27.1–36.7%) of males compared to 17.6% (95%CI 14.6–21% of females, p<0.001). There was no significant difference between the proportion of respondents with an anti-DENV PRNT50 ≥1:10 in Udayapur and Banke. The first objective of this study was to determine whether anti-JEV neutralising antibody levels are maintained over time in Nepal (by comparing residents of districts vaccinated 3 years apart) with the hypothesis that anti-JEV-antibody sero-prevalence would be higher among subjects from Udayapur compared with Rupandehi due to more recent vaccination campaigns. Secondly, we hypothesised that anti-JEV-antibody sero-prevalence would be higher among subjects from Banke compared with Rupandehi due to a higher natural exposure to JEV. In this study, the district targeted by the most recent vaccination campaign (Udayapur) had the lowest sero-prevalence of JEV neutralising antibody. Overall, approximately 70% of participants across the three districts surveyed had anti-JEV PRNT50 ≥1:10, but sero-prevalence was only 41% in Udayapur compared with 81% in Rupandehi and Banke. This difference is not accounted for by reported differences in vaccination coverage, but may be due to a number of factors. Banke and Rupandehi have both participated in more vaccination campaigns than Udayapur (Banke has had three campaigns, Rupandehi two but Udayapur only one) so it is possible that some participants in these two districts had received multiple doses of the vaccine, contributing to the higher sero-positive rate seen in these districts. Previous work has also shown that the impact of vaccination on JE incidence was lower in moderate-risk hill districts compared to high-risk Terai districts (43 compared with 84% reduction in incidence) [40]. Other reasons do potentially exist for these differences, such as differences in vaccine coverage rates (for example vaccine coverage in Udayapur may have been lower than estimated), vaccine handling, or the target population, rather than the underlying level of natural exposure to JEV [40]. However, in a longer term follow up study, although the impact remained greatest in the high-risk Terai districts, JE incidence was 62% lower in moderate risk Terai districts, 69% lower in moderate risk hill districts, and 89% lower in high-risk Terai districts, narrowing the gap somewhat between the high and moderate risk districts, and suggesting a causal role for JE vaccination [22]. The geography and climate of Udayapur is distinct to Banke and Rupandehi. Only 34% of the Udayapur district is classified as being ‘lower tropical’ with an elevation of 0-300m). In contrast, 79% and 89% of Banke and Rupandehi districts respectively are classified lower tropical [41]. It is therefore possible that there is less transmission potential in Udayapur. Additionally, JEV may have circulated in Banke and Rupandehi for longer than Udayapur; Rupandehi was the first reported district in Nepal to see JE in 1978 and it is thought that the virus was brought into Nepal from neighbouring North India, which shares a border with both Banke and Rupandehi but not with Udayapur [17]. Repeated natural exposure to JEV may result in maintenance of neutralising antibody, and hence could be another explanation for our findings. Although historically Rupandehi and Udayapur had similar JE incidence, this may not have remained the case and more recent data are not available. JEV IgM antibodies were found in participants living in all three districts with the highest proportion occurring in Rupandehi. However, the low numbers of IgM positive results in this study make it difficult to form conclusions about differences in recent JEV infections between the districts. We may have underestimated the number of recent infections in Udayapur, because this district was sampled slightly later than the other districts in December; in Nepal the JE season is August–September. This may also account for lower ELISA values in Udayapur. Despite this, these results indicate that JEV transmission is on-going in all three districts. Lastly, the occurrence of IgM antibodies may have been underestimated in Banke and Rupandehi as well, which were also sampled after the JE season (October and December), introducing the possibility that some participants had lost IgM after infection earlier in the same year. Although the fewer number of JE vaccination campaigns in Udayapur may be sufficient to explain the difference in seroprevalence, we cannot completely rule out a role for less natural transmission of JEV between Udayapur and the other two districts. The most likely explanation may be that some combination of these two factors explains the difference. Because we have not directly measured JEV transmission in these districts, we cannot say for certain what the relative importance of each factor is. Nevertheless, given that JEV clearly still circulates in Udayapur (indicated by the presence of JEV IgM), consideration of repeated vaccination campaigns in low/moderate JE incidence areas such as this is warranted. We also hypothesised anti-JEV and anti-DENV neutralising antibody titres would be positively correlated. We found a significant, though modest, correlation between PRNT50 ≥1:10 against DENV and JEV that remained after adjustment for differences between districts, a finding that is in keeping with previous studies, supporting our hypothesis [42]. Additionally, it was found that residents of Banke were significantly less likely to have detectable DENV neutralising antibody compared with those from Rupandehi. This is consistent with the occurrence of 150 cases of dengue fever in Rupandehi during the 2010 outbreak, whereas Banke saw none. Previously it has been thought that dengue fever in Nepal is mostly imported from India, but this study found no association between anti-DENV PRNT50 ≥1:10 and travel to India and only one out of seven with a positive DENV IgM who had crossed the border. This finding is consistent with the establishment of DENV (possibly serotype 2) in some districts of Nepal, at least temporarily, potentially informing the public health response to future dengue cases. Rupandehi had significantly higher DENV sero-prevalence than Udayapur, but Banke had only marginally higher DENV sero-prevalence. Combined with the modest association of DENV and JEV antibody, therefore, maintaining anti-JEV antibody by DENV infection is not a likely explanation for the difference in JEV sero-prevalance between districts. We found that ELISA had a low sensitivity compared with PRNT for detection of anti-JEV antibody. However, PRNT is expensive, labour intensive and not suitable for many settings. Moreover, neutralisation assays can become unreliable at low serum dilations which might be necessary to test the immune response to SA14-14-2 over a longer time period, where dilutions of less than 1:10 may be needed as SA14-14-2 does not always elicit a high level of neutralising antibodies [43]. Therefore, from a public health perspective, JE vaccination campaigns should be directed in response to CSF JEV IgM positive cases rather than population serosurveys. One limitation of this study is that there is no method for differentiation between antibody to JEV acquired by natural infection as opposed to vaccination. Although one study showed potential for a JEV NS1-specific indirect ELISA to differentiate between JE vaccination and past infection [44], this study was based on inactivated JE vaccine and live JE vaccine SA14-14-2 also contains NS1, which may induce antibody responses and therefore would be unable to definitively differentiate between the two. As it was not possible to determine how many vaccinations a participant had received, all were included in the main analysis even if they had moved to the study district after the vaccination campaign. This does not, however, affect our conclusions. Despite being unable to differentiate between the causes of acquisition of immunity to JEV, an understanding of the seroprevalence in the region is important to help guide immunisation campaigns, although, accurate identification of vaccination status would undoubtedly be beneficial. From an immunological perspective, it is not known whether it is simply the presence, or magnitude, of neutralising antibodies correlates with clinical protection against JEV. For example, amnestic antibody responses [45], or cellular immune responses [43], may also contribute and could be a marker of exposure even if antibody neutralisation is absent. Although we cannot differentiate a neutralising antibody negative but otherwise primed (by vaccination or exposure) individual from a completely naïve individual, it is nevertheless likely that there are more susceptible individuals in Udayapur than the other two districts under study. Finally, this study was conducted in adults, whereas JE is mostly seen in children. However, in this regard, our data likely represent an under-estimate of the susceptible population, which would only re-enforce our primary conclusion. Clearly, robust durable neutralising antibody responses have not developed in Udayapur province. Future work could extend and improve upon our findings, including: inclusion of a paediatric cohort; running a longer prospective cohort study; more detailed immunological studies to determine more accurate markers of exposure or immunity beyond neutralising antibody alone. Finally, a large prospective study undertaken across areas of differing JE incidence could address the durability of immunity. However, such a study would be challenging and expensive to conduct. Both JE and dengue are seasonal in Nepal with JE cases peaking in August/September [46] and outbreaks of dengue in 2009 and 2010 occurring in September–October and August–December respectively [17]. Therefore, although our study period overlapped with these months, it is possible that prevalence of IgM antibodies would have been under-estimated compared with earlier in the year. Additionally, vaccine coverage was estimated to be high and similar in all three districts, independent confirmation of this is lacking. Finally, as the study was conducted in 2012 we cannot account for more recent changes in JE epidemiology. For example, cases are now seen in districts with higher elevation and Culex mosquitoes are found at higher elevations possibly due to the warming effect of climate change [47]. However, this does not affect our primary conclusion, that consideration of further JE vaccination campaigns in Udayapur is warranted. Alternatively, strengthening or extending the current routine immunisation schedule could obviate the need for such campaigns, at least in the paediatric population. Although it could be argued the reduction in JE incidence is lower in low/moderate risk compared to high risk districts following vaccination [22, 40], we nevertheless recommend consideration of further JE vaccination campaigns in areas of low/moderate JE incidence and in any district where JEV IgM is detected. However, the need for catch-up campaigns should be informed by data from outbreaks/reintroduction of the virus, and also by the extent of coverage of the routine immunisation program, across all age groups. If routine immunisation coverage were high enough, including in adults, catch up campaigns would no longer be required. Finally, our findings suggest that DENV has the potential to become established in Nepal, a finding of significant public health importance.

10.1371/journal.pcbi.1004518

A Pan-Cancer Catalogue of Cancer Driver Protein Interaction Interfaces

Despite their importance in maintaining the integrity of all cellular pathways, the role of mutations on protein-protein interaction (PPI) interfaces as cancer drivers has not been systematically studied. Here we analyzed the mutation patterns of the PPI interfaces from 10,028 proteins in a pan-cancer cohort of 5,989 tumors from 23 projects of The Cancer Genome Atlas (TCGA) to find interfaces enriched in somatic missense mutations. To that end we use e-Driver, an algorithm to analyze the mutation distribution of specific protein functional regions. We identified 103 PPI interfaces enriched in somatic cancer mutations. 32 of these interfaces are found in proteins coded by known cancer driver genes. The remaining 71 interfaces are found in proteins that have not been previously identified as cancer drivers even that, in most cases, there is an extensive literature suggesting they play an important role in cancer. Finally, we integrate these findings with clinical information to show how tumors apparently driven by the same gene have different behaviors, including patient outcomes, depending on which specific interfaces are mutated.

Until now, most efforts in cancer genomics have focused on identifying genes and pathways driving tumor development. Although this has been unquestionably a success, as evidenced by the fact that we now have an extensive catalogue of cancer driver genes and pathways, there is still a poor understanding of why patients with the same affected driver genes may have different disease outcomes or drug responses. This is precisely the aim of this work-to show how by considering proteins as multifunctional factories instead of monolithic black boxes, it is possible to identify novel cancer driver genes and propose molecular hypotheses to explain such heterogeneity. To that end we have mapped the mutation profiles of 5,989 cancer patients from TCGA to more than 10,000 protein structures, leading us to identify 103 protein interaction interfaces enriched in somatic mutations. Finally, we have integrated clinical annotations as well as proteomics data to show how tumors apparently driven by the same gene can display different behaviors, including patient outcomes, depending on which specific interfaces are mutated.

Cancer patients are extremely heterogeneous in their response to treatments and disease outcomes. The first step towards the understanding of this variability was the identification of the multitude of genes that cause cancer, the so-called cancer driver genes[1]. In that sense, the completion of The Cancer Genome Atlas (TCGA) and other large-scale cancer genomics projects was a watershed event, as it provided the critical mass of data needed to identify driver alterations in most types of cancers[2–15]. Moreover, cancer types that previously were thought to represent homogenous diseases were found to constitute different subtypes with different outcomes depending on the specific driver events in each patient[16]. Since the start of the TCGA project, the catalogue of cancer driver genes has increased and become more accurate[17] thanks not only to the data generated by the project itself, but also to the development of multiple, complimentary algorithms that search for cancer driver genes using different approaches. For example, some of these methods identify cancer drivers by searching for genes with higher than expected mutation rates[18,19], whereas others identify genes that tend to accumulate damaging mutations[20] or contain regions with an unusually high proportion of mutations[21,22]. Nevertheless, the catalogue of cancer driver genes is far from complete and, because of extreme mutation diversity, it is hard to extend it by simply increasing the size of the datasets[19]. A complementary approach towards that goal is to use methods that integrate cancer mutation profiles with other types of biological knowledge to increase the statistical power of the analysis. For example, by integrating the information on the mutation profile of cancer patients with biological networks we can identify pathways and protein complexes that are recurrently mutated in cancer and are, therefore, likely drivers[23]. Note that these complexes can only be identified as drivers when adding the signals of all the components, because each individual protein is rarely mutated and, thus, missed by standard gene-centric approaches. In fact, a recent paper describes the crucial role played by the network topology in the final phenotypic effect of apparently deleterious mutations[24]. Similarly, we can include information on the structure of the protein coded by genes being analyzed to check enrichment in cancer mutations in specific structural regions[22,25–27]. The underlying idea for this approach is that genes (and the proteins they encode) are not monolithic entities, but instead consist of different regions usually responsible for different functions. In that context, it is possible that a given protein acts as a driver only when a specific region is mutated. This idea can be exploited to identify cancer driver genes by analyzing the distribution of mutations within a gene and looking for regions with unusually high mutation rates. Such fine grain approaches are not only capable of finding novel cancer drivers, but they also can help explain some of the variability between tumors or cancer cell lines apparently driven by the same gene[28]. We have previously developed an algorithm, e-Driver, which exploits this feature to identify cancer driver genes based on linear annotations of biological regions such as protein domains[22]. Despite encouraging results, the algorithm still had some limitations, as many structural features, like protein interaction interfaces, may be discontinuous at the sequence level and, hence, can not be analyzed without explicit use of protein structure information. Here we introduce an extended version of e-Driver that uses information on three-dimensional structures of the mutated proteins to identify specific structural features. Then, the algorithm analyzes whether these features are enriched in cancer somatic mutations and, therefore, are candidate driver genes. While technically the analysis can be applied to any structural feature or region, here we focus our attention on protein-protein interaction (PPI) interfaces. Many known cancer driver genes are located in critical regions of the PPI network (interactome), usually in network hubs or bottlenecks[29], warranting closer investigation of interaction interfaces. Moreover, while it is known that many cancer somatic alterations alter PPI interfaces, either destroying existing interactions or creating new ones[30–32], this question has never been systematically analyzed across all known cancer somatic mutations with the specific goal of finding protein interfaces enriched in cancer mutations. Our analysis identified PPI interfaces enriched in somatic cancer mutations in a total of 103 genes (interface driver genes). Thirty-two of these are well-known cancer driver genes, which are strongly enriched in somatic missense mutations and were previously identified using other algorithms and approaches. We also found that interface driver genes have an unusually high number of interactions in all known PPI interaction network models. This effect is especially pronounced for the 32 known cancer drivers, not only when compared to the rest of the genes in the interactome, but also when compared to non-interface cancer driver genes. The role of the remaining 71 genes as cancer drivers will obviously have to be verified experimentally, though we find some attributes as well as literature links that, albeit indirectly, support the prediction of them being cancer drivers. Interestingly, many of the new putative “interface driver genes” are involved in the immune response, particularly in HLA-like and complement systems. The role of the immune system in cancer treatment and evolution is gaining increasing attention[33,34] and our results provide new details regarding which interactions seem to be most affected by somatic mutations. Finally we show how, in many cases, depending on which interface or protein region is altered, tumors apparently driven by the same cancer gene might have radically different behaviors and patient outcomes. We assembled a data set consisting of 5,989 tumors from 23 cancer types from The Cancer Genome Atlas[35] (Table A in S1 Table). The number of samples per tumor type ranged from 56 for uterine carcinosarcoma, to 975 for breast adenocarcinoma (Fig A in S1 Text). Consistent with previous reports[36], the average number of missense mutations per sample is highly variable among cancer types (Fig B in S1 Text), with melanoma having the highest (429 missense mutations per sample) and thyroid carcinoma the lowest (11 missense mutations per sample). We then compiled a list of currently known, high-confidence PPI interfaces using 18,651 protein structures downloaded from PDB (Online methods). In short, we defined a PPI interface as the set of residues from a given chain that are within 5 angstroms of any residue from a different chain in the same set of PDB coordinates (Fig 1a). We identified 122,326 different PPI interfaces between 70,199 PDB chains (Online methods). Finally, we used BLAST to map the residues from the PDB datasets to gene sequences in the ENSEMBL human genome. Overall we mapped the PDB coordinates to 11,154 protein isoforms in 10,028 different human genes. The mapping covers roughly 30% of the human proteome (measured per amino acid), with 6% of the proteome being mapped to at least one PPI interface. Mutations from all cancer datasets (n = 868,508) are distributed randomly across the proteome, with approximately 30% of mutations (n = 285,942) being in regions mapped to structures and around 6% in PPI interfaces (n = 67,174). However, in the case of known cancer driver genes [1,17], regions covered by structures have between 20% and 60% more missense mutations than expected by chance (Figs C and D in S1 Text), as we could map, on average, 40% of all mutations in known driver genes to a structurally solved region. This enrichment, while variable and dependent on the cancer type, is even higher, between two and three-fold, in regions involved in the PPI interfaces. For example, PPI interfaces from cancer driver genes in breast adenocarcinoma, glioblastoma, lower grade glioma rectal adenocarcinoma or uterine carcinosarcoma (Fig D in S1 Text) have more than three times as many mutations as would be expected by chance. These results strongly suggest that, indeed, mutations in PPI interfaces play key roles in carcinogenesis. To analyze the potential role of mutations in other proteins in cancer development, we used e-Driver to analyze individual PPI interfaces for all human proteins in each of the 23 individual cancer projects, as well as in the Pan-cancer dataset consisting of the combination of all of them. Briefly, e-Driver compares the observed number of mutations in a specific protein region with the expected value based on the ratio between the length of the given region and the length of the protein. We had previously used e-Driver to analyze the distribution of cancer somatic mutations in PFAM domains and intrinsically disordered regions and showed, for example, that different domains in the same protein can drive different types of cancer[22]. Here, we adapted e-Driver to analyze features that are discontinuous along the protein sequence, such as PPI interfaces identified from 3D structures of protein complexes. The whole process is exemplified in Fig 1 for PIK3R1 and its interaction interface with PIK3CA. We identified a total of 103 interface driver genes in either one of the cancer projects or in the Pan-cancer analysis (FDR < 0.01, Figs 2, 3 and Tables B-Z in S1 Table). There is significant overlap between the genes identified in this analysis and lists of known cancer genes. For example, 32 interface driver genes (31%) are included in either a list of high-confidence driver genes derived from previous analyses of TCGA data[17] or are part of the Cancer Gene Census[1] (p < 1e-10, odds ratio 9, when only taking into account genes with structurally solved regions). To further validate our findings we repeated the analysis using the PPI interfaces from Interactome3D[37]. While the pipeline used to define the interfaces in Interactome3D is different than the one that we used and the coverage is lower, the resulting list of driver interface genes is very similar (Figs E-G in S1 Text and Table AE in S1 Table), supporting the robustness of our analysis. Note that, as expected, there are many known cancer driver genes (n = 433) that are not picked up by our analysis. These genes might not have been identified either because their mechanism of action does not involve perturbing specific PPI interface, but also because we currently may not have a structure with a PPI interface to match to them and, thus, they were not included our analysis. For example, we can map only 40% of all the mutations in known driver genes to 3D structure models. Many of the remaining 60% of mutations might also be altering interactions, but we will not know that until we increase the structural coverage of the human proteome. Some of the driver interfaces identified here contain known cancer hotspots. For example, NFE2L2, a gene involved in cancer progression and drug resistance, is usually activated by mutations that disrupt the interaction with its repressor KEAP1. We mapped 36 mutations from NFE2L2 to the structure showing its interaction with its repressor KEAP1 (PDB 2FLU, shown in Fig 2). In agreement with previous observations[15], all but two of the mutations (94%) in NFE2L2 involve interface residues, likely disrupting the interaction between the two proteins and activating NFE2L2. Our results also highlight similarities and differences across related driver genes. For example, receptor tyrosine kinases, particularly members of the ERBB and FGFR families, are mutated in many cancers and frequently act as drivers. We found two ERBB proteins, ERBB2 and EGFR, among the interface driver genes. These two proteins are both strongly enriched in mutations in their dimerization interfaces, while the ligand-binding region is rarely mutated (Fig N in S1 Text). We also identified two proteins from the FGFR family: FGFR2 and FGFR3. Again, these two proteins have similar mutation profiles, with both proteins having most of their missense mutations in the region that interacts with the ligand, while leaving the dimerization interface intact. This, however, contrasts with the mutation pattern of the ERBB receptors, where, as we have explained, the ligand-binding region is rarely mutated. Since some of the most successful therapeutic antibodies against EGFR target the dimerization interface identified by our method, it is possible that antibodies against FGF receptors need to target the ligand-binding region in order to be successful[38]. Next we analyzed the 71 interface driver genes that are currently not classified as cancer drivers to determine their potential role in cancer. We found several results supporting our hypothesis that these genes can be cancer drivers. For example, many genes in our new-driver predictions are close network neighbors of known cancer drivers (Fig L and M in S1 Text). A subset of them can also be identified by other established methods (such as OncodriveFM[20] or OncodriveCLUST[21], Tables AC and AD in S1 Table). Furthermore, in several cases there is extensive literature and biological evidence supporting this hypothesis. This is the case, for example, for ARGHAP21. This protein is a small Rho GTPase that is suspected to play a role in epithelial-mesenchymal transition[39] and interacts, probably through the interface identified by e-Driver, with the known oncogene ARHGAP26. Another subset of these 71 potential new cancer driver genes has functions related to immunity. Given the growing body of evidence showing that the immune system plays a key role in cancer progression and patients outcomes[34,40], we analyzed these interfaces in more detail to try to find novel insights about the interplay between tumors and immune cells. For example, a recent pan-cancer analysis identified a subnetwork of proteins around HLA class I proteins as being recurrently mutated in cancer[23]. Our analysis also identified several antigen-presenting molecules as potential cancer drivers, including one class I (HLA-C), one class II (HLA-DRB1), and three HLA-like proteins (CD1C, CD1E and MR1). Note also that HLA-C has been recently identified as a likely driver in head and neck cancer[15]. Another interesting group of immune-related proteins identified in our analysis include several elements of the complement cascade (C3, C4B and C5) or complement regulators and inhibitors (CFHR4, CFI and CPAMD8). The complement molecules C3 and C4 have been previously associated with cancer progression and activation of PI3K signaling[41], whereas C5a is suspected to inhibit CD8 lymphocytes and natural killer (NK) cells, a subset of immune cells involved in the immune response towards tumors [42]. Our analysis with e-Driver not only supports the role of these proteins in cancer, but also suggests a specific mechanism for that role. Cancer driver genes are known to occupy critical positions in the interactome, as well as having more interactions and higher betweenness than the average gene[29]. Since our method identifies additional cancer driver genes, we hypothesized that they would have similar network positions as known cancer drivers. To test this hypothesis, we measured the degree and betweenness centrality of the interface driver genes in 16 different protein interaction and functional networks from 7 different sources (Fig 3 and Figs H-K in S1 Text). In all but one of the networks interface driver genes correlated with higher degrees even after correcting by confounding variables such as number of publications citing the gene[43], whether the gene had a PDB structure or not, or if the gene is a known driver (Table AF in S1 Table). Interface driver genes also correlated with higher betweenness in all but 4 of the networks, after correcting by all the aforementioned variables. Remarkably, while interface driver genes have similar network properties to known cancer drivers (Fig 3b, Table AI in S1 Table), genes that belong to both groups (i.e. known drivers with interfaces enriched in somatic mutations) are located in even more critical positions of the network (Fig 3c, Table AJ in S1 Table). These results are consistent with the hypothesis that the main driver mechanism of the interface driver genes, particularly those with strong driver signatures that are picked by multiple methods, is the alteration of the PPI interfaces and the interactions they mediate. Even if the genes that we identified have more mutations than expected in some of their PPI interfaces, there are tumor samples with mutations in other regions of the same genes. With that in mind, we wondered if there are consistent differences between cancer samples belonging to each of these two groups. To explore this issue, we first used proteomics data[44] and compared the expression levels of different proteins in tumors with mutations in the predicted driver interfaces to that of tumors with mutations in other regions of the same gene. To limit the impact of intrinsic tissue-variability in the protein expression levels, we limited our analysis to tissue-specific driver interfaces. Though we could not analyze most of the interfaces due to lack of statistical power (there were not enough samples with proteomics data in both groups), we did find some interface-specific protein changes. For example, glioblastoma samples with mutations in EGFR’s dimerization interface have higher levels of both EGFR and phosphorylated EGFR (Y992 and Y1173) proteins than patients with other EGFR mutations (Fig 4), suggesting that EGFR signaling is stronger in these patients. Note that these results also agree with the hypothesis that the main molecular mechanism driving cancer in these genes is the disruption of certain interactions, as cancer cells have different signaling levels depending on whether the gene is mutated in the identified driver interfaces or in another region. Another example of interface-specific protein expression changes comes from TP53 and its interface with SV40 (Fig 5). Note that this interface is the same as the one that TP53 uses to dimerize and bind to DNA (Fig 5b). Patients from eight different cancer types (bladder, breast, colon, endometrial, glioma, stomach, lung and head and neck) with mutations in this interface had significantly higher levels of TP53 protein than those with other or no TP53 mutations (Fig H in S1 Text). Moreover, patients with breast cancer had significantly worse outcomes (Fig 5d and Tables AG and AH in S1 Table), suggesting that these mutations are more aggressive than other mutations in TP53 and that maybe different therapeutic approaches are needed in these cases. The association was observed also after correcting by patient age, though because of insufficient statistics we were not able to test other potentially confounding variables such as ER status. Note that traditional gene-centric analyses or the previous version of e-Driver cannot find these differences among patient subpopulations. In this manuscript we have explored the role of missense mutations in PPI interfaces as cancer driver events using our e-Driver algorithm and the mutation profiles of 5,989 tumor samples from 23 different cancer types. Though the interaction interfaces of many cancer driver genes have been studied before[45,46], this is the first time that three-dimensional protein features, such as PPI interfaces, have been systematically used to identify driver genes across large cancer datasets. Previous large scale analyses are either limited to linear features[22,24,25,47], or are not based on known functional regions in three-dimensional structures but, instead, identify de novo three-dimensional clusters of mutations[26]. Our analysis identified several driver PPI interfaces in known cancer driver genes, such as TP53, HRAS, PIK3CA or EGFR, proving that our method can find relevant genes and that alteration of interaction interfaces is a common pathogenic mechanism of cancer somatic mutations. In fact, we found that cancer driver genes, as a group, are strongly enriched (over two-fold in most cancer types, and over three-fold in some cases) in mutations in their PPI interfaces. Moreover, there is a strong correlation between the fact that a cancer driver gene is recurrently mutated on its PPI interfaces and how critical it is to the stability of the interactome in terms of both number of interactions and network betweenness. We also identified a series of driver interfaces in genes that are currently not known as cancer drivers. Some of these genes interact with known cancer drivers or are implicated in key cancer functions, suggesting that they are, indeed relevant to carcinogenesis. Another group of potential cancer drivers identified here are proteins involved in the immune system. With the growing appreciation of the importance of malfunction of the immune system in allowing cancer progression[34,48,49], the immunity genes identified here can be used to develop a series of specific hypotheses of how modifications of their interaction patterns many modify immunity response to specific cancers. Analysis of all the genes with cancer driver interfaces identified in this work is ongoing in our labs, but in the meantime we provide a complete list of such genes in the S1 Table and S1 Text, as well as in our on-line resource Cancer3D[50], inviting other groups to analyze, confirm or refute our predictions. It is important to note that the analysis presented here was limited to high quality interfaces, predicted either from solved structures or from high quality homology models. However, about 70% of the human proteome currently has no high quality structural coverage. This fraction of the proteome includes both low complexity or disordered regions, and protein regions without reliable templates to model their 3D structures. Also, structures of many complexes are still unknown. In these cases, even if we know the structures of the subunits, we cannot define the PPI interfaces and these proteins were not included in our analysis. Finally, even though we did not explore this issue here, there are other mutations that can have an impact on PPI interfaces, such as in-frame indels or silent mutations[51]. Therefore, the results presented here represent only the tip of the iceberg of what can be achieved by including structural data in the analysis of cancer mutation profiles. We expect that our method will improve not only as more cancer genomes are added to existing repositories (increasing the statistical power of the analysis), but also as the structural coverage of the human proteome increases. We expect such increase to come from both new experimentally determined structures in public databases and the use of better modeling tools[52,53]. Another important results of our analysis is that we found that tumors with mutations in the same driver gene can have surprisingly different behavior and outcomes depending on the specific PPI interface affected by the mutation. This adds to a growing body of evidence suggesting that the current gene-centric paradigm in biology, while successful in some cases, will probably not be enough to explain the complex genotype-phenotype relationships underlying a vast array of complex traits[54–59]. In the case of cancer, for example, it is known that the two most common mutations in PIK3CA, E545K and H1047L, contribute to carcinogenesis through different mechanisms[45]. The same is true for different types of mutations in KRAS[60] or, as we have shown here, for mutations in EGFR or TP53. All of the above suggests that in order to predict the outcome of a patient or the best treatment option we will need to have more detailed knowledge about the consequences of a specific mutation than just the identity of the cancer driver gene where it is located. Such increase in detail and knowledge should include, in the case of missense mutations, not only information about the protein domain or PPI interface of the gene being altered, but also data about mutations in other regions of the network, as these can also influence the phenotype of a driver gene through synthetic interactions[61]. Finally, one must keep in mind that, while this work has focused on the analysis and interpretation of missense mutations, there are many other types of variations that can act as cancer drivers and have a significant impact in the outcomes of cancer patients. Examples include promoter mutations[62], copy number variations[63], silent mutations[51] or small insertions or deletions. It is likely that different types of variations of the same gene will have different consequences and, therefore, could need different therapeutic approaches. A clear example of this phenomenon are TP53-driven tumors. As we show here, it is likely that patients with missense mutations in this gene have different outcomes depending on the specific region of TP53 that is mutated, suggesting that they might need different treatments. Nevertheless, a patient whose tumor is driven by a TP53 copy-number loss might benefit from yet another therapeutic approach that would not help any of the above[64]. Therefore, in order to identify the optimal treatment of each patient we will need to integrate and properly analyze all molecular consequences of the different types of mutations present in its tumor. All the supplementary information, the raw data and the algorithms used in this manuscript, as well as the results presented, can be downloaded from http://github.com/eduardporta/e-Driver. The link to the Dropbox folder containing all the raw data can be found in the README.md file. All the statistical calculations were done using R 3.1.0. All figures have been generated using the R package “ggplot2”. We downloaded level 3 mutation data from the TCGA data portal (https://tcga-data.nci.nih.gov) for 5,989 tumor samples that belong to 23 different cancer types (Table A in S1 Table). We then used the Variant Effect Predictor tool to derive the consequences of each mutation in the different protein isoforms where it mapped[65]. We used gene and protein annotations from ENSEMBL version 72. We identified a total of 868,508 missense mutations in 19,196 proteins. Note that we only analyzed the longest isoform of each gene in order to minimize problems related to multiple testing. We identified 18,651 protein structures with multiple chains in PDB (as of May 2014). Then, we analyzed all such structures to find the residues implicated in PPI interfaces. To that end, we defined a protein-protein interface in a chain as all the residues with a heavy atom within 5 angstroms of another heavy atom from a different chain, an intermediate value between the 4 and 6 angstroms seen in other references[31,66]. If a chain was in contact with multiple other chains, we defined a different interface for each chain-chain pair. Note that any specific interface does not have to be linear in sequence and that the same residue can be involved in multiple interfaces from different structures. The complete dataset containing all the PPI structures and models from Interactome3D was downloaded on April 30th 2015. As described previously, protein-protein interfaces are defined as those residues in a chain whose atomic distance falls within 5 angstroms from the partner chain. Since Interactome3D uses Uniprot protein sequences while we use ENSEMBL, we had to map the coordinates from one to the other. In order to do that we compared the two sets of sequences and kept only those interfaces in Uniprot proteins whose sequence matched exactly a protein from ENSEMBL. While this reduced significantly the number of potential interfaces from Interactome3D, from 26,383 interfaces to 11,169, it ensured that the results obtained with each dataset would be comparable. The mapping between ENSEMBL and PDB is the same as the one used in Cancer3D. Briefly, we queried the full PDB (March 2014), including non-human proteins, with every protein from ENSEMBL using BLAST. Every time we identified a PDB-ENSEMBL pair with an e-value below 1e-6 we used the BLAST output to map the residues from the ENSEMBL sequence to the PDB structure[50]. We used e-Driver[22] to identify interfaces that are enriched in somatic missense mutations. The algorithm calculates the statistical significance of deviation from the null hypothesis that the mutations are distributed randomly across the protein using a right-sided binomial test: P(MR, MT)= (MTMR) (PMutReg)MR(1−PMutReg)MT−MR Where “PMutReg” is the ratio between the number of residues involved in the interface and the number of aminoacids in the entire protein, “MR” is the number of mutations in the interface and “MT” is the total number of mutations in the protein. Since it is possible that only a fraction of the protein is covered by the structure, we adjusted the algorithm to limit all the parameters to the structure-mapped region of the protein (for example “MT” refers to the total number of mutations in the region of the protein covered by the specific structure being analyzed, not the absolute total of mutations in the protein). The final step consists in correcting all the p values for multiple testing using the Benjamini-Hochberg algorithm. We considered as positives all of the interfaces with a q value below 0.01. Given the large number of available human PPI networks and the variability in their quality, we decided to use 16 different networks from 7 different sources (Figs H-K in S1 Text): HPRD[67], Biogrid[68], STRING[69], HumNet[70], PSICQUIC[71], one PPI derived from unbiased experiments as well as curated literature[43] and another network derived from in silico predictions of PPI based on structures[72] (which we will call “Kotlyar” from this moment). Three networks, STRING, HumNet and Kotlyar have scores that approximately correlate with the probability of the interaction being true. Therefore, we decided to divide these networks in different subsets, by selecting only those interactions above a certain threshold (Figs H-J in S1 Text). We then calculated the different protein properties (node degree and node and edge betweenness) in each network using the R package “iGraph”. We identified several potentially confounding factors that could explain differences in network properties of the different proteins. For example, there is a clear bias introduced by the fact that we can only analyze proteins with structurally covered regions (either by direct experimental structures or homology modeling). Another potential confounding variable is the number of publications of each protein, as it seems to correlate with the number of interactions[43]. In order to do that, we used the e-tools from Pubmed to retrieve the number of papers mentioning each gene symbol in the title or the abstract. We also took into account whether the gene is a known cancer driver or not, as cancer driver genes are also known to have high degree and betweenness centrality[73]. Finally, we fitted the aforementioned variables, as well as whether the gene was an interface driver or not, into a generalized additive model (using the R package “gam”) and calculated the correlation of each variable with the degree or betweenness centrality in every network. We measured the distance between known cancer driver genes and genes identified by e-Driver not yet known to play a role in carcinogenesis in the different networks. To that end we calculated the distance between both groups of genes using the random walk with restart (RWR) algorithm. The random walk on graphs is defined as an iterative walker’s transition from its current node to a randomly selected neighbor starting at a given source node. This algorithm has been extensively used for the predictions of disease-associated genes[74] as well as the analysis of cancer genomes[75–77]. It also allows the restart of the walk from the source nodes at each time with probability “r”. For a detailed explanation of the effects of different “r” values see Fig L and Fig M in S1 Text. The random walk is described by the equation: pt+1 = (1-r)*W*pt+r*p0 Where W is a column-normalized adjacency matrix of the graph, pt is a vector in which the i-th element holds the probability of being at node i at time t and p0 is the initial probability vector. This vector has value 0 if the gene is not a known driver, and value 1/D if the gene is a known driver, where D is the number of known driver genes in the network. The algorithm iterates the equation until the L1 norm between pt and pt+1 is less than 10−6. Then, we added the probabilities of all the candidate driver genes identified by e-Driver and compared it to 10,000 groups with the same number of random genes to calculate empirical right-sided p-values. We downloaded level 3 clinical and protein expression data, whenever it was available, from the TCGA data portal. Then, for each statistically significant interface, we classified each sample into one of three groups: samples with mutations in the interface, samples with mutations in other regions of the same protein, and samples with no mutations in that protein. Finally, in the case of proteomics data, we used a two-sided Wilcoxon test to identify proteins with statistically significant differences between the first group and the other two. As for the clinical data, we used the Cox proportional hazards model from the R package “survival” to estimate whether mutation of a specific interface was a predictive feature for survival (p < 0.01) after correcting by age.

10.1371/journal.pbio.2000942

The Global Distribution and Drivers of Alien Bird Species Richness

Alien species are a major component of human-induced environmental change. Variation in the numbers of alien species found in different areas is likely to depend on a combination of anthropogenic and environmental factors, with anthropogenic factors affecting the number of species introduced to new locations, and when, and environmental factors influencing how many species are able to persist there. However, global spatial and temporal variation in the drivers of alien introduction and species richness remain poorly understood. Here, we analyse an extensive new database of alien birds to explore what determines the global distribution of alien species richness for an entire taxonomic class. We demonstrate that the locations of origin and introduction of alien birds, and their identities, were initially driven largely by European (mainly British) colonialism. However, recent introductions are a wider phenomenon, involving more species and countries, and driven in part by increasing economic activity. We find that, globally, alien bird species richness is currently highest at midlatitudes and is strongly determined by anthropogenic effects, most notably the number of species introduced (i.e., “colonisation pressure”). Nevertheless, environmental drivers are also important, with native and alien species richness being strongly and consistently positively associated. Our results demonstrate that colonisation pressure is key to understanding alien species richness, show that areas of high native species richness are not resistant to colonisation by alien species at the global scale, and emphasise the likely ongoing threats to global environments from introductions of species.

The introduction of alien species is one of the primary ways in which human actions are changing the environment. Alien species have been responsible for numerous global and local extinctions and are eroding the uniqueness of many natural environments. There is thus a basic need to understand which areas end up with more alien species. Here, we use a major new global database on the distribution of alien birds to show, first, how patterns in the number of species introduced to a location (colonisation pressure) have changed over time. We show that historical introductions were driven largely by European, and especially British, colonialism. However, the rate of bird introductions is increasing, with shifts in the locations of origin and introduction of species probably driven by the cage bird trade. We then combine information on where bird species have been introduced with a global map of alien bird species richness to identify the main drivers of richness. We show that colonisation pressure is the strongest predictor of alien bird species richness, but that there are other anthropogenic and environmental drivers. Most notably, once colonisation pressure has been accounted for, alien bird species richness is higher in areas where native bird species richness is higher.

The number of species naturally inhabiting a location (native species richness [NSR]) is ultimately driven by the combined processes of speciation, extinction, and immigration, and proximately by the suite of environmental, ecological, historical, and evolutionary factors that determine the interplay of these processes [1]. An important feature of the Anthropocene is the extent to which human activities have enhanced immigration [2], such that species are being intentionally or accidentally transported and introduced to areas well beyond the biogeographic barriers that normally prevent their spread, and at unprecedentedly high rates [3]. These species (here termed alien) may establish viable populations and subsequently spread in their new locations (a process termed invasion) [4], altering local and regional-scale species richness as well as patterns of species turnover across areas (i.e., alpha, gamma, and beta diversity, respectively) [5]. Alien species can adversely affect the native biota, driving populations and species to extinction [6], altering ecosystem function [7], and negatively impacting social and economic activities [8], providing significant impetus to understand drivers of the invasion process. The starting point of the process of invasion by alien species is the translocation of individuals by human activities. However, the extent to which the resulting worldwide distribution of alien species richness (ASR) is entirely a consequence of human activities is unknown. ASR is the end point of a multistage process that involves the following: the transportation and introduction of species to areas where they do not naturally occur; the establishment of viable populations around the point of introduction; and the spread of established populations across the wider landscape [4]. Clearly, alien species richness is likely to be highly dependent in the first instance on the number of species relocated to a given area or location by human activities. This is known as colonisation pressure, which is calculated simply as the sum of the number of alien species introduced to a defined location (e.g., country, state, island, 1° grid cell) over some period of time (typically the full period over which introductions have occurred, although sometimes subsets of this period are specified), some subset of which will succeed in establishing an exotic population [9]. The ASR of a given region (or area) at a given point in time will therefore equal colonisation pressure minus the number of those introductions that fail to establish, but plus the number of alien species that have spread into the region or area having been introduced elsewhere. Alien species richness is therefore expected a priori to be a positive function of colonisation pressure [9]. It is also likely to depend on other anthropogenic factors such as the proximity of an area to locations of introduction and the amount of time that introduced species have had to establish and spread (S1 Table). Alien species richness may also be influenced to some degree by the same natural processes that determine native species richness [10]. Natural processes are expected to influence which introduced populations succeed (or fail) in establishing, and which species spread into novel locations from other introduction sites. These components of alien species richness are likely to be functions of the abiotic environment, biotic interactions (primarily with native species because typically ASR < NSR; S1 Table), and stochasticity [11,12]. As a result, they will demonstrate structured spatial variation. However, the relative interplay of these different natural processes on alien species richness remains poorly known [3]. The primary reason is that for most alien taxa, we have very limited information on colonisation pressure, meaning that it is impossible to determine accurately the form of its expected effect on alien species richness. This omission is a fundamental barrier to a mechanistic understanding of other drivers of alien species richness. For example, an observed negative association between native and alien species richness could be evidence of biotic resistance to the spread of alien species (S1 Table), but equally could simply reflect higher colonisation pressure in areas of lower native species richness. Thus, determinants of variation in alien species richness cannot meaningfully be analysed without knowledge of variation in colonisation pressure. Here, we present what is, to our knowledge, the first global analysis of the determinants of alien species richness that takes into account colonisation pressure. Our study uses birds (class Aves) as a model system and is underpinned by a database comprising 27,723 distribution records for 971 alien bird species worldwide (see Methods). Uniquely, for this species group, the high quality of our collated data means that our database includes relatively detailed information on where species have been introduced (both successes and failures). This data quality allowed us to quantify spatial and temporal variation in the key variable of colonisation pressure at a range of spatial and temporal scales and to test factors that underpin this variation. We show that the locations of origin and introduction, and the identities, of bird species introduced in the historical period (1500–1903 AD) were largely determined by planned liberations associated with European (mainly British) colonial expansion, whereas recent introductions (1983–2000 AD) were largely unplanned and associated with increased levels of economic activity. We then combine information on colonisation pressure with abiotic, biotic, and anthropogenic variables to identify which of the hypothesised determinants influence worldwide variation in alien species richness (S1 and S2 Tables). We show that global patterns of alien species richness in birds are determined by both human and biotic factors. Our results reveal that attempts to model alien species richness at regional scales without data on colonisation pressure have problems inferring causality due to uncertainty over whether drivers are acting on colonisation pressure or alien species richness. We censored introduction data to the period 1500–2000 AD, because 1500 is a standard cutoff point for studying biological invasions [13,14], and introductions occurring after the year 2000 may not yet have entered the literature. Our analysis of introduction drivers covers a total of 3,661 alien bird introduction records (first known occurrence of a given species in a given country), which were distributed over time as shown in Fig 1. These 3,661 records were split into four quartiles, on the basis of introduction date, for analysis. The rate and drivers of alien bird introductions altered markedly over the period 1500–2000 AD. The first quartile of bird introductions ordered by date spans 403 y (1500–1903 AD; see Methods), during which a total of 245 species were introduced to 167 countries (922 introductions in total). In contrast, the fourth quartile covers just 17 y (1983–2000 AD), with a total of 324 species introduced to 235 countries (935 introductions in total). The rate increased sharply around the middle of the nineteenth century, such that 69.3% of first-quartile introductions were after 1850 (Fig 1). This increase coincides with the founding of the first Acclimatisation Societies [15,16], which largely drove introductions in this period. These organisations were specifically aimed at promoting introductions of beneficial species, such as game, and were especially prevalent in the then British territories [15,16]. Thus, families with the most introduced species are the main game-bird families—Anatidae (duck, geese, swans), Phasianidae (pheasants, grouse, partridge) and Columbidae (pigeons, doves) (S3 Table). Species in the first quartile of introductions were more likely to originate from Europe (Fig 2A, S4 Table). Early introductions were concentrated in fewer countries than expected (observed: 167; expected median [range]: 212 [189–231], p < 0.0002; see Methods), and these countries were more likely to be constituents of the then British Empire (Figs 2B and 3A). The rate of bird introductions increased again from around the middle of the twentieth century, such that half of all known bird introductions occurred after 1956 (Fig 1). Recent introductions (1983–2000 AD) can be better explained by globalisation and economic growth [17]; the number of introductions to a country in the fourth quartile is positively correlated with its per capita GDP (Fig 3D), but not whether or not the introduction was to a former British colony (Fig 3C). The sources of introduced species have shifted significantly over the twentieth century, away from Europe and to the Indian subcontinent, Indochina, and sub-Saharan Africa (Fig 2C, S1 Fig, S4 Table). Three of the four bird families with the most introduced species in this period include popular cage birds—Psittacidae (parrots), Estrildidae (soft-bill finches), Sturnidae, (mynas, starlings) (S3 Table)—reflecting a shift towards unplanned introductions (or releases) of species in the cage and pet trade [18]. Introductions are now spread across more countries than in the first quartile, and more countries than expected by chance (observed: 235; expected median [range]: 214 [194–235], p = 0.0002). The locations of introduction of alien birds in the most recent quartile reflect the geographic focus of the bird trade [18], with notable hotspots in the Far East (e.g., Singapore, Hong Kong, Taiwan), the Near East, Spain, and Florida (Fig 2C and 2D). The establishment (or failure) and subsequent spatial spread of populations of introduced bird species modifies the global patterns of introductions (Fig 2B and 2D, S1 Fig) to give rise to the current global distribution of alien bird species richness (Fig 4). As expected, colonisation pressure is a strong positive determinant of bird alien species richness and is, indeed, the variable most closely associated with alien species richness in our models (Table 1, S5 and S6 Tables, S2 Fig). As a result, the global map of alien bird species distributions shows that regions with high alien species richness tend to be located in temperate regions at midlatitudes (Fig 4); these are regions where former British colonies, rapidly developing countries, and countries with high per capita GDP are located, and where colonisation pressure has been concomitantly high (Fig 2B and 2D). Notable hotspots of alien bird species richness include areas in the United States (including the Hawaiian Islands), Caribbean, United Kingdom, Japan, Taiwan, Hong Kong, New Zealand, Australia, Persian Gulf States, and the Mascarene Islands. Alien species richness is also higher in areas with a longer history of introductions, with a significant positive effect of the number of years since the first bird introduction to a region, and shows a negative relationship with distance to a historic port (Table 1, S5 and S6 Tables, S2 Fig). These effects are independent of a measure of general human activity, the human footprint index (Table 1, S5 and S6 Tables, S2 Fig). The alien species richness of birds is not only a function of human historical factors but also shows significant imprints of the natural environment. Notably, there is a strong positive relationship between alien species richness and native species richness at the global scale (Table 1, S5 and S6 Tables, S2 Fig). Within the global regions to which alien bird species have been introduced or spread, areas with fewer native bird species tend also to house fewer aliens, with alien species richness peaking at mid to high levels of native species richness (S2 Fig). This may indicate that whatever drives native species richness may also drive alien species richness. Native species richness for a wide range of terrestrial taxa, including birds, tends to be higher in warmer, wetter regions [19], perhaps because these areas have higher levels of energy availability [20,21] or less physiologically stressful environments [22]. Native bird richness also tends to be higher in regions with greater elevational ranges [19]. Univariate analysis suggests that alien species richness also follows these trends, as regions with medium to high median temperatures, medium to high levels of precipitation, and medium to high elevational ranges are home to more alien bird species (S5 Table). Nevertheless, these abiotic effects are outperformed by native species richness as a determinant of alien species richness in our models (Table 1). Cross-validation, holding out individual biogeographic realms (see Methods), confirms that the primary predictors in our global models are robust to sub-sampling of the data: colonisation pressure or native species richness was the strongest effect in all of the realm models (S6 Table), supporting the general influence of both anthropogenic and environmental drivers of species establishment and spread. Conversely, cross-validation reveals that time since introduction only enters models for alien species richness for the Indo-Malayan and Palearctic realms. A range of additional environmental variables are also retained in the most likely realm-level models, but their effects are always weak relative to those of colonisation pressure and native species richness (S6 Table). We repeated our analyses excluding colonisation pressure to test whether lack of knowledge of this variable results in different conclusions about the determinants of alien species richness. This analysis shows that the influences of time since introduction and distance to historic port on alien species richness strengthen substantially (S7 Table). The resulting minimum adequate model additionally includes a positive effect of annual precipitation on alien species richness, suggesting that alien bird species richness is higher in areas of medium to high annual precipitation, as also tends to be the case for native species richness [19]. However, there is no evidence for direct abiotic effects when colonisation pressure is included (Table 1, S5 and S6 Tables). The minimum adequate model excluding colonisation pressure is also a substantially worse fit to the data (ΔwAIC = 1390.3 relative to the best model with colonisation pressure). Biological invasions by alien species are one of the primary ways in which human activities are causing global environmental change, providing a strong incentive to understand the invasion process. The annual rate of first records of alien species worldwide has increased more or less constantly over the last 200 y and now averages >1.5 new records per day [23]. Nevertheless, the extent to which global variation in the richness of alien species is entirely the result of variation in human activities, or whether it also includes the imprint of the same sorts of natural processes that determine the richness of native species, is impossible to distinguish in the absence of studies that incorporate the effect of where species have been introduced—i.e., colonisation pressure [9]. Our global database on alien birds has enabled us to show, first, what underlies patterns in the global distribution of colonisation pressure, and second, how incorporating this information into analyses of alien species richness allows the relative influence of anthropogenic and natural drivers to be identified. The drivers and locations of alien bird introductions have changed over the last 500 y. Early introductions were primarily planned with the aim of establishing new populations of beneficial species, although many bird species were also introduced for purposes of ornamentation [15,16]. The identities, sources, and introduction locations of the species concerned reflect these motivations, and the fact that early translocations (1500–1903 AD) were largely a European, and especially a British, endeavour. However, changes in the attitudes and legislation in the countries responsible for historical introductions have led to a reduction in colonisation pressure in those locations [18,24]. Recent introductions are higher on average in countries with higher per capita GDP, which is related to a country’s volume of trade as well as the disposable income of its populace. The identities, sources, and recipient locations of the species in recent introductions identify their origins primarily in unplanned releases of species in the cage bird trade, notably many parrots, finches, and mynas (S3 Table). The global locations with high current colonisation pressure (Fig 2D) reflect the present-day geographic foci of the bird trade, itself associated with increasing affluence and disposable income worldwide [17,25], but especially in rapidly developing countries in the Near and Far East [26,27]. Bird keeping is a popular hobby and an expression of social status in many countries, driving a growing and lucrative pet trade, especially in Eastern Asia, the Near East, and parts of Europe (Fig 2D) [28]. For example, bird introductions to Spain and Portugal exhibited a major increase in the early 1980s coinciding with economic upsurges in these countries and increasing volumes of trade [29]. Nevertheless, the relationship between GDP and number of introductions in recent years is triangular in form (Fig 3D); high GDP does not necessarily lead to higher colonisation pressure. Countries bucking this trend mainly include those in continental Europe, and notably also New Zealand, reflecting the changes in attitudes and legislation discussed above. The identities of bird species involved in introductions have changed markedly over time (S3 and S4 Tables) but remain predominantly Old World in origin (c.f. Fig 2A and 2C, S1 Fig). The native ranges of recently introduced birds (Fig 2C) overlap the species richness hotspots of the Himalayan arc, tropical southeast Asia, and east Africa [30], but the distribution of red in Fig 2C suggests an overrepresentation of birds from open land biomes and an underrepresentation of forest birds. This could be because tropical forest species tend to have smaller geographic ranges, while bird species with larger geographic ranges are more likely to be introduced [3]. The biogeographic realm with the highest native bird species richness is the Neotropics [30], but this region remains relatively underexploited as a source for alien birds. A relatively high proportion of narrow-ranged species combined with the region’s relative distance from the main cage bird markets of the Near and Far East may be responsible for this situation. However, that may change as the bird trade exhausts supplies from other parts of the world [28], especially as the New World is home to many brightly coloured birds, adept songsters, and parrots, that would likely be highly marketable species. New World species such as Aratinga jandaya, Mimus polyglottos, and Tangara seledon have already been recorded for sale in Taiwanese bird shops [31]. It may also change if the demand for cage birds in Neotropical regions continues to grow. Fifty-nine bird species have already been recorded as introduced in Brazil, primarily from other parts of the Neotropics [32]. The historical drivers of colonisation pressure largely determine modern variation in alien species richness (Table 1), reflecting the direct effects of human activities on species distributions. The best global model (Table 1) explains 98.6% of the variation in bird alien species richness, accurately predicting alien species richness to c.2 species per grid cell (S6 Table), with colonisation pressure having the strongest effect (Table 1, S5 Table). Holdout cross-validation shows that colonisation pressure also explains high variation in alien species richness when removing four out of six biogeographic realms (S6 Table). Because (as previously noted) ASR is equal to colonisation pressure minus the number of introduced species that fail to establish, but plus the number of alien species that spread in from other areas, alien species richness is expected to be a positive function of colonisation pressure. Nevertheless, this result is not a given: patterns of failure and spread could have acted to erase the initial spatial variation in human introductions. The clear signature of colonisation pressure is therefore a cause for concern given the negative impacts of some alien species [5,6,7] and the rate at which colonisation pressure is increasing (Fig 1) [23]. Alien species richness is also higher where bird species have had more time to establish and spread following introduction. The early trade in alien birds was conducted largely by sea, and so areas proximal to historical sources of imported birds have higher alien species richness as a result. These results reinforce the need for greater controls on species introductions to prevent invasions, such as the 1993 Biosecurity Act in New Zealand or the recent EU Regulation No 1143/2014 on Invasive Alien Species. Also of concern is that the effects of time since first introduction and distance from a historic port imply an invasion debt [24], as modern sources of alien species are yet to obscure the imprint of historical processes. Interestingly, however, these effects are less consistent within biogeographic realms (S6 Table), suggesting that their global effect may be largely due to the different temporal and spatial invasion histories of different realms (Fig 2B). Nevertheless, alien species richness is not just a function of human activities but also shows the imprint of natural processes. Notably, we find a consistent and strong positive association between alien species richness and native species richness, both globally (Table 1) and across realms (S6 Table); this runs counter to hypotheses of biotic resistance whereby diverse native assemblages resist the spread of alien species, but rather confirms previous analyses at island and continental scales showing that “the rich get richer” [33,34]. Previous studies have not controlled for colonisation pressure, opening the possibility that this phenomenon was purely a consequence of where species were introduced. Our results confirm that this association is independent of colonisation pressure. Most introductions have occurred at mid-latitudes (Fig 2B and 2D), which tend to have intermediate levels of native species richness [30], yet within the regions to which alien bird species have been introduced or spread, areas richer in native bird species tend also to be richer in aliens. Positive richness associations at broad scales might imply that native and alien bird species are responding to similar environmental variables [33], but what these variables are remains an important and open question. We also cannot rule out the alternative explanation that native and alien species are responding to different but spatially coincident processes. For example, alien bird richness may reflect some unmeasured aspect of human behaviour while humans tend to live at higher densities in areas rich in native bird species [35]. Nevertheless, our results show that, as a predictor of alien species richness, native species richness outperforms a variety of abiotic factors known to relate to bird species richness, including temperature, precipitation, and elevation (Table 1, S5 and S6 Tables). Repeating the analysis as if data on colonisation pressure were not available results in a final model that is a substantially poorer fit to the data, which identifies stronger effects of time since first introduction and distance to historic port, and has additional effects of precipitation on alien species richness (c.f. Table 1 and S7 Table). Because alien species richness is expected to be a positive function of colonisation pressure, analyses of alien species richness carried out without data on this will assign higher importance to variables that are surrogates for colonisation pressure and will also give higher weighting to variables that may influence the number of introduced species that fail to establish, as well as the number of alien species that spread in from other areas. Upweighted variables may also influence alien species richness, as is the case for time since first introduction and distance to historic port in our analyses, but might include variables that would not otherwise rate as important, as is the case for precipitation. This model does identify that human activities matter to global variation in alien species richness; nevertheless, we can be more specific that the key activity in this regard is colonisation pressure. While the distribution of alien species richness and, in particular, the dearth of alien bird species in the arid zones (Fig 4) would indeed seem to imply that abiotic factors may be an important influence on where alien species richness is high or low, there is no evidence for direct abiotic effects when colonisation pressure is included (Table 1, S5 and S6 Tables). Instead, at least for birds, there were fewer attempts to establish alien species in areas of low precipitation. Abiotic factors determine the fundamental niche of a species, delineating the set of environmental factors that determine where a species can and cannot maintain a population. Therefore, in some cases, the physical environment will influence the success or failure of alien species introductions and whether or not alien species established elsewhere can spread into a region [36]. Nevertheless, there are many areas of the world where abiotic conditions are suitable for species but where they are not naturally found because of biotic resistance (e.g., the presence of competitors) or dispersal limitation. Human-mediated translocation overcomes the second of these constraints, while the positive relationship between alien and native species richness suggests that the first may not be as strong a bar as has sometimes been hypothesised (S1 Table). The relatively weak influence of abiotic variables in our analyses may be because the likelihood that an area is within the fundamental niche of an alien is higher when it is within the fundamental niches of more native species, and, hence, the influence of the environment is subsumed within the effect of native species richness. Either way, our results highlight the need for further investigation of how environmental suitability and native richness might interact to determine the potential of different areas to accept alien species. Environmental factors are unlikely to have a major impact on colonisation pressure, which has been largely driven by human activities, but will affect the number of introduced species that fail to establish and the number of alien species that spread into a region having been introduced elsewhere. The failure of individual introductions has been shown to depend on the number of individuals introduced (propagule pressure), traits of the species involved, and characteristics of the environment where the introduction took place [12,37]. For birds, establishment failure decreases as propagule pressure increases up to c.50–100 individuals, beyond which it is more strongly influenced by environmental suitability [12]. Our analysis does not consider which species fail or spread, only the number, but propagule pressure may affect this if it is systematically lower in some regions. In fact, propagule pressure is likely to be a positive function of colonisation pressure, especially for unplanned introductions [9], which may influence the precise form of the relationships we observe but will not alter the fact that alien species richness is a positive function of colonisation pressure. Previous global studies have identified the donor and recipient regions of alien plants and regions with high alien plant richness [38] and shown that the distributions of gastropods after human transport are primarily explained by the prevailing climate and, to a smaller extent, by distance and trade relationships [39]. However, these studies have not had access to information on colonisation pressure, which we have shown to display interesting and informative spatial and temporal variation, and to be the key determinant of variation in the richness of alien bird species. Moreover, without information on colonisation pressure, statistical models are substantially poorer fits to data on alien species richness and differ in the factors they identify as richness drivers and the weights they assign to factors. Notably, alien species richness is not simply a consequence of higher overall levels of human activity (higher population, greater habitat disturbance, and increased access) in some regions, as it is independent of human footprint index, which quantifies this (S1 Table). Information on colonisation pressure is rarely available, but without it, erroneous conclusions regarding the determinants of alien species richness are likely. Including colonisation pressure allows us to understand, first, how changing histories of socioeconomic drivers affect the distribution and extent of bird population introductions (Fig 2), and then that alien species richness is a consequence of a combination of anthropogenic factors and biotic acceptance of aliens into areas already rich in native bird species. Our analysis focuses on human-mediated introductions of bird species to locations outside their native geographic range. Our database (the Global Avian Invasions Atlas, or GAVIA; the acronym refers to the scientific name of the bird genus that includes divers or loons, although none of these species have ever been introduced) comprises 27,723 distribution records for 971 bird species for which there is some evidence of translocation outside their native range (i.e., aliens; see status categories below), based on almost 700 published references and substantial unpublished information derived from consultation with organisations and experts worldwide. Each entry in GAVIA corresponds to a single record of a single species recorded as introduced and alien in a specific location as published in a single reference. Records therefore correspond to descriptions or depictions of part or all of the alien geographic distribution of a species, rather than point occurrence records. For those records with a sufficient level of detail (for example, a sub-state or specific location of introduction was provided), or where the original reference contained a distribution map, vector range maps were created to represent the species’ alien ranges at different points in time. The full bird taxonomy used in GAVIA was that used by the International Union for the Conservation of Nature (IUCN) Red List of Threatened Species (www.iucnredlist.org, downloaded August 2010). The country and regional designations used in GAVIA were downloaded from the Global Administrative Areas (GADM) database (www.gadm.org, downloaded August 2010). GAVIA contains records for all statuses of introduction event, from those that were unsuccessful through to those that are deemed to be established, i.e., which are thought on the basis of population trends or expert local opinion to have a self-sustaining population in the area of introduction. More recent introductions are likely to be better catalogued, but the fact that bird distributions are very well known and that many of the historical introductions were planned and documented in detail [15,16] means that our data are likely to be of generally high quality [23]. The full GAVIA database is archived on Figshare at http://dx.doi.org/10.6084/m9.figshare.4234850 and described in Dyer et al. [40]; the specific data used in this paper are provided in supplementary files S1–S5 Data, as identified below. Six categories were used in GAVIA to describe the invasive status of each alien species: (1) Established: The species has formed self-sustaining populations in the area of introduction; (2) Breeding: The species is known to be breeding/have bred in the area of introduction, but is not yet thought to be self-sustaining; (3) Unsuccessful: The species has not formed self-sustaining populations (casual, incidental); (4) Died Out: The species was once established but has now completely died out in the area of introduction; (5) Extirpated: The species was once established but has now been actively eradicated in the area of introduction; (6) Unknown: The status of the species in the area of introduction is not known and further clarification is necessary to determine which of the other five categories is appropriate. “Established species” refers to aliens in category 1 only, and “introduced species” to aliens in any category. We use GAVIA to conduct two distinct analyses. First, we use data on the first introduction records of alien bird species between 1500 and 2000 AD to describe spatial and temporal variation in bird introductions worldwide. Specifically, we (i) divide records for the first appearance of each alien species in each country into quartiles based on record date. Focussing on the first (1500–1900 AD) and last (1983–2000 AD) quartiles, we then (ii) produce maps showing which countries alien birds had been introduced to and (iii) maps of the native distributions of those species. We present analyses to show (iv) whether colonial history or economic activity (GDP) can explain which countries received more alien bird introductions, (v) whether the number of countries receiving birds has increased over time, and (vi) whether the regions of origin of those bird species has changed over time. Finally, (vii) we test whether the identities of species introduced in the historical and modern eras are a random subset of all bird species. The methods for these analyses are presented in the next section (Spatial and Temporal Variation in Bird Introductions). Second, we analyse global variation in the species richness of alien bird species at the 1° scale, incorporating information on colonisation pressure (first records arising from introduction rather than spread) as a key predictor of this variation. Specifically, we (viii) use data on the alien distributions of bird species to construct a global map of bird alien species richness at the scale of 1° grid cells and (ix) identify a set of anthropogenic and environmental variables that have been predicted to determine variation in alien species richness. We first (x) test for collinearity and (xi) spatial autocorrelation amongst the variables and use the outcomes of these tests to influence our final choices of predictor variables and methods. We then used (xii) Bayesian additive regression models and (xiii) simultaneous autoregressions to relate the chosen predictor variables to alien species richness while addressing the problem of spatial autocorrelation. We (xiv) use model selection approaches based on Akaike’s information criterion to identify key determinants of alien species richness and (xv) test the robustness of the predictors using holdout cross validation. The methods for these analyses are presented in the section entitled “Alien Bird Species Richness” below. All analyses were conducted using R (version 3.2.1) [41]. The subset of introduction records used here are those that represent the first record of a species in each country unit (countries, or states/provinces for Australia, Canada, and the United States) for the period 1500–2000 AD with known outcomes (statuses 1–5 above). Each species is therefore only counted in each country unit once, regardless of whether that introduction was successful or not. We censored introduction data to those falling within this time period because 1500 is a standard cutoff point for studying biological invasions [13,14], and there is evidence that introductions occurring after the year 2000 have not yet entered the literature and therefore represent an incomplete sample [23]. A very low proportion of recorded bird introductions are dated before 1500 AD (c.0.2%). We excluded from our analyses natural colonisations and translocations for conservation purposes. These criteria resulted in a total of 3,661 alien bird introduction records (first known occurrence of a given species in a given country unit) from 715 bird species (73.6% of species with introduced populations), which were distributed over time as shown in Fig 1. The data for Fig 1 are given in S1 Data. (i) To facilitate comparison between historical and modern introductions, the 3,661 records were split into four quartiles on the basis of introduction date. The first quartile includes 922 records encompassing the period 1500–1903 AD, the second quartile 895 records from 1904–1956 AD, the third quartile 909 records from 1957–1982 AD, and the fourth quartile 935 introductions during the period 1983–2000 AD. Here, we present analyses on the first (“historical”) and fourth (“modern”) quartiles; the second and third quartiles are intermediate in their characteristics to these (c.f. Fig 2 and S1 Fig). (ii) We constructed maps showing the countries into which species had been introduced in the historical and modern periods (Fig 2B and 2D) using introduction records from the GAVIA database, overlaid on the current map of countries (or country units) to avoid issues arising from historical changes in country boundaries. These maps show country-level estimates of colonisation pressure for the periods 1500–1903 AD (Fig 2B) and 1983–2000 AD (Fig 2D). (iii) We constructed richness maps for the native distributions of these introduced species (Fig 2A and 2C) using native geographic range maps extracted from the database used by Orme et al. [30]; they were created by projecting the range maps onto a hexagonal grid of the world, resulting in a geodesic discrete global grid, defined on an icosahedron and projected onto the sphere using the inverse Icosahedral Snyder Equal Area projection. This resulted in a hexagonal grid composed of cells that retain their shape and area throughout the globe. These maps were created using ESRI ArcGIS version 10.2.2 [42]. (iv) To assess the influence of British colonial history on alien bird introductions (Fig 3), we obtained a list of former British colonies [43]. The countries in which alien species were recorded for the first time during the historical and modern periods of introduction were assigned to either “British colony” or “non-British colony” categories. A two-sample Wilcoxon test was used to compare the number of alien bird introductions in the historical time period (first quartile of introduction dates) in former British colonies, versus the number of introductions in non-British colonies. This was repeated for introductions in the modern era (fourth quartile) to determine whether the influence changed over the historical span of bird introductions. To assess the influence of economic growth on alien bird introductions (Fig 3), data on GDP per capita (in 1990 international Geary-Khamis dollars [Int$]) were downloaded from ourworldindata.org (downloaded 18/03/15) [44]. We used per capita GDP because data are available for historical and modern time periods, it relates to the income of the populace, and countries with higher per-capita GDP tend also to have higher volumes of trade (r = 0.18, N = 207, p = 0.0066, for country-level GDP data from the UN and trade data from the CIA World Factbook, both sourced from Wikipedia on 25/08/16). Both income and trade are known to influence alien invasion pressure [17,26]. A subset of GDP data for the year 1900 was used for the countries present in the historical era, and the year 2000 for countries present in the modern era. GDP data were not available for all countries, particularly during the historical era, and countries without data were excluded, leaving 32 countries in the historical era analysis and 118 in the modern era. For the historical era, GDP per capita ranged from a minimum of Int$545 for China to a maximum of Int$5,899 for Switzerland (mean = Int$2,276; median = Int$1,980). For the modern era, GDP per capita ranged from a minimum of Int$509 for Sierra Leone to a maximum of Int$28,702 for the United States (mean = Int$7,162; median = Int$4,564). A linear regression was used to compare the number of alien bird introductions in the historical era and GDP per capita in the year 1900 (Int$). This was repeated for the modern era and GDP per capita in the year 2000 (Int$). (v) Bespoke simulations were used to test for differences between the observed and expected number of country units where alien species had been first recorded in either time period. Each iteration of the simulation involved selecting 922 introductions at random, and without replacement, for the historical era (and 935 for the modern era) from the full dataset of all introductions in the period 1500–2000 (n = 3,661), and calculating the number of country units to which those introductions in this randomly chosen subset were assigned. This process was repeated 10,000 times for each time period, and the observed number of country units was judged significantly different from the expected if the observed number fell outside of the 2.5%–97.5% quantiles. Additionally, we calculated the number of overlapping countries between the historical and modern eras, for each of the iterations paired by iteration number (1–10,000), in order to determine if species were being introduced into the same or different countries in the different time periods. The observed overlap was judged to be significantly different to the expected if the observed number fell outside of the 2.5%–97.5% quantiles. The same procedure was also used to test for differences between the observed and expected number of species introduced in each time period. (vi) In order to examine how the source locations of species introduced have changed between the two time periods (S4 Table), the native range of each species was intersected with the eight biogeographic realms defined by Olson et al. [45], using ESRI ArcGIS version 10.2.2 [42], and each was assigned to the realm where the native range of that species was found. No species from the Antarctic realm are included in these data, leaving seven realms in the analysis. For those species ranges that spanned more than one realm, the realm in which the largest part of the range fell was selected. Where the range was distributed equally across two or more realms, the species was excluded from the analysis (n = 27). This resulted in 225 species with assigned native realms in the historical era, and 298 species in the modern era (S4 Table). A Pearson’s Chi-squared test with a simulated p-value (based on 2,000 replicates) was used to determine whether the number of species sourced from each biogeographic realm was significantly different from that expected by chance between the historical era and the modern era. (vii) Bespoke simulations were also used to test for differences between the observed and expected number of introduced species from each bird family in both time periods (S3 Table). For these randomisations, a list of the total global avifauna was used (n = 10,245 species [46]). Each iteration of the simulation involved selecting 245 species at random, and without replacement, for the historical era (324 for the modern era) from the total global avifauna and summing the number of these randomly chosen species in each family. Larger numbers of species are expected to be selected by chance from more speciose families. A total of 10,000 iterations of the simulation procedure were run for each time period, and the observed number of introduced species in any given family was judged significantly greater than the expected number if at least S% of the randomly derived values for that family were less than the observed, where S = (β/ 2) x 100. The β is calculated by applying a sequential Bonferroni correction to α, and α = 0.05 [47]. Lists of the numbers of introductions by country for the quartiles of introductions, as ranked by date, are provided in S2 Data, along with lists of introduced bird species for the first and fourth quartiles. Data on GDP and number introductions by country for quartiles 1 and 4 are provided in S3 Data. The data used to plot Fig 2 and S1 Fig are given in S4 Data. (viii) Global analyses of ASR were based on the vector range maps and introduction records from the GAVIA database, and additional raster data on environmental and anthropogenic variables. For consistency with studies of native bird species richness patterns [19,30], all data were converted to a global grid using a Behrmann equal area projection at a cell resolution of 96.486 km, equivalent to 1° longitude and approximately 1° latitude at the equator. This was performed using the R packages sp [48,49] and raster [50]. The global grid contained 360 by 152 cells, omitting the partial cells at latitudes higher than 87.13°. Each grid cell was assigned latitude and longitude values, which represented the centre point of each cell. Alien species richness here is based on the 362 bird species with records of established (as defined above) alien populations (i.e., only those records in GAVIA with invasive status category 1) containing sufficient detail to convert to range maps using the software ESRI ArcGIS version 9.3 [51]. The most recently reported established range for each species was used to calculate alien species richness. The range maps were converted to grid cell counts using the R packages rgdal [52], sp [48,49], and raster [50]. Species were scored as present in a grid cell if any of the established introduced range fell within the cell boundaries. This ensured that even small established introductions, or those occurring on small islands, were counted. Overall alien species richness was derived by summing all species present within each cell and was mapped using ESRI ArcGIS version 10.2.2 (Fig 4) [42]. Alien species richness was natural log+1 transformed for analysis. Overall, the global grid contained 54,720 cells, but cells not containing any alien bird species records would inflate covariation measures (the double zero problem [53]), and therefore those with either no alien bird introductions or to where no alien bird species had subsequently spread (i.e., where both colonisation pressure = 0 and ASR = 0) were excluded. Any cell with missing data for any of the variables described below was also excluded from the analysis, leaving a total of 10,258 grid cells. (ix) A set of anthropogenic and environmental predictor variables were selected for use in model building based on their suitability for hypothesis testing (S1 Table). Available raw data for each of the candidate variables were re-projected and re-sampled to the same equal area grid as the alien species richness data using spatial tools from R (for details see S2 Table) [41]. (x) Tests of collinearity between the predictor variables found relatively high correlations within the temperature variables, elevation variables, and habitat complexity variables, and between human population density, human footprint index, and distance to the nearest city (S8 Table). The predictor variables temperature minimum, maximum, and range, median elevation, and the habitat complexity in surrounding 24 cells were thus excluded from models a priori. As the human footprint index incorporated human population density, human infrastructure, and road access, population density mean and median and distance to city were also excluded a priori. This resulted in nine predictor variables for analysis. (xi) Spatial autocorrelation is a common phenomenon in environmental data, where similarities in the values of predictor and response variables arise as a function of proximity of sampling locations. Species distribution data in particular are inherently spatially structured [60] due to a combination of intrinsic processes such as population growth and dispersal [61], areas of false presence or absence records due to errors in distributional data [62], or where the environmental processes that drive species richness patterns show spatial autocorrelation themselves [61]. There is strong spatial autocorrelation in both the response and predictor variables in our data (Moran’s I ≥ 0.74; p < 0.001 in all cases, with the exception of habitat complexity, for which Moran’s I = 0.21; p < 0.001), and therefore regression methods that assume each grid cell is an independent data point are inappropriate here. Given the problem of spatial autocorrelation, we analysed variation in alien species richness in two alternate ways. (xii) First, we conducted the analysis using a spatially structured random effect. We used stochastic partial differential equations (SPDE) [63] to fit a Gaussian random field to the data to approximate the patterns of spatial autocorrelation. We included this effect in a Bayesian additive regression model inferred using integrated Laplace approximations (INLA) in R (R-INLA) [64]. We tested several different versions of the spatial mesh, choosing complexity using Watanabe-AIC (wAIC) [65] to estimate the benefit of increasing the complexity of the spatial term versus its ability to explain complex spatial patterns. As the dependent variable (alien species richness) had a right-skewed distribution, we tested several candidate error “families” (Gaussian, log-Gaussian, Poisson, negative binomial) and compared these models using the conditional predictive ordinate (CPO) measure of fit, which gives the probability of each individual data point given the model (and ranges from 0 in the case of poor fit, to N [the sample size] in the case of perfect fit), and the probability integral transform (PIT) values. The log-Gaussian distribution was the best fit to the data with zero CPO failures, a right-skewed CPO histogram, showing that very few values had a low probability given the data and a uniform PIT distribution [66]. (xiii) To validate the INLA results, we also analysed variation in alien species richness using simultaneous autoregressions (SAR). Neighbourhood size was defined as the distance that captured the centre point of all eight surrounding grid cells (150 km). Neighbourhood connection matrices were calculated with row-standardised weights. Two specifications of the error covariance matrix were considered: SARlag (spatial autocorrelation in the response) and SARerr (spatial autocorrelation in the error term). A Lagrange multiplier test was used to find the best error specification, and the SARerr model showed higher support (SARerr: Lagrange multiplier LMerr = 20,876, p < 0.001; SARlag: Lagrange multiplier LMlag = 15,219, p < 0.001). SARerr models are recommended as most reliable when dealing with spatially autocorrelated species distribution data, having been found to perform well and provide the most precise parameter estimates regardless of the kind of spatial autocorrelation and whether model selection is via R2 or AIC [60]. All SAR models were constructed with the R package spdep [67,68]. (xiv) We ran single predictor models of all variables in order to compare the significance and directions of slopes for different predictors modelled in isolation (S5 Table). We included quadratic as well as linear terms for six of the predictors in our models to allow for non-linear relationships (colonisation pressure, time since introduction, native species richness, elevational range, median temperature, and precipitation). The inclusion of quadratic terms was determined by comparing single predictor models for each linear term with a model containing both the linear and quadratic form. For the SAR regression mode, if the AIC (Akaike’s information criterion) [69] improved by >4 then the quadratic form was included in model building. For the INLA model we used the analogous Watanabe-AIC (wAIC) [65], also with a threshold of 4. A multivariate minimum adequate model (MAM) was generated by forward stepwise procedures. The single predictor model with the lowest AIC or wAIC value was used as a starting model, with each “next best” predictor added in turn. The criterion for inclusion of additional model terms was improvement of the AIC value by >4. The use of an AIC type approach in model selection procedures observes principles of parsimony and avoids the model over-fitting that can be a result of data dredging [69]. Once the MAM with the lowest overall AIC score was identified, each predictor not included was once again added in turn to ensure that the best combination of predictor variables was selected. For the INLA models, ∑CPO, another goodness-of-fit measure, was also calculated, where higher scores represent more data points having high probability given the model (S3 Fig). A maximum model fit for INLA or SAR models was also assessed with a pseudo-R2 value calculated as the squared Pearson correlation between predicted and observed values (S3 Fig) [60]. These values are for illustrative purposes only and should not be overinterpreted. For the INLA models, spatial plots of residuals show that autocorrelation is virtually eliminated by including a spatial term (S4 Fig). For SAR models, spatial correlograms were used to examine the patterns of spatial autocorrelation for alien bird species richness and for the model residuals, and confirm that the method also largely eliminates spatial autocorrelation in the MAM (S5 Fig). (xv) To assess our model selection methods and the robustness of model parameter estimates, holdout cross validation was performed (S6 Table) in a manner akin to k-fold cross validation [70], but following Newbold et al. [71] using realms to test the biological predictions of the models. This allowed us to provide a more realistic biological test of model predictions, rather than using hold-out models on randomly divided sets of data. Each of the 10,258 grid cells was assigned to a biogeographic realm [45]. The grid cells from the Antarctic and Oceanic realms were excluded from this part of the analysis due to low sample size. Five of the six remaining realms were used as the training set upon which stepwise model selection was conducted, as described above. The sixth realm was then used as the testing set, and a cross validation metric, root mean squared error (RMSE), was calculated to assess the ability of the final model at predicting the held-out realm. This process was repeated five more times with a different realm being used as the testing set each time, and then the RMSE across all samples was averaged to obtain the mean cross validation error, which provides an estimate of the predictive accuracy of our full data model if challenged by new data. INLA and SAR models gave qualitatively similar results, identifying the same key predictors of global spatial variation in alien species richness. Therefore, we focus on the results of the INLA analysis, as this was the more conservative of the two approaches. The equivalent SAR results are available on request. The data frame for the analysis of alien species richness is provided in S5 Data.

10.1371/journal.pntd.0005932

Structural basis for the high specificity of a Trypanosoma congolense immunoassay targeting glycosomal aldolase

Animal African trypanosomosis (AAT) is a neglected tropical disease which imposes a heavy burden on the livestock industry in Sub-Saharan Africa. Its causative agents are Trypanosoma parasites, with T. congolense and T. vivax being responsible for the majority of the cases. Recently, we identified a Nanobody (Nb474) that was employed to develop a homologous sandwich ELISA targeting T. congolense fructose-1,6-bisphosphate aldolase (TcoALD). Despite the high sequence identity between trypanosomatid aldolases, the Nb474-based immunoassay is highly specific for T. congolense detection. The results presented in this paper yield insights into the molecular principles underlying the assay’s high specificity. The structure of the Nb474-TcoALD complex was determined via X-ray crystallography. Together with analytical gel filtration, the structure reveals that a single TcoALD tetramer contains four binding sites for Nb474. Through a comparison with the crystal structures of two other trypanosomatid aldolases, TcoALD residues Ala77 and Leu106 were identified as hot spots for specificity. Via ELISA and surface plasmon resonance (SPR), we demonstrate that mutation of these residues does not abolish TcoALD recognition by Nb474, but does lead to a lack of detection in the Nb474-based homologous sandwich immunoassay. The results show that the high specificity of the Nb474-based immunoassay is not determined by the initial recognition event between Nb474 and TcoALD, but rather by its homologous sandwich design. This (i) provides insights into the optimal set-up of the assay, (ii) may be of great significance for field applications as it could explain the potential detection escape of certain T. congolense strains, and (iii) may be of general interest to those developing similar assays.

Sub-Saharan Africa is plagued by many diseases, which impede its socio-economical development. One of these diseases, Animal African Trypanosomosis, affects livestock and is caused by the parasites of the Trypanosoma genus (T. vivax and T. congolense). Animal African Trypanosomosis leads to considerable economic losses and renders sustainable livestock industry in Sub-Saharan Africa very difficult. In order to proceed with the selective treatment of infected animals, they need to be properly diagnosed. We recently described the use of an assay to specifically detect T. congolense infections in both experimentally and naturally infected animals. The diagnostic assay employs a Nanobody (Nb), which is the smallest antigen-binding unit derived from camelid heavy-chain only antibodies. Our previous results showed that the Nb-based diagnostic test specifically recognizes T. congolense fructose-1,6-bisphosphate aldolase, a glycolytic enzyme that is well conserved amongst other Trypanosoma species. In this paper, we studied the molecular mechanisms underlying the high specificity of the Nb-based diagnostic assay. The principles derived from this work may be important for the design and improvement of similar diagnostic tests.

The Trypanosoma genus represents a diverse group of extracellular hemoflagellated parasites of which some members can infect and cause disease in humans and livestock. An infection with African trypanosomes generally leads to the development of pathologies called Human African Trypanosomosis (HAT) and Animal African Trypanosomosis (AAT), respectively. In the case of AAT (also called “Nagana”), the predominant causative agents are T. congolense and T. vivax. Estimates place 50 million animals at risk of infection in Sub-Saharan Africa and indicate that the annual AAT-driven economic losses to the local livestock industry are close to US$ 4 billion [1]. Evidently, AAT has a profound negative impact on the development of endemic regions. While drug treatments to combat AAT exist, these are deployed in an indiscriminate fashion on a large scale due to the lack of inexpensive, specific and easy-to-use diagnostic tests. These attributes are of great significance for the development of point of care tests (POCTs) for rapid detection of AAT in a low-income setting where the disease is endemic. Importantly, the practice of indiscriminate administration of anti-trypanosomal drugs to both healthy and diseased animals has led to the emergence of drug-resistant parasite strains [2,3]. For this reason, there are ongoing efforts by the research community to develop both DNA- and protein-based tests to improve diagnosis of AAT in the field. An important category of assays for the diagnosis of trypanosomosis is the one of immunodiagnostics such as enzyme-linked immunosorbent assays (ELISAs) and lateral flow assays (LFAs) [4–9]. Both the antibody- and antigen-based immunodiagnostics have their advantages and drawbacks. The antibody-based tests, which rely on the detection of circulating parasite-induced host antibodies, have two main disadvantages: (i) a low specificity due to antibody cross-reactivity [10] and (ii) the inability of differentiating between past and ongoing infections as a consequence of long lasting circulating antibodies after parasite clearance [11–13]. These problems are alleviated by antigen-based assays, which aim to detect circulating parasite antigens. However, antigen-based diagnostic tests face their own issues. First, during an active infection, antigen levels should be high enough in order to be detected [14]. Second, the detection and capturing antibodies of the assay should bind to different epitopes from the host antibodies, which usually form immuno-complexes with the circulating antigen or at least be able to outcompete them [15]. Finally, the assay’s antibodies should be species-specific, which is not straightforward given that some antigens are highly conserved among different Trypanosoma species. Recently, we described an antigen-based immunoassay for diagnosis of active T. congolense infections using Nanobody (Nb) technology [16]. The immunoassay is designed in the format of a homologous sandwich ELISA employing a single Nb (Nb474). The target of the assay was identified as T. congolense fructose-1,6-bisphosphate aldolase (TcoALD) and, hence, validates this enzyme as a diagnostic biomarker for T. congolense infections [16]. Fructose-1,6-bisphosphate aldolase is an enzyme involved in the glycolytic pathway and most members of this protein family occur in solution as tetramers. This is because of the low dissociation constants for the dimer-tetramer equilibria, resulting in stable tetramer formation [17]. In trypanosomatids, most glycolytic enzymes (including aldolase) are located in specialized organelles called glycosomes [18]. At first glance, the potential of an intracellular, glycosomal enzyme as an infection biomarker may seem counterintuitive. However, it has recently been discovered that, within the context of host-parasite interactions, trypanosomes produce extracellular vesicles containing many different proteins including fructose-1,6-bisphosphate aldolase [19]. As such, TcoALD is part of the T. congolense “secretome”, i.e. the collection of all molecules secreted/excreted by the parasite [20], which is probably why it can act as a biomarker for active T. congolense infections. The Nb474-based ELISA is highly specific for T. congolense as infections with other trypanosomes such as T. brucei brucei, T. vivax, and T. evansi are not detected. Concomitantly, the Nb474 sandwich ELISA only yields a positive signal when incubated with TcoALD and is negative for the detection of TbALD and LmALD (T. brucei brucei and Leishmania mexicana glycosomal fructose-1,6-bisphosphate aldolase) [16]. This is remarkable given that glycolytic enzymes such as fructose-1,6-bisphosphate aldolase display a relatively high degree of sequence conservation, especially among different trypanosome species (94.1% for TcoALD and TbALD). These findings raise questions about the molecular details of the Nb474-TcoALD interaction determining the specificity of this particular assay. In this study, we present the structural basis for the high specificity of the Nb474-based T. congolense homologous sandwich ELISA. Using a combination of X-ray crystallography, site-directed mutagenesis, ELISA and surface plasmon resonance (SPR), we demonstrate that the high specificity of the Nb474-based immunoassay is determined by its sandwich design. The results may serve as a basis for the improvement of the Nb474-based ELISA and the design of similar antigen-based diagnostic tests. The generation of Nb474 by alpaca immunization, the recombinant production of its C-terminally His-tagged variant in E. coli and subsequent purification by IMAC and size exclusion chromatography (SEC) have been described recently [16]. The recombinant production of C-terminally His-tagged TcoALDWT in E. coli was performed as described [16]. To purify TcoALDWT from an overnight production culture, cells were first harvested by centrifugation (10 min; 8000 rpm, JLA-8.1000 rotor; 14°C). The bacterial pellets were resuspended in lysis buffer (50 mM Tris-HCl, 500 mM NaCl, pH 8.0) and aliquoted in volumes of 50 ml. The aliquots were flash-frozen using liquid nitrogen and stored at -80°C. Prior to purification, aliquots were thawed on ice. Cells were lysed using a sonicator (Ultrasonic disintegrator MSE Soniprep 150; 5 sonication cycles of 1 min at 15 microns amplitude with a 2 min pause between each cycle) and the cell lysate was centrifuged (20 min, 18000 rpm, JA-20 rotor, 4°C). The supernatant was collected and filtered (0.45 μm). The purification of TcoALDWT was performed on an AKTA Prime Platform (GE Healthcare) using IMAC and SEC. A 5 ml HisTrap HP nickel-sepharose column (GE Healthcare) was equilibrated with buffer A (50 mM Tris-HCl, 500 mM NaCl, pH 8.0) for at least five column volumes. The sample was loaded on the column using the same buffer at a flow rate of 1 ml min-1. After loading, the column was further washed with 5 column volumes of the same buffer. TcoALDWT was then eluted by a linear gradient of buffer B (50 mM Tris-HCl, 500 mM NaCl, 1 M imidazole, pH 8.0) over 20 column volumes. The fractions containing TcoALDWT were pooled and concentrated to a final volume of 2 ml for the subsequent SEC step on a Superdex 200 16/60 column (GE Healthcare), which was pre-equilibrated with at least one column volume of buffer C (50 mM MES, 500 mM NaCl, pH 6.7). The sample was eluted at a flow rate of 1 ml min-1. Fractions containing TcoALDWT were pooled and stored at 4°C. Each of the purification steps was monitored by SDS-PAGE and Western blot under reducing conditions. The purification and storage conditions for TcoALDWT were optimized via differential scanning fluorimetry (DSF, see S2 Fig). The TcoALDA77E, TcoALDL106Y, and TcoALDA77E/L106Y mutants were generated by modifying the TcoALDWT sequence. Synthesis and cloning of the mutant sequences was outsourced to a commercial company (GenScript). These mutants were produced and purified as described for TcoALDWT. The stoichiometry of the Nb474-TcoALD complex was determined by analytical SEC. The experiments were performed using a Superdex 200 HR 10/30 (GE Healthcare) column, pre-equilibrated with buffer C for at least one column volume. Five hundred μl samples containing 1 mg TcoALD mixed with varying molar ratios of Nb474 (Nb474:TcoALD ratios of 1:4, 2:4, 3:4, 4:4, and 6:4, respectively) were allowed to incubate for at least 1 h prior to injection. The samples were eluted with a flow rate of 0.5 ml min-1 and the elution peaks of all chromatograms were subjected to SDS-PAGE analysis. The column was calibrated with the Bio-Rad molecular mass standard under the same conditions. The (Nb474-TcoALD)4 complex was generated by mixing Nb474 and TcoALD in a Nb474:TcoALD ratio of 6:4, allowing the sample to equilibrate for at least 1 h prior to purification on a Superdex 200 16/60 column (GE Healthcare) pre-equilibrated with at least one column volume of buffer C. The sample was eluted at a flow rate of 1 ml min-1. Fractions containing the (Nb474-TcoALD)4 complex were pooled and stored at 4°C. Differential scanning fluorimetry (DSF) experiments were performed to optimize the purification and storage conditions of TcoALD. DSF was performed on a CFX Connect Real-Time System Thermal Cycler (Bio-Rad). Data were collected from 10°C to 95°C at a scan rate of 1°C min-1. The fluorescence signal was recorded every 0.5°C. Experiments were carried out in 96-well plates and the total sample volume was 25 μl. To determine the optimal protein-dye ratio, a grid screen of various concentrations of SYPRO orange dye (Life Technologies) (0x, 5x, 10x, 50x, 100x) and TcoALD (0 μM, 1 μM, 5 μM, 10 μM, 25 μM, 50 μM) was carried out. After identification of a suitable condition (10x SYPRO orange dye and 5 μM of TcoALD), buffer and additive screens were performed as previously described [21]. All experiments were conducted in duplicate. The (Nb474-TcoALD)4 complex was concentrated to 0.5 mg ml-1 using a 5,000 molecular weight cut-off concentrator (Sartorius Vivaspin20). Crystallization conditions were screened manually using the hanging-drop vapor-diffusion method in 48-well plates (Hampton VDX greased) with drops consisting of 2 μl protein solution and 2 μl reservoir solution equilibrated against 150 μl reservoir solution. Commercial screens from Hampton Research (Crystal Screen, Crystal Screen 2, Crystal Screen Lite, Index), Molecular Dimensions (MIDAS, JCGS+), and Jena Bioscience (JBScreen Classic 1–10) were used for initial screening. The His-tags of both proteins were retained for crystallization. The crystal plates were incubated at 20°C. Diffraction-quality crystals of the complex were obtained in Crystal Screen Lite (Hampton Research) no. 18 (100 mM sodium cacodylate pH 6.5, 200 mM magnesium acetate, 10% PEG 8000) and the crystals grew after approximately 14 days at 20°C. The (Nb474-TcoALD)4 crystals were cryocooled in liquid nitrogen with the addition of 25% (v/v) glycerol to the mother liquor as a cryoprotectant in 5% increments. Data were collected on the PROXIMA2 beamline at the SOLEIL synchrotron (Gif-Sur-Yvette, France) and were processed with XDS [22]. The quality of the collected data sets was verified by close inspection of the XDS output files and through phenix.xtriage in the PHENIX package [23]. Twinning tests were also performed by phenix.xtriage. Analysis of the unit-cell contents was performed with the program MATTHEWS_COEF, which is part of the CCP4 package [24]. The structure of the (Nb474-TcoALD)4 complex was determined by molecular replacement with PHASER-MR [25] using the structure of T. brucei aldolase (PDB ID: 1F2J, [26]) as a search model. This provided a single solution (top TFZ = 96.4 and top LLG = 14236.4). From here, refinement cycles using the maximum likelihood target function cycles of phenix.refine [27] were alternated with manual building using Coot [28]. The final resolution cut-off was determined through the paired refinement strategy [29], which was performed on the PDB_REDO server [30]. The crystallographic data for the (Nb474-TcoALD)4 complex are summarized in Table 1 and have been deposited in the PDB (PDB ID 5O0W). Molecular graphics and analyses were performed with UCSF Chimera [31]. The amino acid sequences of trypanosomatid homologs of TcoALD were obtained by a protein BLAST search of the TriTrypDB [32] using TcoALD (Uniprot ID: G0UWE7) as a query sequence. A total of 24 trypanosomatid aldolase sequences (including TcoALD) were employed to generate a sequence alignment with MAFFT [33] using the Geneious Pro program suite (Biomatters Ltd). The amino acid sequences of TcoALD homologs from T. congolense MOSROM_ALL, T. congolense SA268, T. congolense KAPEYA357, and T. congolense ZER-AGRIUMBE were kindly provided by dr. Hideo Imamura. The details on these sequences and how they were obtained have recently been published [34]. Genomic DNA samples from T. congolense TSW103, T. congolense WG84, T. simiae Ban7, and T. godfreyi Ken7 were kindly provided by Prof. dr. Wendy Gibson. More information concerning these sequences can be found in the work by Masiga et al. [35]. The gene encoding aldolase was extracted from these genomic DNA samples via PCR. Four different primers were designed to amplify the aldolase-coding genes based on the nucleotide sequence of the T. congolense IL3000 aldolase gene (Genbank accession number CCC93713.1): one set of primers to amplify the entire gene (TcoALDcFwd: 5’-ATGTCCAGGCGTGTGGAAGTTC-3’; TcoALDcRev: 5'-CTAGTAGGTGTTGCCAGCAAC-3'), a short region from the gene encoding Met1 to L181 (TcoALDcFwd: 5’-ATGTCCAGGCGTGTGGAAGTTC-3’; TcoALDOcshRev: 5'-CGAGCGTTTCAGCGTTGAA-3'), or a short region from the gene encoding Y162 to Y372 (TcoALDcshFwd: 5-ACAAGATTCAGAACGGCAC-3'; TcoALDcRev: 5'-CTAGTAGGTGTTGCCAGCAAC-3'). The PCR mix contained the following components: 0.4 mM forward primer, 0.4 mM reverse primer, 0.4 mM dNTPs, 1x GoTaq G2 buffer (Promega), 1.5 U GoTaq G2 DNA polymerase (Promega), 5 ng genomic DNA. The PCR was performed according to the following protocol: (i) 30 cycles of denaturation (95°C, 5 min), denaturation (94°C, 1 min), annealing (55°C, 1.5 min), elongation (72°C, 1 min), (ii) elongation (72°C, 10 min), (iii) storage (4°C). Amplified PCR products were resolved by electrophoresis on 1% agarose (Lonza) in TBE buffer (90 mM Tris, 90 mM borate, 2.5 mM EDTA). Electrophoresis was conducted at 100V for 30 minutes. Amplicons were cleaned-up using the PCR clean-up kit (Sigma-Aldrich) following the protocol recommended by the manufacturer. Gene sequences were obtained through DNA sequencing with 50 pmol of each primer. Sequencing of the samples was outsourced to the VIB Genetic Service Facility. A total of 16 aldolase sequences were employed to generate a sequence alignment with MAFFT [33] using the Geneious Pro program suite (Biomatters Ltd). The Nb474H/Nb474B homologous and Nb474B/mouse anti-His heterologous sandwich ELISAs were performed in similar manner as previously described [16]. Briefly, Nb474H (homologous ELISA) or Nb474B (heterologous ELISA) was coated on the plate as capture reagent by applying 100 μl (diluted to a concentration of 0.02 μg ml-1 in PBS) per well. The plate was incubated overnight at 4°C and the excess of non-coated Nb was removed by washing the plate three times with PBS containing 0.01% Tween20 (PBS-T). Next, blocking of residual protein binding sites was performed by adding 300 μl blocking buffer (5% milk powder in PBS) to each well and the plate was kept for 2 h at room temperature. Subsequently, the plate was washed three times with PBS-T, after which the TcoALD wild type and mutant variants were allowed to interact with the coated Nb by applying 100 μl (diluted to a concentration of 1 μg ml-1 in blocking buffer) per well. After incubation for 1 h at room temperature, the plates were subsequently washed three times with PBS-T. Then, 100 μl Nb474B (diluted to a concentration of 0.02 μg ml-1 in blocking buffer) or 100 μl mouse anti-His (diluted to a concentration of 0.05 μg ml-1 in blocking buffer) was added to the plate as a primary detection reagent for the homologous and heterologous ELISAs, respectively. After an incubation of 1 h at room temperature, the plate was washed 5 times with PBS-T. The conjugate, 100 μl of streptavidin-HRP (diluted to a concentration of 1 μg ml-1 in rinsing buffer) or 100 μl goat anti-mouse-HRP (diluted to a concentration of 0.05 μg ml-1 in blocking buffer), was then added to the plate for the homologous and heterologous ELISAs, respectively, followed by incubation for 1 h at room temperature. After a final washing step (5 times with PBS-T), the ELISAs were developed by addition of 100 μl of 3,3’,5,5’-tetramethylbenzine (TMB) substrate and subsequent incubation for 25 min at room temperature. The enzymatic reaction was stopped by adding 50 μl 1 M H2SO4 to the reaction mixture. The plates were read at OD450 nm with a VersaMax ELISA Microplate Reader (Molecular Devices). Surface plasmon resonance (SPR) experiments were performed on a BIAcore T200 system (GE Healthcare). The interactions between Nb474 and the TcoALD variants were analyzed on a CM5 chip. Nb474 was immobilized in flow cell 2 using the following procedure. Using a flow rate of 5 μl min-1 the carboxylated dextran matrix was activated by a 7-min injection of a solution containing 0.2 M N-ethyl-N′-(3-diethylamino)propyl carbodiimide (EDC) and 0.05 M N-hydroxysuccinimide (NHS). A Nb474 solution of 1 μg ml-1 (50 mM sodium acetate pH 5.0) was subsequently injected until the desired amount of protein was immobilized (approx. 50 R.U.). The surface immobilization was then blocked by a 7-min injection of 1 M ethanolamine hydrochloride. The surface in flow cell 1 was used as a reference and treated only with EDC, NHS and ethanolamine. Sensorgrams for different concentrations of the TcoALD variants expressed as monomer concentrations (0.02 nM, 0.05 nM, 0.10 nM, 0.20 nM, 0.35 nM, 0.50 nM, 0.75 nM, 1.00 nM, 2.00 nM, 5.00 nM, 10.00 nM for TcoALDWT; 0.78 nM, 1.56 nM, 3.12 nM, 6.25 nM, 12.50 nM, 25.00 nM, 50.00 nM, 100.00 nM, 125.00 nM, 250.00 nM, 500.00 nM for TcoALDA77E; 0.12 nM, 0.24 nM, 0.49 nM, 0.98 nM, 1.95 nM, 3.90 nM, 7.81 nM, 15.62 nM, 31.25 nM, 62.50 nM, 125.00 nM for TcoALDL106Y; 0.78 nM, 1.56 nM, 3.12 nM, 6.25 nM, 12.50 nM, 25.00 nM, 50.00 nM, 100.00 nM, 125.00 nM, 250.00 nM, 500.00 nM, 750.00 nM, 1.00 μM for TcoALDA77E/L106Y) plus a 0 concentration (injection of running buffer) were collected at a flow rate of 30 μl min-1 and a data collection rate of 1 Hz. For the Nb474 binding/washing experiments, the ligand (Nb474) was first saturated by an injection of an adequate concentration of the first analyte (TcoALD variant; 10 nM for TcoALDWT, 125 nM for TcoALDA77E, 31.25 nM for TcoALDL106Y, and 750 nM for TcoALDA77E/L106Y). Immediately after injection of the first analyte (i.e., no dissociation phase), different concentrations of the second analyte (Nb474; 0.5 nM, 1.0 nM, 1.5 nM, 5.0 nM, 10.0 nM, 100.0 nM, 500.0 nM, 1.0 μM) plus a 0 concentration (injection of running buffer) were injected at a flow rate of 30 μl min-1 and a data collection rate of 1 Hz. All analytes were dialyzed into the running buffer (20 mM HEPES, 150 mM NaCl, 0.005% Tween, 3.4 mM EDTA, pH 7.4) prior to data collection. Analyte injections were performed with association and dissociation phases of 480 s and 660 s, respectively. This was followed by a 5 μl pulse injection of regeneration buffer (0.2% SDS). Prior to data analysis, reference and zero concentration data were subtracted from the sensorgrams. The data collected for Nb474 binding to the pre-formed Nb474-TcoALDWT and Nb474-TcoALDA77E complexes were analyzed with a 1:1 Langmuir binding model. All experiments were performed on the same sensor chip using the same flow channels. The amino acid sequences of glycolytic enzymes such as fructose-1,6-bisphosphate aldolase are generally well conserved. Indeed, among trypanosomatids, the pairwise sequence identity for aldolase is 86.7% (S1 Fig). For TcoALD and TbALD, the sequence identity is 94.1%. Nonetheless, the Nb474-based immunoassay is highly specific for TcoALD. Thus, we were interested in identifying the TcoALD epitope recognized by Nb474. First, we produced TcoALD through recombinant protein production in E. coli and optimized its purification conditions via DSF. The details of these procedures are given in Materials and Methods and S2 Fig (panels A-D). Next, we determined the stoichiometry of the Nb474-TcoALD complex via analytical SEC (Fig 1). An excess of Nb474 could only be detected at a molar ratio of 6:4 between Nb474 and the TcoALD monomer, and not at the other tested ratios 1:4, 2:4, 3:4, and 4:4. This suggests that one Nb474 binds a single TcoALD monomer. The analytical SEC reveals another interesting feature of the Nb474-TcoALD interaction. First, TcoALD appears to occur as a dimer in solution. The TcoALD monomer has a theoretical molecular mass of 42.6 kDa (170.4 kDa for a TcoALD tetramer). Instead, TcoALD migrates with a higher apparent molecular mass (MMapp = ~66 kDa, Fig 1H), suggesting a dimer population (TcoALD2). Second, a comparison of the analytical SEC profiles recorded for the different ratios between Nb474 and the TcoALD monomer suggests that adding Nb474 to TcoALD promotes tetramer formation. Rather than shifting the TcoALD2 peak to the left, the titration of Nb474 reduces the intensity of the TcoALD2 peak and gives rise to a peak corresponding to an entity of larger molecular mass. At an estimated molecular mass of ~230 kDa for the peak at a 4:4 molar ratio (Fig 1F), this complex likely corresponds to a hetero-octameric (Nb474-TcoALD)4 complex in which four Nb474 occupy identical sites on the TcoALD tetramer (TcoALD4). Indeed, the theoretical molecular mass of such a complex is 233.52 kDa, which is in accordance with the molecular mass calculated based on the analytical SEC data (Fig 1H). For crystallization purposes, the (Nb474-TcoALD)4 complex was prepared using a 6:4 molar ratio as described above and purified by SEC. Crystals of the (Nb474-TcoALD)4 complex and their diffraction are shown in S2 Fig (panels E-F). The details of the crystallographic experiment are summarized in Table 1. The crystal structure of the (Nb474-TcoALD)4 complex confirms that TcoALD4 indeed contains 4 binding sites for Nb474 (Fig 2). Nb474 binds an epitope on the TcoALD surface that is located relatively far away from the aldolase A and B dimer interfaces. This results in large distances between the Nb474 epitopes on TcoALD4 relative to the A and B dimer interfaces (~ 69 Å and ~ 79 Å from one epitope to another, respectively). The Nb474-TcoALD interaction is mediated by residues from all three complementarity determining regions (CDRs; Fig 2). The bulk of the contacts are provided by CDR1 and CDR3, while a single amino acid from CDR2 (Arg53) is involved in TcoALD recognition. A detailed overview of all the interactions is given in S3 Fig and S1 Table. A superposition of the crystal structures of TbALD (PDB ID: 1F2J, [26]), LmALD (PDB ID: 1EPX, [26]), and the (Nb474-TcoALD)4 complex allows to pinpoint those residues that are located in the vicinity of or on the TcoALD epitope recognized by Nb474 and are distinct between the three trypanosomatid aldolases (Fig 3A). These residues are located at positions 76, 77, 96, 98, 99, 101, 109, 328, and 332 for TcoALD and TbALD. For LmALD, all positions are shifted by -1. We reasoned that mutating some of these TcoALD residues to their TbALD/LmALD counterparts would influence Nb474 binding and provide a starting point to explain the assay’s specificity. Within the above-mentioned selection of amino acids, we first identified those residues that are conserved in both TbALD and LmALD, but differ in TcoALD. These amino acids would most likely contribute to a loss of binding energy given that the Nb474-based immunoassay does not provide a binding signal for both TbALD and LmALD [16]. This narrowed the selection of residues to mutate down to four positions: A77/E77, R96/K96, L106/Y106, and S328/E328 for TcoALD/TbALD (E76, K95, Y105, and E327 for LmALD). Positions 96 and 328 were further omitted because, based on the structural comparison depicted in Fig 3B, these are too far from the Nb474 paratope to have any influence on Nb474-TcoALD interaction. This resulted in a final selection of two positions that were targeted for site-specific mutagenesis: A77/E77 and L106/Y106 for TcoALD/TbALD (E76 and Y105 for LmALD). Hence, three TcoALD mutants (TcoALDA77E, TcoALDL106Y, and TcoALDA77E/L106Y) were generated. The different TcoALD variants were tested in the Nb474-based homologous sandwich ELISA and compared. As can be seen from Fig 4A, TcoALDA77E is still detected, although to a lesser extent compared to TcoALDWT, whereas TcoALDL106Y and TcoALDA77E/L106Y display no signal. Two hypotheses could be presented to explain these observations. The first poses that the lack of detection of the TcoALD mutants is caused by a loss of recognition by Nb474 due to the introduced mutations. The second explanation states that the mutations somehow weaken the Nb474-TcoALD interaction and that a “self-competition” or “washing” effect is at play because of the homologous sandwich design of the assay. In order to distinguish between both hypotheses, a second, heterologous ELISA was carried out with Nb474B as a capturing agent (Fig 4B). Compared to TcoALDWT, the three mutants display a lower but clear signal, with TcoALDA77E/L106Y providing the lowest intensity. When combined, the results of both ELISAs suggest that Nb474 still interacts with the TcoALD mutants, thus favoring the second hypothesis. The interaction between Nb474 and each of the TcoALD variants was investigated further via SPR. For this experiment, Nb474 and the TcoALD variants were employed as ligand and analytes, respectively. From Fig 5 (panels A-D), it is clear that all TcoALD mutants bind to Nb474 and that the kinetics of the Nb474-TcoALD interaction are affected by the A77E, L106Y, and A77E/L106Y mutations. Unfortunately, this could not be quantified by any interaction model, which is why the interpretation of the presented SPR data is performed in a semi-quantitative fashion. For TcoALDWT, saturation is readily observed at an enzyme concentration of 10 nM (maximal binding signal Rmax of ~110 R.U.; Fig 5A). This indicates that the affinity of the Nb474-TcoALDWT interaction is quite high (nM to pM range), which is supported by the necessity of solutions containing 0.2% SDS to regenerate the Nb474-coated sensor chip surface. The three TcoALD mutants only attain a similar maximal binding signal at higher analyte concentrations (Fig 5, compare panels A-D), suggesting that the binding of Nb474 to the TcoALD mutants is weakened by the introduced mutations. Interestingly, although TcoALDA77E only reaches a binding signal of ~110 R.U. at a concentration of 500 nM (Fig 5B), the dissociation phases for the Nb474-TcoALDWT and Nb474-TcoALDA77E interactions seem very similar. In the case of TcoALDL106Y, a signal of ~110 R.U. is attained at a concentration of 125 nM (Fig 5C), while the dissociation of this complex appears to occur faster. Finally, binding of Nb474 to TcoALDA77E/L106Y does not reach the maximal signal observed for the other TcoALD variants, even at a concentration of 1 μM (Fig 5D). An additional SPR experiment was designed to mimic the homologous sandwich ELISA (Fig 5, panels E-H). The basic set-up is the same as mentioned above: Nb474 and the TcoALD variants were selected as ligand and analytes, respectively. For each TcoALD variants, the analyte concentration was chosen to saturate the Nb474-coated sensor chip surface. Upon saturation, Nb474 was injected onto the sensor chip surface at different concentrations and allowed to interact with the pre-formed Nb474-TcoALD complex. For both TcoALDWT and TcoALDA77E, this results in additional binding of Nb474 and formation of Nb474-TcoALDWT-Nb474 and Nb474-TcoALDA77E-Nb474 sandwiches (Fig 5, panels E and F). Interestingly, these binding curves can be fitted with a 1:1 Langmuir binding model. It appears that the Nb474-TcoALD interaction is virtually unaffected by the A77E mutation as the affinity constants for both binding events are quasi identical (Table 2). In the case of TcoALDL106Y and TcoALDA77E/L106Y, the injection of additional Nb474 leads to dissociation of the pre-formed Nb474-TcoALD complex as evidenced by a reduction in RU signal over time (Fig 5, panels G and H). The ELISA data in conjuncture with the SPR results indicate that the high specificity for TcoALD displayed by the Nb474-based immunoassay is not determined by the initial interaction between TcoALD and the capturing Nb474, but rather from its homologous sandwich design. The above-mentioned mutation studies imply that T. congolense strains carrying mutations at positions 77 and/or 106 would be detected less efficiently (or not at all) by the Nb474-based homologous sandwich ELISA. To probe the aldolase sequence variation within the Trypanosoma subgenus Nannomonas, of which T. congolense is a member, the aldolase amino acid sequences were determined for the following parasites: T. congolense (Savannah type), T. congolense (Forest type), T. congolense (Kilifi type), T. simiae, and T. godfreyi. A sequence alignment reveals that the sequence identity of aldolase within Nannomonas is relatively high (90.6%; S4 Fig). Interestingly, while position 106 remains unaltered throughout all sequenced Nannomonas members (Leu106), position 77 displays a larger sequence variation: T. congolense Savannah and Kilifi subtypes contain an Ala77, T. congolense Forest subtypes harbor a Val77, and T. simiae and T. godfreyi both possess a Glu residue at position 77 (S4 Fig). Animal African Trypanosomosis is neglected on a global scale, yet it continues to impose a heavy societal and economic burden on Sub-Saharan Africa [1]. However, the disease can be contained provided that the necessary control and surveillance programs are put in place. For HAT, such multidisciplinary initiatives have “eliminated the disease as a public health problem” [36], which means that in most areas HAT can be targeted for eradication. Together with vector control strategies and adequate treatment schemes, tools for rapid diagnosis of AAT are of utmost importance. Luckily, such assays targeting T. congolense and T. vivax infections (the two important causative agents of AAT in livestock) are under development [4–7,9]. Recently, we described the generation of a highly specific Nb-based homologous sandwich ELISA targeting TcoALD to detect active T. congolense infections [16]. Evidently, the principle of a homologous sandwich ELISA can only work if the target antigen is a multimer [37,38]. Members of the fructose-1,6-bisphosphate aldolase family usually occur in solution as stable tetramers [17]. Mutations at the A and B dimer interfaces influence the dimer-tetramer equilibria by destabilizing the tetramer, but, interestingly, without affecting the enzyme’s catalytic activity [17,39,40]. While aldolase dimers retain the same catalytic potential compared to tetramers, they appear to be less thermostable [39]. In the case of TcoALD, the analytical SEC data presented here indicates that the enzyme does not occur as a tetramer in solution, but rather seems to behave as a dimer. This suggests that the dissociation constants for the dimer-tetramer equilibria are higher for TcoALD compared to archetypal aldolases. This may also explain why TcoALD was observed to be labile during our first purification trials. Using DSF, we markedly improved the thermal stability of TcoALD (melting temperatures Tm of ~40°C and ~50°C in the initial and final buffer conditions, respectively). Interestingly, in their work on rabbit muscle aldolase, Beernink and Tolan measured Tm values of ~45°C and ~60°C for aldolase dimers and tetramers, respectively [39]. Nb474 clearly influences the TcoALD dimer-tetramer equilibria. As shown by analytical SEC, the titration of Nb474 against TcoALD shifts the equilibrium towards the formation of an aldolase tetramer. The end-point of the titration is reached at a Nb474:TcoALD molar ratio of 4:4, suggesting the formation of a hetero-octameric (Nb474-TcoALD)4 complex, which is confirmed by X-ray crystallography. The crystal structure of the (Nb474-TcoALD)4 complex provides a molecular basis as to why the homologous sandwich ELISA format works in the case of TcoALD. The Nb474 epitope is located on the extremities of the TcoALD tetramer, thereby easily allowing all four copies of Nb474 to bind their epitopes without mutual interference. A detailed analysis of the Nb474-TcoALD interface reveals a multitude of interactions between both proteins, mainly mediated by CDR1 and CDR3 residues. The SPR data demonstrate that these interactions result in a high-affinity recognition event (KD in the pM range), which explains why Nb474 is such a good capturing agent [16]. In some cases, the mutation of a single residue on the antigen’s epitope can cause total loss of antigen recognition by the Nb [41]. In contrast, the mutation studies presented here indicate that this is not the case for the Nb474-TcoALD interaction. Changing specific TcoALD epitope residues to their TbALD/LmALD counterparts (A77E, L106Y, A77E/L106Y) does not result in a loss of TcoALD recognition by Nb474. Instead, an interaction still takes place, albeit with different kinetics, suggesting that the mutations have a significant effect on TcoALD binding. Unfortunately, this could not be quantified by any interaction model. This indicates that the interactions taking place on the sensor chip surface are relatively complex. Indeed, based on the analytical SEC results, we suspect that multiple events occur simultaneously on the sensor chip surface. First, TcoALD occurs as a dimer in solution making it a bivalent analyte. Hence, this possibly leads to avidity effects on the sensor chip, whereby one TcoALD2 is able to bind two Nbs simultaneously. The use of an analysis model designed to take such effects into account was attempted [42], but this did not improve the fit. Second, since Nb474 binding promotes TcoALD2 tetramer formation, this would mean that, on the sensor chip surface, binding of a TcoALD2 to a Nb allows the subsequent recruitment of an additional TcoALD2 onto a formed Nb474-TcoALD2 complex. Moreover, the effect of the introduced mutations on the TcoALD dimer-tetramer equilibria is unknown. Generally, in such complex cases, the ‘analyte’ should be immobilized on the sensor surface to become ‘ligand’ and the ‘ligand’ should be used in the mobile phase to become ‘analyte’. However, employing the TcoALD variants as ligands is not an option as they do not survive the harsh regeneration condition used during the experiment (0.2% SDS). Given the complexity of the interactions on the sensor chip surface, we therefore prefer not to fit the data with any model to avoid overparametrization and misinterpretation of the real KD value describing the Nb474-TcoALD interaction. Hence, we interpreted the SPR data in a semi-quantitative manner. A determination of the dissociation affinity constants for the Nb474-TcoALDWT and Nb474-TcoALDA77E becomes possible when the SPR experiments are carried out according to the format of the homologous sandwich ELISA. Surprisingly, both interactions have very similar affinities (73.83 pM and 66.97 pM, respectively) despite the mutation of an Ala to a bulkier, charged Glu residue. This can be explained by a closer examination of the Nb474 paratope (S5A Fig). Nb474 contains a cavity, which is perfectly aligned with the position of Ala77 on the TcoALD epitope. Hence, given a local rearrangement, a Glu side chain could be easily accommodated. We hypothesize that Nb474 immobilized in an ELISA well or on a sensor chip surface has less conformational freedom to accommodate the Glu77 side chain, which leads to less efficient binding of TcoALDA77E. This explains why, during the SPR experiments, a 50-fold increase in analyte concentration was needed for TcoALDA77E compared to TcoALDAWT in order to reach the same binding signal. However, once bound, the Nb474-TcoALDA77E dissociation displays the same kinetics as for the Nb474-TcoALDWT interaction as evidence by the SPR data. In contrast, non-immobilized Nb474 has the conformational freedom to accommodate Glu77 on TcoALDA77E efficiently, thereby displaying very similar binding kinetics as observed for interaction with TcoALDWT. In the case of TcoALDL106Y and TcoALDA77E/L106Y, investigation of the homologous sandwich ELISA format with SPR reveals that non-immobilized Nb474 outcompetes immobilized Nb474 for antigen binding and thus washes the antigen off the Nb474-coated surface. Although the affinity constants for the Nb474-TcoALDL106Y and Nb474-TcoALDA77E/L106Y could not be measured directly, these observations suggest that these mutations weaken the Nb-antigen interaction. For the L106Y mutation, this can again be explained by examination of the structure. The presence of a Tyr residue at this position would disrupt the salt bridge between Asp106 of Nb474 and TcoALD Arg109 and Arg110 (S5B Fig). The A77E/L106Y double mutant most likely experiences a combined effect of both mutations, which is why TcoALDA77E/L106Y yields the lowest binding signals in all experimental set-ups. Together, the SPR and crystallographic data explain the results of the Nb474-based homologous sandwich ELISA. Compared to TcoALDWT, a low signal was observed for TcoALDA77E, whereas TcoALDL106Y and TcoALDA77E/L106Y could not be detected. While the Nb474-based immunoassay is highly specific for diagnosing T. congolense infections, our mutations studies imply that the detection of all T. congolense strains may not be guaranteed. In our previous work [16], we tested the Nb474-based ELISA on the sera of mice infected with different T. congolense strains of the Savannah subtype. While some infected sera gave rise to very high signals (T. congolense strains TC13, IL1180, Ruko 14cl3, and MF3cl2), others displayed low binding (T. congolense strains STIB68, TRT55, MF5cl4). It is difficult to assess whether these differences arise from i) varying expression levels of TcoALD between the distinct T. congolense strains, ii) the occurrence of mutations on the epitope recognized by Nb474 with effects similar to the A77E, L106Y, and A77E/L106Y mutations studied in this paper, or iii) a combination of both. Our results concerning the aldolase sequences within the Nannomonas subgenus seem to suggest the first hypothesis. The aldolase amino acid sequence conservation among all T. congolense subtypes tested in this work (Savannah, Forest, Kilifi) is very high (95.2%). Most importantly, the amino acids at positions 77 and 106 are relatively well conserved (Ala77 and Leu106 for Savannah and Kilifi subtypes; Val77 and Leu106 for Forest subtype). While the T. congolense Forest subtype contains a Val at position 77, this is not expected to severely impact detection in the Nb474-based immunoassay based on our findings. Given that Val and Ala are chemically and structurally much more similar than Glu and Ala, the Nb474-TcoALD interaction is likely to be much less perturbed by the Ala77Val than the Ala77Glu mutation. Hence, this would suggest that the Nb474-based immunoassay would detect all T. congolense infections. However, the potential occurrence of T. congolense strains carrying mutations that would escape detection in the Nb474-based ELISA is not unconceivable. This finding calls for an extensive and detailed molecular characterization of the different T. congolense strains and sequence their genomes. Finally, it is interesting to note that the pig-infective T. simiae and T. godfreyi parasites have an aldolase with an Ala77Glu and Leu106 genotype, suggesting that the Nb474-based could be employed to detect infections of these trypanosomes. The data presented here also provide insights into the practical set-up of the Nb474-based ELISA. The amounts of capturing and detecting Nb474 yielding the highest signal were determined using a checkerboard ELISA format without prior knowledge of the Nb474-target interaction and its affinity [16]. The outcome of this effort is shown as a heat map in Fig 6. The highest ELISA signal is obtained when relatively low amounts of both capturing and detecting Nb474 are used (~ 2 ng for both, respectively). In the case where the optimal amount of capturing Nb474 is kept fixed (~ 2 ng), any deviation (higher or lower) from the optimal 2 ng amount of detecting Nb474 reduces the intensity of the observed ELISA signal. A decrease results in less detecting Nb474 binding to the Nb474-TcoALD sandwich, which is why a reduction in signal intensity is observed. Based on the results presented in this paper, an increase in the amount of detecting Nb474 above the optimal 2 ng would enhance the “self-competition” or “washing” effect, which was exacerbated in the case of the TcoALDL106Y and TcoALDA77E/L106Y mutants. Likewise, in the case where the optimal amount of detecting Nb474 is kept fixed (~ 2 ng), any deviation from the optimal 2 ng of capturing Nb474 reduces the signal intensity in the ELISA. Employing relatively low amounts of capturing Nb474 is possible due to the high affinity and slow dissociation kinetics of the Nb474-TcoALD interaction as evidenced by the SPR data. A decrease in the amount of capturing Nb474 compared to the optimal case leads to less antigen being captured, which is why a reduction in signal intensity is observed. An increase in the amount of capturing Nb474 would enhance the avidity effects in the ELISA wells, whereby one TcoALD multimer would be able to bind several Nbs simultaneously. Hence, no TcoALD epitopes would be available for binding of detecting Nb474, which results in the observed lower ELISA signal with increasing amounts of capturing Nb474 (Fig 6). The main focus of this paper was to determine the molecular mechanisms underlying the high specificity of a Nb-based homologous sandwich ELISA that allows detection of T. congolense infections. As reported previously, the assay targets glycosomal aldolase [16]. While aldolase proteins are relatively well conserved throughout all domains of life, they seem to be immunologically sufficiently distinct, even within the same genus. In this and previous work [16], we have demonstrated that the Nb474-based sandwich assay specifically detects the presence of T. congolense aldolase, while this is not so for aldolase from other trypanosomes. Coincidentally, a similar finding has been documented for the differential diagnosis of Plasmodium infections. A monoclonal antibody-based immunoassay targeting malarial aldolase results in the specific detection of Plasmodium vivax infections, while remaining negative for samples containing Plasmodium falciparum [43]. These examples demonstrate that parasite-encoded aldolases are suitable biomarkers for the stringent detection of parasite infections. This may present an interesting research avenue for the development of immunoassays for the specific detection of other pathogens. The results presented in this paper indicate that, while the concept and use of such assays are relatively simple, their underlying biochemistry can be quite complex. This may be of particular interest to those developing similar assays.

10.1371/journal.pcbi.1003562

Atomistic Picture for the Folding Pathway of a Hybrid-1 Type Human Telomeric DNA G-quadruplex

In this work we studied the folding process of the hybrid-1 type human telomeric DNA G-quadruplex with solvent and ions explicitly modeled. Enabled by the powerful bias-exchange metadynamics and large-scale conventional molecular dynamic simulations, the free energy landscape of this G-DNA was obtained for the first time and four folding intermediates were identified, including a triplex and a basically formed quadruplex. The simulations also provided atomistic pictures for the structures and cation binding patterns of the intermediates. The results showed that the structure formation and cation binding are cooperative and mutually supporting each other. The syn/anti reorientation dynamics of the intermediates was also investigated. It was found that the nucleotides usually take correct syn/anti configurations when they form native and stable hydrogen bonds with the others, while fluctuating between two configurations when they do not. Misfolded intermediates with wrong syn/anti configurations were observed in the early intermediates but not in the later ones. Based on the simulations, we also discussed the roles of the non-native interactions. Besides, the formation process of the parallel conformation in the first two G-repeats and the associated reversal loop were studied. Based on the above results, we proposed a folding pathway for the hybrid-1 type G-quadruplex with atomistic details, which is new and more complete compared with previous ones. The knowledge gained for this type of G-DNA may provide a general insight for the folding of the other G-quadruplexes.

G-quadruplexes are high-order DNA/RNA structures. They are involved in the regulation of telomere maintenance, DNA replication, transcription and translation, and are also attractive drug designing targets for treating cancers and promising building blocks for molecular nanodevices. The knowledge of their formation process will improve our understanding of how they achieve their functional structures and then facilitate designing of artificial G-quadruplexes with novel functions. The study of their formation process is also of academic importance, since they involve many different physical chemical factors or interactions, including the hydrogen bonds, the electrostatic effect associated with metal ions, and the syn/anti reorientation of the glycosidic bonds. These make the G-quadruplex a fascinating model system for studying the structure formation of bio-molecules. Furthermore, the study of their formations may enrich the free energy landscape theory that has been well developed for protein folding, but yet to be verified in the other biomolecular systems. Here we computationally study the folding process of the hybrid-1 type human telomeric DNA G-quadruplex and infer a new folding picture, which may also cast a light to the formation of the other G-quadruplexes.

G-quadruplexes are high-order DNA or RNA structures formed from guanine-rich sequences, and their building blocks are G-tetrads that arise from Hoogsten hydrogen-bonding between four guanines. The G-tetrads stack on top of each other and form four-stranded helical structures. Bioinformatics analysis suggests that G-quadruplex motifs are prevalent in genomes. Recently, experimental evidence is accumulating for the in vivo presence of G-quadruplexes in DNA telomeres, in gene promoter regions [1], and even in messenger RNAs [2], [3], suggesting that they are involved in the regulation of telomere maintenance, replication, transcription and translation. G-quadruplexes are also attractive drug designing targets for treating cancers and platforms for delivering drugs [4]. Despite of their functional importance, the folding processes by which they achieve the functional structures have not been well understood as that of DNA and RNA duplexes [5]–[12]. It is believed that there are significant differences between G-quadruplexes and duplexes in the balance of forces, mainly the hydrogen bonds and electrostatic interactions [13]. Therefore, the study of the folding of G-quadruplex will improve our understanding of the balance between different forces in determining the structures and dynamics of such a typical folded oligonucleotide, and may facilitating designing new quadruplexes with novel functions. Moreover, the knowledge may enrich the energy landscape theory that has been well developed for protein folding, but yet to be verified in the other biomolecular systems. However, the folding of G-quadruplexes is a difficult problem due to its sensitivity to the terminal nucleotides, the dependence on ion types and concentration, and particularly due to the little known interplay between metal ions and folding dynamics; the syn/anti reorientations of the glycosidic bonds of the nucleotides further complicate the folding process. There are lots of experimental works on different forms of G-quadruplexes, studying their native structures, thermodynamical properties, folding kinetics and cooperativity, as well as the roles of ions in the stability and folding process. A detailed discussion of these works is beyond the scope of this article and can be found in several excellent reviews [13]–[19]. Recently, new progress has been made on the folding intermediates of DNA quadruplexes [20]–[23], particularly those achieved by single-molecular techniques including optic tweezers and magnetic tweezers [24]–[26]. For example, Wei et al. investigated the folding kinetics of human telomeric G-quadruplexes using magnetic tweezers and detected a G-triplex [25]; they also observed reversible transitions from the G-quadruplex to the G-triplex as well as from the G-triplex to the unfolded coil, and then suggested that the G-triplex is an in-pathway intermediate. Molecular modeling and simulations are able to complement experiments by providing much detailed information or insights [27]–[31]. For example, Sugiyama et al. systematically investigated the intermediates of human telomeric G-quadruplexes using ab initio calculations and MD simulations; the folding pathways and the roles played by ions were discussed [29], [30]. Limongelli et al. studied the folding of a 15-mer G-quadruplex using metadynamics; they identified a stable G-triplex and then validated it with a number of experiments [32]. Despite of many pioneer works, the atomistic picture for the folding pathways of quadruplexes is still lacking due to the temporal and spatial resolution limits of experimental techniques, the exclusion of conformation dynamics or entropies in theoretical analysis, or insufficient sampling of the phase space in previous all-atom computer simulations. In this work we studied the folding process of a 24-nt human telomeric DNA sequence (PDB ID 2GKU) (Figure S1) [33] with explicitly modeled solvents and ions using an advanced sampling technique and large-scale simulations. This sequence was selected since it forms a unique native structure of hybrid-1 type in KCl solution at room temperature and has many experimental results to be compared with [23], [30], [33], [34]. The folding time of this sequence was measured to be longer than 10 ms by stopped-flow and spectroscopic techniques [34], well beyond the timescale of traditional all-atom MD simulations. To overcome the barrier crossing problem, we combined the power of large-scale simulations and a novel advanced sampling technique named bias-exchange metadynamics, which is very efficient at accelerating barrier-crossing events by periodically modifying the effective energy felt by the system with small repulsive Gaussian potentials and thus enforcing the escape from local minima [35]. For a further acceleration of the sampling and increase of its coverage in the phase space, multiple (six) copies of metadynamics were run simultaneously with each biased on a different collective variable (CV) [36]. The conformations and velocities of different replicas were allowed to exchange periodically according to a metropolis criterion. From the data obtained by bias-exchange metadynamics, we calculated the free energy landscape, identified several intermediate states, and further studied their stabilities and dynamics by performing massive conventional MD simulations. Based on the above results we proposed an atomistic picture for the folding process of the hybrid-1 type G-DNA and discussed its relevance to the previous experimental and theoretical results. The convergence of the bias-exchange metadynamics was tested by monitoring the random walk of the replicas in their CV spaces, the exchange probability as a function of simulation time, and the evolution of FEL during simulation (Figure S3). For the four biased replicas, the CVs sampled all the possible values of Q and , and a broad region of dRMSD () and (); and the replica walked back and forth many times in the relevant space. These features indicated that the simulation sampled a sufficient large region of CV spaces. The number of successfully exchanged events was almost linear as a function of time for all replicas, showing that the exchange happened at a steady rate throughout the simulation. The average exchange probabilities were in the range of 4–5% for the four biased replicas and about 21% for the neutral replicas. The lower values for four biased replicas were expected since they were biased at different CVs and had very different energetics. The FELs were calculated solely from the neutral replicas to avoid potential problems from the applied biases in the other replicas. It was found that the general shape of the FELs did not change after , and the two FELs calculated respectively from two neutral replicas at were almost indistinguishable. Besides, the highest free energy barrier between basins was around several kcal/mol, reflecting a good sampling quality of the relevant phase space. The FELs at will be used for the following analysis. The free energy landscape shown in Figure 1 roughly manifests a diagonal shape, indicating the cooperativity between the formation of native contacts and the binding of metal ions. From the FEL six basins of attraction are identified and labeled from I to VI, respectively. Their representative structures are also shown in the figure, obtained based on a clustering analysis [37] of the belonging conformations, which are determined using their CVs. For the first basin it is found that the structures are pretty heterogeneous. For example, the largest cluster has a rather compact structure, i.e., the first two G-repeats ( and ) roughly form a hairpin, upon which docks the 3′ terminal via non-native interactions. The second and third largest clusters are both characterized by hairpins, however, formed between and and between and , respectively. Most stable interactions observed in the first basin are non-native ones, supported by the hydrogen bond map averaged on all the structures belonging to this basin (Figure S4). The ions binding on the G-DNA are weak, with the binding probabilities generally lower than 0.15. Besides, the binding probabilities are almost uniform on all nucleotides; there is no specific binding detected (Figure S4). Based on the above analysis, the basin-I is designated as the denatured state. The last basin (basin-VI) occupies a narrow area and is characterized by the highest values of Q and among all basins. Plus, clustering analysis showed that the belonging structures are homogenous and similar to the native one. Therefore the basin-VI is concluded to be the native state of the G-DNA. In addition to the basin-I and basin-VI, there are other four basins of attraction on the FEL. Obviously these are intermediate states and hold the key for understanding the folding process of the G-DNA. For a better characterization of these intermediates, we feel that a clustering analysis of the BEMD data is not accurate enough, since it is not trivial to determine the width of a basin and whether a structure belongs to that basin solely based on the CVs, due to possible overlaps between basins in a low-dimensional projection of the free energy landscape. Therefore we further performed multiple conventional MD simulations initialized from these intermediates. Such simulations are free of the above mentioned problems, and most importantly, they are able to provide true dynamics of the intermediates, which is lost in BEMD due to the added potentials. In the following sections we will discuss the structures and dynamics of the intermediates by combining the data from BEMD with that from conventional simulations. The structure of the intermediate-II is heterogeneous, mainly characterized by a well formed hairpin at 3′-terminal and an unstable hairpin at the 5′-terminal according to Figure 1. Conventional MD simulations initialized from the largest cluster confirmed such an observation. As shown by the hydrogen bond map in Figure 2(A) and the detailed structure in Figure 3(A), the intermediate-II is compose of a well-formed native hairpin between the G-repeats and (shorted as hereafter) and a non-native hairpin formed by the first two G-repeats via G9∶G3 and G10∶T1; and the interactions between two hairpins are ignorable. Dynamically, the structures are under constant fluctuations, with the RMSDs up to 1 nm with respect to their initial conformations. The fluctuations are mainly associated with relative motions between two hairpins (Figure S5 and Video S1). The consistence of the conventional simulations with the BEMD data suggests that the former has covered the most relevant phase space of the intermediate-II, although the initial structures were chosen only from the largest cluster, whose population was about 20% in this intermediate. Besides, it is interesting to note that is in an antiparallel conformation, in contrast to its parallel conformation in the native structure. The latter structure is probably not stable in this stage without the supporting from the nearby interactions, due to the tension associated with the parallel conformation and the reversal loop. The ion binding pattern of this structure shown in Figure 4(A) is similar to that of the denatured structures, i.e., the binding probabilities are low and almost distributed evenly on all nucleotides. There is no strongly binding sites observed. The intermediate-III is a native triplex composed of the last three G-repeats (denoted as hereafter), revealed by BEMD (Figure 1) and confirmed by multiple conventional MD simulations. As shown in Figure 2(B), there are two large groups of hydrogen bonds, including that between the G-repeat and (G21∶G17 and G22∶G16), and between and (G17∶G9 and G16∶G10). The detailed structure in Figure 3(B) consistently shows that is spatially close to , and is close to . The initial structures for running conventional MD simulations represent roughly 36% populations in the intermediate-III, however, the trajectories still cover a broad region of phase space (Figure S9). Conventional simulations demonstrate two different dynamics that drive the G-DNA toward different destinations. The first kind is a docking of the on the triplex, which essentially makes the structure transform into the intermediate-IV (Figure S9 and Video S2). Another dynamics is characterized by a flanking motion of the with respect to the triplex, constrained by a native base pair A20∶T1 and an non-native interaction G9∶G4 (Figure 2–3 and Video S3). These two interaction pull the first G-repeat close to the triplex so that it will not drift away from the triplex. The flanking motion keeps the G-DNA in the original basin and results in a fluctuating structure, which will eventually transform into the intermediate-IV via the first kind of dynamics described above. In the intermediate-III, metal ion binding pattern becomes interesting. As shown in Figure 3 and 4, a ion is trapped between the first and second G-tetrads, resulting in high ion binding probabilities of almost 90% of the nearby nucleotides. Clearly, the binding of a positive ion compromises the strong negative charges along the backbone and further stabilizes the base pairs by coordinating the O6 atoms of the nearby bases. The second ion between the second and third G-tetrad seen in the native structure is absent in the intermediate-III, therefore the nearby native base pairs G23∶G15 and G15∶G11 are hardly detectable, although the three nucleotides are almost in position (Figure 2(B) and Figure 3(B)). The structure of the intermediate-IV is characterized by a incomplete docking of on the triplex , supported by both BEMD and conventional simulations (Figure 1–3). The last nucleotide G5 in does not reach its correct position in the third G-tetrad but forms non-native hydrogen bonds with G9 instead. The G3 and G4 nucleotides in , however, bind correctly to the trapped ion in the central channel and their ion binding probabilities increase to about 80% from below 30% (Figure 4), leading to a basically formed quadruplex. At this folding stage, the lower ion binding site in the central channel is still unoccupied, although seven out of eight of its nearby nucleotides are in position. Dynamically speaking, the whole structure is very stable, indicated by the not-larger-than 0.25 nm RMSDs with respect to the initial structures (Figure S7). This dynamics is believed to be representative of that of the intermediate-IV since the initial structures for running conventional simulations represent 80% population of the basin. The intermediate-V is different from the preceding intermediates primarily in the trapped ion in lower site in the quadruplex channel (Figure 1). The trapped ion results in a further strengthening of the native base pairs and increase of ion binding probabilities of the nucleotides in the third G-tetrad (Figure 2–4). The structure is very similar to the native one and thus the intermediate should be viewed as a sub-state of the native basin of attraction. Indeed, we observed two direct folding trajectories from this intermediate to the native states in the conventional simulations (Figure S8 and S9). The folding of G-DNA is complex partially due to the involvement of the syn/anti reorientations of the glycosidic bonds. To reveal how such motions interplay with the folding process, we analyzed the syn/anti patterns and dynamics of the intermediates based on multiple conventional MD simulations. The torsion angle used to determine the syn/anti configurations for a specific glycosidic bond was calculated based on the following four atoms: O4′ and C1′ in the sugar ring, and N9 and C8 in the base. The results are shown in Figure 5 and in Figure S11, S12, S13, S14. It can be seen that in four intermediates the glycosidic bonds generally take correct syn/anti configurations when the corresponding nucleotides form native and stable base pairs with the others. Here by correct we mean the glycosidic bonds take the same configurations as in the native structure. However, fluctuating glycosidic bonds are also observed. For example, in the intermediate-II there are two nucleotides (G22 and G23) fluctuating between and configurations although they are within the basically formed 3′-terminal hairpin. Even after the structure transforms to the intermediate-III, G23 is still under fluctuation. The typical time scale for the syn/anti transitions is of order of ten nanoseconds, according to conventional MD simulations (Figure S11, S12, S13, S14). Interestingly, two nucleotides with wrong syn/anti configurations are observed although they have formed base pairs with others, which are G17 in the intermediate-II and G11 in the intermediate-III, tentatively attributed to their outer position in the formed structure and associated larger flexibility. From the intermediates-III to V, more and more stable base pairs are formed and the fluctuating bonds become fewer accordingly. In the last two intermediates, we also observed fluctuating bonds but no wrong syn/anti configurations. The combined power of bias-exchange metadynamics and large scale conventional MD simulations enabled us to explore the free energy landscape of the DNA G-quadruplex and the structure and dynamics of the intermediates. The relevance of the results described above to the previous experimental and theoretical data is discussed in the following sections. Recently, the existence of a triplex as a folding intermediate in several different quadruplexes has been established by many experimental approaches, including CD, DSC, and ITC analysis [20], FRET [23], optic tweezers [24], and magnetic tweezers [25]. However, the detailed structure of the triplex, particularly the binding patterns of the associated metal ions, is still unclear due to temporal and spatial resolution limits of experimental techniques. The triplex detected in our simulations (the intermediate-III) is relevant to that detected in previous experiments. For example, our triplex is characterized by a docking of on the hairpin , with the first G-repeat at the 5′ terminal fluctuating around. This structural feature has also been observed in the thermal denaturation experiments of several human telomere DNA sequences including Tel22 and 2GKU by Gray et al. [23], who found that these two DNAs have common unfolding pathways and the intermediate triplex states have greater solvent exposure of the 5′-segment. The folding/unfolding of Tel22 in the presence of ions has also been studied by another group using DSC and CD measurements; they confirmed the existence of a triplex as intermediate state and determined a release of 1.5 ions from the folded to the triplex states [20]. As a comparison, our calculation shows that the average numbers of bound ions in the triplex and in the native states are 1.2 and 3.0, respectively (Figure S10); the difference of 1.8 ions agrees quite well with the experimental value. Furthermore, Mashimo et al. systematically calculated the energies of various possible topologies of triplex using ab initio molecular dynamics and fragment molecular orbital method [30], and then for the type-1 quadruplex such as 2GKU they suggested a triplex that has a similar structure to ours. Therefore the triplex detected in our computations is relevant to previous experimental and theoretical ones. Moreover, our analysis provides more atomistic details on its structure, particularly in the patterns of metal ion binding. In the folding studies of the hybrid-1 type G-DNAs, the formation time and folding dynamics of the parallel conformation in the 5′-terminal and the associated reversal loop are always a myth. In previous literatures, it was often suggested that these local structures form at the end of the folding stage via a flip of the first G-repeat. However, there is little direct proof supporting this suggestion. Here thanks to the powerful BEMD, we observed two intermediates that provided insights into the underlying dynamics. In the structure of the intermediate-III, the first G-repeat is constrained by the interaction A20∶T1 and G9∶G4 in such a place, that only a flanking motion of with A20∶T1 as a pivot is needed to form the parallel conformation (Figure 3). In the conventional MD simulations started from this intermediate, we indeed observed two direct trajectories that transformed from the intermediate-III to IV with such a motion (Figure S6 and Video S2). As a result, in the intermediate-IV the parallel conformation and the reversal loop have been mostly formed (Figure 3). Therefore the formation of the local structures is not a once-for-all event occurring in the final folding stage, as often implicated in previous literatures. Instead, the formation starts early with the triplex (the intermediate-III), and is basically finished when the triplex transforms into the quadruplex (the intermediate-IV); and the final formation is accomplished after the trapping of the second ion in the central channel (the intermediate-V). We believe that the new picture can be easily verified by experiments, since it suggests that the A20∶T1 and G9∶G4 interactions play a key role during the transition from the triplex to the quadruplex. It is highly possible that a knockout of these interactions will significantly impede the formation of the reversal loop and slow down the folding rate of the G-DNA. Early folding events are also important for understanding the whole folding process [38], [39]. Previously, Mashimo and colleagues proposed that the type-1 quadruplex first folds into the hairpin and then to the triplex based on ab initio calculations and molecular simulations [30]. Although our work agrees with theirs on the formation of as an intermediate, it suggests a different initial structure, versus . To determine which structure is more kinetically connected to the triplex, we performed 10 high temperature unfolding simulations starting from the native structure (Figure S15). It was found that 8 of them unfold into structures containing , while only one into that containing . This may be attributed to the larger entropy of the partially formed structure containing , compared with that containing . Physically, the entropy of the latter is lower in that it has two spatially close strands of length 7- and 8-nt, respectively, and the excluded volume effect between them lowers the structural entropy; in contrast, the former has a long unpaired strand of length 14-nt and a free nucleotide A24; the excluded volume effect between them is obviously minimal. Besides, the hairpin has lower enthalpy, according to two additional simulations performed for the two hairpins (Figure S16). Therefore, it is more likely that the early folding of the quadruplex starts from the hairpin . The roles of non-native interactions in the folding process of G-DNA deserve further discussing. Before that, it is worth pointing out that in the research field of protein folding, non-native interactions are known to be important, particularly for the intrinsically disordered proteins. For example, Wang and colleagues studied the binding-induced folding of IA3, which is an intrinsically disordered protein that inhibits the yeast aspartic proteinase saccharopepsin by folding its own N-terminal residues into an alpha helix upon binding [40]. With their developed multi-scaled approach [41], [42], they found that the non-native interactions facilitate binding by reducing significantly the entropic search space in the landscape. Here in the folding of the G-DNA, the roles played by the non-native interactions were found to be similar. As described in the result section, the non-native interaction G9∶G4, together with the native A20∶T1, pull the first G-repeat close to the triplex and so that it will eventually dock on the triplex. Without these interactions, the first G-repeat may drift away and has to search in a much larger phase space. The above arguments can be easily verified by an experiment that measures the folding rates of the G-DNAs mutated on the corresponding nucleotides. The structural formation and binding of metal ions are cooperative during the whole folding process. Physically, the effects of trapping of cations in the central channel of the quadruplex are twofold. First, the trapped cations compromise the strong negative charges of the backbone and facilitate their approaching to each other. Second, the metal ions are able to coordinate the O6 atoms of the nearby bases thus bridge the interactions between them. According to our simulations, the total number of bound ions increases monotonically from the intermediate II to V (Figure S10). In each intermediate, the formed base pairs need the binding of cations to strengthen their stabilities. For example, in the triplex structure of the intermediate-III, although G11, G15, and G23 are almost in their native position, they do not form stable base pairs according to Figure 2, as is correlated with the absence of the second ions in the central channel. This feature is more clear in the intermediate-IV, where the above three nucleotides become even closer while the base pairs between them are still minimal (Figure 2 and 3), attributed to the same reason. Only after the G-DNA proceeds to the intermediate-V, the second ions is trapped in the central channel and then the surrounding base pairs become significantly stable. From another point of view, the trapping of cations also needs the formation of the local structures. This can be seen in the intermediate-IV, probably due to the lack of the protection from G5, the second ion is able to leak out of the channel from the bottom (Figure 3) and thus cannot be trapped there stably (Figure 4). Therefore it is concluded that the folding and binding of ions are cooperative and mutually supporting each other. The syn/anti reorientations are among the most important factors that affect the folding rate of G-DNAs. There are two different syn/anti reorientation dynamics according to our simulations. In general, the glycosidic bonds either stay at the correct configurations if the corresponding bases form native pairs with the others, or keep fluctuating if the bases are relatively free. In other words, in the correctly formed native structural elements defined by base pairs and backbone arrangements, wrong yet persistent syn/anti configurations are seldom observed, possible due to the steric inconsistence between local backbone arrangements and wrong syn/anti configurations. This feature is consistent with a previous work by Sugiyama et al., who systematically studied all the possible loop conformations as well as the syn/anti arrangements for type-1 and type-2 quadruplexes using ab initio molecular dynamics and fragment molecular orbital calculations, and found that all the intermediate states leading to the native structure have correctly arranged syn/anti configurations [30]. Another support came from a recent simulation on G-DNAs by Šponer's group [43], where they concluded that for folding to a specific G-DNA topology in a single molecular event, the molecule must have an appropriate combination of syn/anti nucleotides, otherwise the likely result will be a misfolded structure. However, exceptions to the above pictures do exist according to our simulations. In the early folding intermediate-II and III, two nucleotides with wrong syn/anti configurations are observed although they form native pairs with the others (Figure 5). The exception may be explained by the outer positions of the nucleotides in the tertiary structure and the associated lacking of additional supports from nearby nucleotides or bound ions. Consistent with this argument, when the G-DNA folds to the intermediates-IV and further to V, more and more stable base pairs form and no wrong syn/anti configurations are observed any more. Caution should be given regarding the limitations of the present simulation. First, although the current simulation detected intermediates only having less seriously wrong syn/anti patterns, mainly in the early folding stage and in the outer positioned nucleotides, it could not rule out the existence of other type of intermediates with most glycosidic bonds in wrong syn/anti configurations, since none of the replicas in BEMD was biased on the glycosidic bonds to enhance sampling on the relevant phase regions. Second, even if the existence of such intermediates could be ruled out by future computations, the syn/anti reorientations would still play an important role by significantly retarding the folding rate, since the molecule has to explore different combinations of syn/anti configurations in a much larger phase space to find the right bottle neck leading to the native states. Third, although the BEMD simulation was shown converged here, it should be noted that the convergence was subjected to the present setup of the applied CVs. In another word, the present simulation does not preclude that there is an orthogonal CV that samples intermediates not detected by the present setup. To finally confirm that the folding intermediates detected here are the only true intermediates would require significantly more expensive unbiased simulations and/or metadynamics with alternative and independent CVs. It is also worth mentioning that we had tested many different combinations of CVs and performed several times of BEMD simulations; the present four CVs were not chosen randomly. A much more large-scale unbiased simulation for the specific DNA is being prepared. It is of particular interest to make a qualitatively comparison of the folding of quadruplexes with that of proteins. It seems that the former is more complex, since even for this small G-DNA of 24 nucleotides, four intermediate states have been identified. While in proteins, two-state folding is frequently observed for small globule proteins. Whether this is due to the particular topology of the quadruplex or the balance of interactions is not known yet. It is also interesting to characterize the main feature of the energy landscape of the G-quadruplexes and see if the energy funnel theory applies for these molecules. To this end, a topology-centered coarse-grained model of DNA quadruplex may be of help. The folding of quadruplexes is also complicated by the indispensable cooperation with metal ions, since the strong negative charges associated with the nucleotides have to be compensated by cations from solvents. According to our simulations, the metal ions progressively bind to the DNA as the quadruplex builds up, suggesting that the two process are cooperative. At last, the complexity is further increased by the involvement of the syn/anti reorientations of the glycosidic bonds, which increase the searching space and may also trap the G-DNA in some local minima. In summary, enabled by the combined power of the bias-exchange metadynamics and large-scale conventional MD simulations, we studied the folding process of a hybrid-1 type human telomeric DNA G-quadruplex. We obtained for the first time its folding free energy landscape and identified several intermediates. Further analysis of these intermediates showed that the structure formation and metal ion binding are cooperative and mutually supporting each other. The roles of the syn/anti reorientations in the folding process were also investigated. It was found that the nucleotides already taking their native positions usually have correct syn/anti configurations. However, intermediates with wrong syn/anti configurations were also detected, particularly in the early folding stages. Based on the above results, we suggest a new atomistic folding picture for the G-DNA, as shown schematically in Figure 6 and described as follows. The G-DNA first forms a hairpin from the 3′-terminal, on which the second G-repeat docks, accompanied by the trapping of the first ion in the central channel. The result of the docking is a triplex. At this folding stage the first G-repeat is constrained nearby the triplex by both native and non-native interactions and fluctuating around the triplex. After the first G-repeat docks upon the triplex eventually, an incomplete quadruplex forms, and the reversal loop also basically forms at this stage. However, the second binding site in the central channel is yet to be occupied, and therefore the third G-tetrad is somewhat unstable. After another ion is trapped inside the channel, the whole quadruplex is strengthened and the folding is completed. We believe this is a more detailed and complete picture compared with previous ones, and it represents a step forward in understanding the folding of the hybrid-1 type G-DNA. The knowledge gained here may also provide insights into the structure formation processes of the other types of DNA G-quadruplexes. In the preparation of the simulation system, we solvated the PDB structure (2GKU) within a box of 6087 TIP3P water molecules and added 3 and 24 ions to achieve charge neutral and an equivalent ion concentration of . The amber99sb_parmbsc0 force field was used, which combined the amber99sb force field with new parmbsc0 nucleic acids torsions [44]. The ion parameters were taken to be their defaults values in the force field, which are (sigma) and (epsilon) for , and (sigma) and (epsilon) for . The electrostatic interaction was treated using PME with a cutoff of . The same cutoff was used in the calculation of the van der Waals (VDW) interactions. All bonds were constrained using the LINCS algorithm and the MD time step was set to . Berendsen algorithm was used for both temperature and pressure coupling. All simulations were performed with Gromacs (version:4.5.3) [45] and its plugin PLUMED (version 1.3) [46]. The whole system was first subjected to a minimization of 1000 steps and then an equilibrium run with a NPT ensemble at 1atm and for 2 nanoseconds for a preparation of the initial structure. After that a long conventional MD simulation of length was performed started from this structure, in order to check the stability of the system setup and the native structure. It was found that during the simulation the fraction of native contacts (Q) was always higher than 0.92, all the native hydrogen bonds formed well, and the binding probabilities of ions on 12 native sites were very close to unity, showing that the native structure is stable under the force field. The details of the trajectories are given in Figure S2. The folding time of the specific DNA studied here is well beyond the timescale of traditional all-atom MD simulations. To overcome the barrier crossing problem, we adopted the bias-exchange metadynamics. In metadynamics, the overall external Gaussian potential acting on the system at time is given by(1)where is the value taken by the Collective Variables at time , is the Gaussian height, the Gaussian width, and determines the frequency of adding Gaussian potentials. The basic assumption of metadynamics is that after a sufficiently long time provides an estimate of the underlying free energy:(2) The bias-exchange metadynamics was run at with four copies biased on four different CVs, respectively, as well as two neutral replicas without any bias. Four CVs included the fraction of native contacts formed between the 12 guanines (Q), the dRMSD of the backbone (C4* atoms) with respect to the native structure, the number of binding/contacts between ions and the O6 atoms of the guanine bases (), and the radius of the gyration (). The parameters used for calculating these CVs were taken to be their default values in PLUMED [46]. The replicas were allowed to exchange their conformations and velocities periodically according to a metropolis-like criterion to further speed up the barrier crossing process. The criterion was given by(3)where and were the coordinates of walker and , respectively, and was the metadynamics potential acting on the walker . Among all replicas five of them were started from the native structure, while one neutral replica was started from an extended structure obtained from unfolding simulations at a high temperature. During the BEMD run, the conformations and velocities of different replicas were exchanged periodically according to a metropolis criterion. The height of the repulsive Gaussian potentials were and their widths were set to 2.5/130, 0.02 nm, 0.5, and 0.2 nm, for Q, dRMSD, , and , respectively. Note that the above number 130 is the total number of native contacts. The deposition rate of the Gaussian potentials is . The attempting frequency for replica exchanges was set to . The overall simulation time of the metadynamics was 4.2 microseconds, with each replica lasting for . The convergence of the calculation was shown in Figure S3 and discussed in the main text. To reveal the structures of the intermediates as well as their dynamics, we resorted to additional conventional MD simulations. For each of the four intermediates, we randomly selected three structures in the largest cluster obtained by a clustering analysis, and then for each structure we performed 100 ns MD simulations. The system setup, the force field and the parameters for running conventional simulations were the same as described above. There were in total 12 trajectories and the overall simulation time was 1.2 µs. The details of the trajectories are given in the Figure S5, S6, S7, S8, S9, S10, S11, S12, S13, S14. We adopted a simple algorithm to cluster the conformations obtained in MD simulations. We compared the -th frame in the trajectories with the representative structures of the clusters obtained previously one by one; if a dRMSD smaller than a threshold was detected, the -th frame was deemed belonging to the corresponding cluster; if the -th frame did not belong to any existing clusters, it was assumed to be the representative structure of a new cluster. The threshold for determining if two structures belonged to the same cluster was set to be 0.3 nm in all analysis except in the clustering of high-temperature unfolding trajectories described later, where the threshold is set to be . Unfolding simulations were started from the native structure and performed at 1atm and to enhance the barrier crossing events. The system setup was the same as described above. In total 10 such simulations were performed with each lasting for . After the simulations were finished, we performed clustering analysis to get the unfolding pathways, as shown in Figure S15. We performed two additional MD simulations for two hairpins and to compare their stability. For the hairpin , we chopped the hairpin fragment from the native structure starting from A14 to A24 and deleted the other nucleotides; the remaining length was 11-nt. Similarly, for we retained the structure from A8 to T18; the remaining length was also 11-nt. For each hairpin, we solved it in explicit waters and added ions to achieve charge neutral and the same concentration. The obtained two systems had the same number of ions and almost the same number of water molecules (3,928 versus 3,930). Both hairpins were restrained to their native structures with weak harmonic potentials. The MD simulations were performed at and 1atm for each. After the simulations were finished, we calculated the enthalpy of the hairpins by excluding the restraint energy, as shown in Figure S16.

10.1371/journal.pgen.1002096

The ISWI Chromatin Remodeler Organizes the hsrω ncRNA–Containing Omega Speckle Nuclear Compartments

The complexity in composition and function of the eukaryotic nucleus is achieved through its organization in specialized nuclear compartments. The Drosophila chromatin remodeling ATPase ISWI plays evolutionarily conserved roles in chromatin organization. Interestingly, ISWI genetically interacts with the hsrω gene, encoding multiple non-coding RNAs (ncRNA) essential, among other functions, for the assembly and organization of the omega speckles. The nucleoplasmic omega speckles play important functions in RNA metabolism, in normal and stressed cells, by regulating availability of hnRNPs and some other RNA processing proteins. Chromatin remodelers, as well as nuclear speckles and their associated ncRNAs, are emerging as important components of gene regulatory networks, although their functional connections have remained poorly defined. Here we provide multiple lines of evidence showing that the hsrω ncRNA interacts in vivo and in vitro with ISWI, regulating its ATPase activity. Remarkably, we found that the organization of nucleoplasmic omega speckles depends on ISWI function. Our findings highlight a novel role for chromatin remodelers in organization of nucleoplasmic compartments, providing the first example of interaction between an ATP-dependent chromatin remodeler and a large ncRNA.

Chromatin structure and function are regulated by the concerted activity of covalent modifiers of chromatin, nucleosome remodeling factors, and several emerging classes of non-coding RNAs. ISWI is an evolutionarily conserved ATP-dependent chromatin remodeler playing essential roles in chromosome condensation, gene expression, and DNA replication. ISWI activity has been involved in a variety of nuclear functions including telomere silencing, stem cell renewal, neural morphogenesis, and epigenetic reprogramming. Using an in vivo assay to identify factors regulating ISWI activity in the model system Drosophila melanogaster, we recovered a genetic interaction between ISWI and hsrω. The hsrω gene encodes multiple non-coding RNAs (ncRNAs), of which the >10 kb nuclear hsrω-n RNA, with functional homolog in mammals, is essential for the assembly and organization of hnRNP-containing nucleoplasmic omega speckles. These special nuclear compartments play essential roles in the storage/sequestration of hnRNP family and other proteins, thus playing important roles in mRNA maturation and other regulatory processes. Here we show that the hsrω-n ncRNA interacts in vivo and in vitro with ISWI to directly regulate its ATPase activity. We also provide in vivo data showing that omega speckle nuclear organization depends on ISWI function, highlighting a novel role for chromatin remodelers in nuclear speckles organization.

ISWI, the catalytic subunit of several ATP-dependent chromatin remodeling complexes, is highly conserved during evolution and is essential for cell viability [1]. ISWI-containing complexes play central roles in DNA replication, gene expression and chromosome organization [2]. ISWI uses ATP hydrolysis to catalyze nucleosome spacing and sliding reactions [1]. Loss of ISWI function in Drosophila causes global transcription defects and dramatic alterations in higher-order chromatin structure, including decondensation of chromosomes [3], [4]. In vitro and in vivo studies in several model organisms have also shown the involvement of ISWI-containing complexes in other nuclear functions like telomere silencing, stem cell renewal, neural morphogenesis and epigenetic reprogramming during nuclear transfer in animal cloning [2], [5], [6]. The diverse functions associated with ISWI depend upon the ability of other cellular and nuclear factors to regulate its ATP-dependent chromatin remodeling activity [7]–[9]. In order to identify new regulators of ISWI function, we developed in vivo assays to identify genetic interactors of ISWI in D.melanogaster [10], [11]. Using an eye-based assay to identify factors antagonizing ISWI activity, we recovered, among other genes, a genetic interaction between ISWI and hsrω [10]. The hsrω gene is developmentally expressed in almost all cells types and is one of the most strongly induced heat shock genes in flies [12]–[14]. The hsrω locus encodes multiple non-coding RNAs (ncRNA), of which the large >10 kb nuclear species (hsrω-n) is essential for the assembly and organization of the hnRNP-containing omega speckles [14]. These specialized nuclear compartments are distinct from other nuclear speckles and are localized in the nucleoplasm close to chromatin edges [14]. Omega speckles play essential roles in storage and sequestration of members of the hnRNP family and other proteins involved in RNA processing and maturation in normal as well as environmentally or genotoxically stressed cells (for a list of hsrω interactors see [13]–[15]. Here we show that the hsrω ncRNA interacts in vivo and in vitro with ISWI to directly regulate its ATPase activity. Additionally, we provide in vivo data showing that omega speckle nuclear organization depends on ISWI function. Our work thus suggests that ISWI and the omega speckle associated hsrω ncRNAs interact and reciprocally affect each other's activities. Our findings highlight a novel role for chromatin remodelers in organization of nuclear speckles. Loss of hsrω function by RNAi [15] results in a striking amelioration of morphological defects in eyes exclusively composed of ISWI-null mitotic clones (Figure 1A, 1B and Figure S1A–S1D, S1J; Table S1A). Mutations in the sqd gene, which encodes the Squid hnRNP, a component of omega speckles, also suppresses ISWI mutant eye defects (Figure S1F–S1I and S1K; Table S1A) [10]. Absence of ISWI in larval salivary gland cells causes a dramatic decondensation of the male X polytene chromosome [4]. Remarkably, hsrω-RNAi suppresses the ISWI null male X chromosome condensation defects as well (Figure 1C, 1D). Tissue-specific mis-expression of the catalytically inactive ISWIK159R transgene also produces eye phenotypes and global chromosome decondensation [3], [4], [11]. Silencing of hsrω-n activity not only suppresses ISWIK159R eye phenotypes (Figure 1E–1H) but also restores normal polytene chromosome condensation (Figure 1I, 1J). In agreement with the above observations, the larval lethality of ISWI-null mutants is also partially suppressed by hsrω-RNAi (Figure 1K; Table S1B), strongly indicating that reduction of hsrω-n transcripts improves ISWI activity. On the other hand, over-expression of hsrω through the hsrωEP93D allele [15] antagonizes ISWI activity, resulting in enhanced chromosome condensation defects and eye phenotypes in ISWI-null background (Figure 1L and Figure S1E; Table S1A). The suppression of chromosome condensation and eye defects in ISWI nulls by hsrω-RNAi is not due to a reduction in the efficiency of the GAL4/UAS driving system used to produce the ISWI-null and ISWIK159R mutant phenotypes (Figure S2A, S2B). Furthermore, the effect of hsrω-RNAi is highly specific for the loss of ISWI function (Figure S2C, S2D). Given the role played by omega speckles in nuclear RNA processing [13], we also checked if the levels of ISWI or ISWIK159R proteins and their corresponding mRNA were affected by hsrω-RNAi, which could account for the suppression of ISWI-null or ISWIK159R defects. However, depletion of hsrω transcripts by RNAi does not detectably affect ISWI protein or mRNA levels in either of these cases (Figure S3). In order to understand the molecular basis of the specific suppression of ISWI phenotypes by hsrω-RNAi, we examined the distribution and organization of omega speckles in the ISWI mutant third instar larval Malpighian tubule nuclei, which show abundant omega speckles using either RNA∶RNA in situ hybridization to hsrω-n ncRNA or immunostaining for some of the omega speckle associated hnRNPs [14]. Interestingly, the organization and distribution of omega speckles in ISWI mutants is profoundly altered when compared with wild type cells. Instead of typical speckles, the hsrω-n transcripts form “trail”-like structures in ISWI-null mutant nucleoplasm, indicating a severe defect in their maturation or organization (Figure 2A, 2B). Interestingly, Squid, NonA and other omega speckle associated hnRNPs also form “trail”-like structures in ISWI mutants (Figure 2C–2F, Figure 3A–3D, and Figure S4), which shows that distribution of not only the hsrω-n ncRNA but also of the omega speckle-associated hnRNPs is compromised in ISWI mutant nuclei. As shown earlier [15], the omega speckles do not form in the absence of hsrω-n transcripts and the omega speckle-associated hnRNPs remain diffused in the nucleoplasm (Figure 3E–3F). Interestingly, when the ISWI as well as hsrω-n ncRNA are absent, omega “trails” are not formed (Figure 3G–3H), strongly indicating that ISWI mutant specific omega “trails” are dependent on the presence of the hsrω-n ncRNA. Analysis of live cells expressing a SquidGFP transgene [16] clearly identifies the GFP-positive “trails” in live ISWI mutant cells similar to those observed in fixed cells (Figure S5). This shows that the ISWI omega “trails” are not a fixation artifact. Significantly, comparable hsrω RNA “trails” were not seen (Figure S6) in the presence of other mutants like jil1, ada2 and gcn5 which also display chromosome condensation defects similar to those observed in the ISWI mutants [17], [18]. This excludes the possibility that the omega “trails” in ISWI mutant nuclei result from a “squeezing” effect of the nucleoplasm due to a massive “fallout” of chromatin associated proteins following global chromosome decondensation. Studies in several model organisms have shown that ISWI plays a global role in transcriptional activation as well as repression [1], [3], [4]. Therefore, we examined if ISWI mutation altered the levels of hsrω-n ncRNA or the omega speckle-associated proteins. However, no significant difference in their levels was found between ISWI mutant and wild type cells (Figure S7). The >10 Kb hsrω-n ncRNA that organizes the omega speckles contains a small 0.7 Kb intron [14], [19]. It has been recently observed [20] that a spliced form of the hsrω-n transcript is also associated with the omega speckles. Therefore, we checked if the ISWI mutant condition affects splicing of this RNA which may result in the “trail”-like organization. RT-PCR and Northern blot analyses clearly indicate that ISWI mutation does not affect splicing of the hsrω-n ncRNA (Figure S8A–S8C). In light of the significant role played by ISWI in gene expression, we checked whether an engulfment of the nuclear RNA export machinery in ISWI mutants affected RNA transport from nucleus, which in turn could modify the omega speckles into “trails”. In situ hybridization to cellular RNA with poly-dT probe did not reveal any difference in the nuclear vs cytoplasmic distribution of poly-A RNAs between wild type and ISWI mutant cells (Figure S8D). Thus, ISWI mutant nuclei do not seem to have a general RNA export defect, which could have been responsible for the observed omega “trails”. Omega speckles are thought to provide a dynamic system to sequester and release specific RNA processing factors in normal as well as stressed cells [13]. Following heat shock, hsrω is one of the most highly transcribed genes [13], [21] and omega speckles coalesce into bigger growing clusters that finally get restricted to the hsrω gene locus, providing a dynamic sink for proteins that need to be transiently withdrawn from active nuclear compartments under stress conditions [14]. As already noted above, ISWI mutant condition causes the omega speckles to form nucleoplasmic “trails” in unstressed cells (Figure S9A, S9B). Although heat shock caused clustering of the omega speckles or “trails” in wild type and ISWI mutant cells, respectively, the numbers of clusters in the latter cells were much less (Figure S9C, S9D), suggesting that speckle dynamics under heat shock is also compromised because of ISWI mutant background. Finally, the “trail”-like organization of hsrω ncRNA and its associated proteins in ISWI mutants is not limited to Malpighian tubule or salivary gland polytene cells (Figure 2A, 2B and Figure S10A, S10B), but is also observed in ISWI mutant diploid cells (), indicating that disorganization of omega speckles is a general consequence of loss of ISWI function. Unlike the association of ISWI with different bands and interbands on polytene chromosomes [4], [11], the hsrω-n ncRNA localizes in the nucleoplasm in proximity or at the edges of chromosome spreads, without any apparent overlap with the chromatin associated ISWI (Figure 4A, 4B). However, examination of confocal images of intact nuclei revealed some chromosome-nucleoplasm sites where ISWI and the hsrω-n ncRNA are adjacent and seem to form connecting bridges between nucleoplasm and chromatin (Figure 4C, 4D). Barring a few exceptions, Squid and other omega speckles associated hnRNPs also showed no overlap with ISWI on polytene chromosome spreads (Figure 4E, 4F and Figure S11A, S11B). Significantly, like the hsrω ncRNA they too were found to partially overlap with ISWI in several nucleoplasmic foci in intact nuclei (see Figure 4G, 4H and Figure S11C, S11D), suggesting that ISWI may indeed partially interact directly or indirectly, at least transiently, with omega speckles in the three-dimensional nuclear space. In order to directly investigate whether the chromatin remodeling factor ISWI physically interacts with omega speckles, we used an affinity purified ISWI antibody [4] to conduct classic RNA immunoprecipitation. Our semi-quantitative RT-PCR analysis revealed the presence of hsrω-n ncRNA in larval nuclear extracts immunoprecipitated with ISWI antibody (Figure 5A, 5B). To rule out a non-specific association of ncRNAs with ISWI, we used the same immunoprecipitate to detect U4 and Rox1 [22] ncRNAs by RT-PCR. Significantly, neither of these two otherwise abundant ncRNAs were detectable (Figure 5A, 5B) in the mmunoprecipitate. This confirms the specificity of the physical interaction between ISWI and hsrω-n RNA in native larval extracts. Further, to exclude the possibility that the physical association observed between ISWI and hsrω was due to fortuitous interactions occurring during nuclear extract preparation, we conducted the CLIP assay (Cross-Linking & Immuno Precipitation) using the affinity purified anti-ISWI antibody [4] on fixed larval nuclear extracts. The CLIP data confirmed a highly specific interaction between ISWI and the hsrω ncRNA in the nucleus (Figure 5C), as observed with the native extracts (Figure 5B). Moreover, as shown in Figure 5D, RNA pull down assay confirmed that ISWI is also specifically pulled down by immobilized hsrω-n ncRNA along with the other known omega speckles associated hnRNPs [13] while a control generic RNA does not pull down ISWI or the other hnRNPs (Figure 5D). Classic gel shift assay using in vitro transcribed hsrω-n ncRNA repeat unit (280b) and full length recombinant ISWI clearly shows that ISWI effectively retards hsrω-n ncRNA mobility, but that of a generic control RNA is retarded poorly (Figure 5E). Moreover, the mobility shift of the hsrω-n RNA by ISWI binding is specifically competed by hsrω-n but not by a generic RNA (Figure 5F). This further confirms the specific nature of ISWI/hsrω physical interaction in vitro. A functional significance of the physical interaction between ISWI and hsrω-n ncRNA is indicated by the stimulation of ISWI ATPase activity. Remarkably, as also reported previously [23], while the generic control RNA very poorly stimulates the ISWI ATPase activity, the hsrω-n ncRNA was found to specifically stimulate the ISWI ATPase activity to levels greater than those normally seen with DNA but lower than nucleosome-stimulation (Figure 5G) [23]. The 280b hsrω-n nuclear ncRNA repeat unit used for the binding and ATPase assays is predicted to organize into a stable double stranded RNA molecule containing a few loops (Figure S12). This secondary organization is common to many RNAs, but this structure is also reminiscent of a double stranded DNA molecule. Therefore, it remained possible that the recognition of a double stranded nucleic acid (RNA or DNA) may provide a basis for the observed binding and stimulation of ATPase activity of ISWI by the hsrω-n ncRNA. When we checked the ability of the double stranded DNA sequence encoding the hsrωncRNA to elicit ISWI ATPase activity, we found that ISWI was stimulated to levels similar to those reported for other generic linear double stranded DNA molecules [23] (Figure S13A). Furthermore, co-presence of hsrω-n ncRNA and nucleosomes in a classic ATPase assay with ISWI clearly shows that both substrates compete for ISWI binding and its ATPase activity stimulation (Figure S13B). The ISWI protein has two functional domains (Figure 6A), the N-terminal (ISWI-N) ATPase domain and the C-terminal (ISWI-C) nucleosomal DNA recognizing domain [24]. Results presented in Figure 6B and 6C, show that the hsrω-n binds with the ISWI-N fragment and stimulate its ATPase activity, suggesting that ISWI could interact with hsrω-n ncRNA through its ATPase domain. Therefore, we further checked if the presence of ATP, ATP-γ-S (a non-hydrolizable form of ATP) or ADP could affect ISWI binding or determine a conformational change in the ISWI/hsrω complexes resolved by gel shift. Our data show that all the three nucleotides have no effect on ISWI binding (Figure 6D), probably suggesting that the ATPase activity of ISWI may not be necessary for physical interaction between ISWI and hsrω RNA. Factors that coordinate nuclear activities occurring on chromatin and the nucleoplasmic compartments remain unidentified and uncharacterized. Therefore, an important open question in nuclear organization field is how nuclear speckles localize and organize themselves near transcriptionally active genes to cross talk with chromatin factors for processing of the nascent RNAs. Our data indicate that ISWI may provide a functional ‘bridge’ between chromatin and nuclear speckle compartments. Indeed, ISWI can directly or indirectly contact the omega speckles in intact nuclei, through hsrω-n ncRNA or some of the associated hnRNPs. Our confocal analysis suggested a functional ‘bridge’ between a chromatin factor (ISWI) and nucleoplasmic omega speckle components (hsrω ncRNA and hnRNPs). However, not all omega speckles show partial overlap with ISWI. Indeed, these molecular “bridges” between chromatin and nucleoplasm are probably transient, since time-lapse movies on live cells with fluorescently tagged chromatin and omega-speckle components clearly show very high mobility of these speckles (see Video S1), which probably may explain the absence of classic co-localization between ISWI and omega speckle components. The observed direct physical interaction between ISWI and hsrω-n ncRNA together with the stimulation of ISWI-ATPase activity in light of the partial overlap revealed by confocal microscopy suggests that ISWI may interact with hsrω-forming speckles only transiently, probably to help the hsrω ncRNA to properly associate with or release the various omega speckle-associated hnRNPs. Loss of ISWI may impair the correct maturation, organization or localization of omega speckles resulting in the observed omega “trail” phenotype. Our data also provide a possible explanation for the suppression of ISWI defects by hsrω-RNAi. In ISWI mutants carrying normal levels of hsrω transcripts, the limited maternally derived ISWI [3] is shared between chromatin remodelling and omega speckle organization reactions so that its sub-threshold levels in either compartments severely compromises both functions (see Video S2). However, when hsrω transcript levels are reduced by RNAi in ISWI null background, most of the maternal ISWI may become available for chromatin remodelling reactions, so that a minimal threshold level of chromosome organization can be achieved. This would permit initiation of close to normal developmental gene activity programs resulting in suppression of the ISWI eye and chromosome defects or in the postponement of the larval lethality to pupal stage. Additionally, it is known that when hsrω ncRNA is down regulated through RNAi, levels of free hnRNPs and other chromatin factors (i.e. CBP) are also elevated [25]. Therefore, we cannot formally exclude the possibility that these changes may also counteract ISWI defects by as yet unknown mechanisms. Our work provides the first example of modulation of an ATP-dependent chromatin remodeler by a ncRNA, and to our knowledge the first in vivo and in vitro demonstration of a role of chromatin remodeler in organization of a nuclear compartment. However, the mechanism underlying stimulation of the ATPase activity of ISWI by the hsrω-n ncRNA, which may facilitate the organization of omega speckles, remains to be understood. Given the evolutionary derivation of the ISWI ATPase-domain from RNA-helicase-domains [1], a provocative hypothesis is that ISWI could “remodel” speckles by structurally helping the assembly or release of specific hnRNPs with the hsrω-n ncRNA to generate mature omega speckles. Chromatin remodelers, nuclear speckles and their associated long ncRNAs are emerging as essential components of gene regulatory networks, and their deregulation may underlie complex diseases [15], [25]–[27]. The functional homology of the human noncoding sat III transcripts with the Drosophila hsrω ncRNA [13], [27], highlights the relevance and translational significance of studies unraveling the functional connections between ncRNA-containing nuclear compartments and chromatin remodelers. Flies were raised at 22°C on K12 medium [28]. Unless otherwise stated, strains were obtained from Bloomington, Szeged or VDRC (Vienna Drosophila RNAi Center). Genetic tests for dominant modifier (enhancement or suppression) of ISWI-EGUF and ISWIK159R phenotypes were conducted as previously described [10], [11]. The tissue specific expression of the UAS-ISWIK159R [4], the UAS-hsrωRNAi3 and the EP93D transgenic lines [15] was obtained with ey-GAL4 (for eyes and larval salivary glands) or Act5C-GAL4 driver (for larval Malpighian tubules and testis). The surface architecture of adult eyes was examined by the nail polish imprint method [26]. For the larval lethality assay, numbers of larvae of different genotypes that pupated and the numbers of pupae emerging as flies in a given cross were separately counted. Mouse monoclonal antibodies against the following proteins were used at the indicated dilutions: Hrb87F (P11) [14] dilution 1∶5 for IF and 1∶100 for WB; Squid (1B11) [29] dilution 1∶100 for IF and 1∶2000 for WB; NonA [30] dilution 1∶50 for IF and 1∶1000 for WB; PEP [31] dilution 1∶2000 for WB. Affinity purified rabbit ISWI antibody [4] was diluted 1∶200 for IF and 1∶2000 for WB. FITC- and Rhodamine- conjugated anti-mouse and anti-rabbit secondary antibodies (Jackson Immuno Research) were diluted 1∶200 for IF and 1∶2000 for WB, respectively. The biotin-labeled anti-sense hsrω-n RNA 280b riboprobe was generated from the pDRM30 plasmid [32] and used for FRISH. For gel mobility assays the sense hsrω-n RNA riboprobe was generated from the same plasmid. Single and double immunofluorescence on polytene chromosome spreads were conducted as previously described [11]. Larval tissues (salivary glands, Malpighian tubules and testis) were dissected from third-instar larvae grown at 22°C. Fully or partially squashed tissue preparations were used for FRISH and Immuno-FRISH assays as previously described [14] with some modifications (Text S1). Total proteins from salivary glands and Malpighian tubules were extracted as previously described [11]. The SDS-PAGE separated proteins were transferred onto nitrocellulose membrane (Whatman Schleicher & Schuell) for Western detection using SuperSignal West Femto substrate (Pierce). Chemiluminescent signals were acquired with the ChemiDoc XRS imager (BioRad). Native larval nuclear protein extracts from third instar w1118 larvae were prepared as previously described [11] and RNA-immunoprecipitations were conducted as published earlier [33] with small modifications (Text S1). Recombinant full length ISWI or ISWI-N or ISWI-C proteins [23], [24] were incubated with in vitro transcribed sense 280b tandem repeat unit of the hsrω-n ncRNA or a generic RNA of the same size (RNActr, Roche) as a control, in increasing ratios of 1∶1, 5∶1,10∶1 and 20∶1 nmoles. The hsrω-n ncRNA or the RNActr were incubated with the desired protein for 30 min at 25°C in RB2 buffer (20% Glycerol, 0.2 mM EDTA, 20 mM Tris-HCl pH 7.5, 1 mM MgCl2, 150 mM NaCl, 1 mM DTT and RNAsin). After incubation, the RNA/protein complexes were resolved on 1.4% agarose gel in 0.5× TBE at 4°C for 105 minutes at 70 volts. RNA molecules were visualized by ethidium bromide staining. ATP, ATP-γ-S and ADP (Roche) were added in the gel shift assay at a final concentration of 100 µM. Excess of cold hsrω-n repeat unit or a generic RNActr transcript was used as competitor for ISWI/hsrω binding detected by gel mobility shift using P33 radiolabeled hsrω280b sense repeat unit and recombinant ISWI. RNA/protein complexes were resolved as above. After gel drying, RNA/protein complexes were visualized using the BioRad Phosphoimager system. ATPase assay was conducted as previously described [23]. Extent of ATP hydrolysis was calculated with the following formula [P33/(P33+AMP-P−P33)]*100 (Figure 5G). The ATPase activity of 4 nmoles of full length ISWI was assayed for 1 hour; 4 nmoles of ISWI-N and ISWI-C were assayed for 30 minutes in the presence of 2 nmole of either supercoiled plasmid DNA, 280 bp hsrω-repeat unit encoding double stranded DNA, hsrω-n 280 bp tandem repeat ncRNA or a 300 bp generic RNA (RNActr; Roche) as a control.

10.1371/journal.ppat.1002014

Lung Adenocarcinoma Originates from Retrovirus Infection of Proliferating Type 2 Pneumocytes during Pulmonary Post-Natal Development or Tissue Repair

Jaagsiekte sheep retrovirus (JSRV) is a unique oncogenic virus with distinctive biological properties. JSRV is the only virus causing a naturally occurring lung cancer (ovine pulmonary adenocarcinoma, OPA) and possessing a major structural protein that functions as a dominant oncoprotein. Lung cancer is the major cause of death among cancer patients. OPA can be an extremely useful animal model in order to identify the cells originating lung adenocarcinoma and to study the early events of pulmonary carcinogenesis. In this study, we demonstrated that lung adenocarcinoma in sheep originates from infection and transformation of proliferating type 2 pneumocytes (termed here lung alveolar proliferating cells, LAPCs). We excluded that OPA originates from a bronchioalveolar stem cell, or from mature post-mitotic type 2 pneumocytes or from either proliferating or non-proliferating Clara cells. We show that young animals possess abundant LAPCs and are highly susceptible to JSRV infection and transformation. On the contrary, healthy adult sheep, which are normally resistant to experimental OPA induction, exhibit a relatively low number of LAPCs and are resistant to JSRV infection of the respiratory epithelium. Importantly, induction of lung injury increased dramatically the number of LAPCs in adult sheep and rendered these animals fully susceptible to JSRV infection and transformation. Furthermore, we show that JSRV preferentially infects actively dividing cell in vitro. Overall, our study provides unique insights into pulmonary biology and carcinogenesis and suggests that JSRV and its host have reached an evolutionary equilibrium in which productive infection (and transformation) can occur only in cells that are scarce for most of the lifespan of the sheep. Our data also indicate that, at least in this model, inflammation can predispose to retroviral infection and cancer.

The identification of cells that give origin to cancer is critical in order to design effective therapeutic strategies. To this end, the early stages of cancer are the most informative but they are seldom associated with clinical symptoms and therefore pass unnoticed in human patients. Studies on animal tumors are invaluable to this research area. In this study, we determined the cells originating an infectious lung cancer of sheep (ovine pulmonary adenocarcinoma, OPA) that is similar to some forms of human pulmonary adenocarcinoma. OPA is caused by a virus known as Jaagsiekte sheep retrovirus (JSRV). We show that OPA is caused by JSRV infection of proliferating type 2 pneumocytes (lung alveolar proliferating cells, LAPCs). We show that young animals possess abundant LAPCs and are highly susceptible to JSRV infection while healthy adult sheep exhibit a relatively low number of LAPCs and are resistant to OPA induction. However, adult sheep were susceptible to JSRV infection when the presence of LAPCs was stimulated by induction of a mild injury to the respiratory epithelium. Thus, our study identifies the cells originating lung adenocarcinoma in OPA and shows that inflammation to the respiratory epithelium can predispose to retrovirus infection and cancer.

Retroviruses have been instrumental in understanding the genetic basis and the fundamental molecular mechanisms leading to cancer [1]. Studies on the pathogenesis of retrovirus induced malignancies have also contributed to our understanding of the cells that give origin to cancer and the role played by stem and progenitor cells in these processes [2]. The “cancer stem cell” (CSC) hypothesis postulates that cancer is initiated and sustained by adult stem cells [3]–[4]. A growing body of experimental evidence is supporting the presence of CSCs in haematological malignancies and in some solid tumours. However, the presence and significance of CSCs is object of considerable debate particularly in slow turnover organs such as the lungs [5]–[7]. Identifying the cells that give origin to cancer is critical both to understand the basic carcinogenetic processes but also to devise appropriate therapeutic strategies. Most retroviruses induce transformation of hematopoietic cells but there are a few notable exceptions causing sarcomas, nephroblastomas, mammary carcinomas, nasal and lung adenocarcinomas in a variety of animal species [8]. Ovine pulmonary adenocarcinoma (OPA) is a naturally occurring (and experimentally inducible) lung cancer of sheep caused by a retrovirus known as Jaagsiekte sheep retrovirus (JSRV) [9]–[11]. OPA is a common disease of sheep in most geographical areas of the world. Interestingly, the disease shares several clinical and histological features with some forms of human lung adenocarcinomas. Therefore, OPA represents an excellent animal model with great potential to contribute significantly to our understanding of retroviral pathogenesis, lung tumorigenesis and pulmonary biology [9], [12]–[13]. JSRV is the only oncogenic virus that causes a naturally occurring lung adenocarcinoma. Interestingly, in contrast to the overwhelming majority of oncogenic retroviruses, JSRV is a replication-competent virus that possesses a structural protein (the viral envelope, Env) that acts as a dominant oncoprotein [14]–[16]. Expression of the JSRV Env is sufficient to induce cell transformation in vitro in a variety of cell lines [13]–[15], [17]–[22] and importantly in vivo in both experimental mice models and in lambs [23]–[24]. Thus, productive virus infection and cell transformation are mutually dependent in OPA and this creates an “evolutionary dilemma” as, at face value, abundant viral replication is entirely dependent on tumor development in the host. The JSRV Env is believed to induce cell transformation via the activation of a variety of signal transduction pathways including the PI-3K/Akt and Ras-MEK-MAPK [13], [20], [22],[25]–[27]. Experimentally, intratracheal inoculation of concentrated JSRV viral particles in young lambs induces OPA in the overwhelming majority of animals with a very short incubation period (varying from a few weeks to a few months) [28]–[29]. There is a clear age-dependent susceptibility to experimentally induced OPA in lambs while it is not possible (or extremely difficult) to reproduce the disease in adult sheep [29]. These data suggest that there is a different availability of the target cells of JSRV transformation in animals of a different age. The age-susceptibility to OPA induction does not appear to be related to expression of the receptor in target cells or to a differential immune response towards the virus. Indeed, the cellular receptor for the virus (Hyaluronidase-2, Hyal-2) is ubiquitously expressed [16], [29] and this virus can infect several cell types in vitro and in vivo [30]–[33]. In addition, JSRV naturally or experimentally-infected animals do not mount a significant immune response, likely as a result of tolerance induced by expression of JSRV-related endogenous retroviruses (enJSRVs) which are present in the genome of all domestic and wild sheep [34]–[37]. In OPA affected sheep, abundant expression of JSRV proteins are confined to the tumor cells although viral RNA and DNA can be detected by sensitive PCR assays in a variety of cells of the lymphoreticular system [30]–[31], [38]. In sheep naturally infected with JSRV and with no neoplastic lesions, JSRV can be detected only in lymphoid tissues [39]. OPA tumours, similar to some human lung adenocarcinomas, are formed by secretory cells of the distal pulmonary tract; predominantly alveolar type 2 pneumocytes and less commonly the non-ciliated bronchial cells of the terminal bronchioli (Clara cells; see note at the end of the text on the usage of this term) [40]–[42]. Interestingly, a putative bronchioalveolar stem cell (BASC) has been identified in mice lungs although its presence in other species, including humans, has not been established with certainty [43]. It has been proposed that BASCs have the capacity to originate both Clara cells and type 2 pneumocytes and to be the cell origin of lung adenocarcinoma in mice in response to oncogenic K-ras [43]. However the significance of BASCs in physiological and pathological processes and the origin of lung adenocarcinoma are under debate [44]–[45]. In order to identify the target cells of JSRV infection and transformation we performed a series of in vivo studies in experimentally infected lambs and adult sheep. Furthermore, we derived a JSRV-based vector in order to assess the ability of this virus to infect non-dividing cells in vitro. In this study we identified the cells target of JSRV infection and transformation and provide important insights into lung biology, pulmonary carcinogenesis and retroviral pathogenesis. All experimental procedures carried out in this study are included in Project Licence 60/3905 approved by the Home Office of the United Kingdom in accordance to the “animals (scientific procedures) act 1986”. Experiments carried out at the Istituto G. Caporale were also detailed in protocol number 3315 approved by the Italian Ministry of Health (Ministero della Salute) in accordance with Council Directive 86/609/EEC of the European Union. Viral stocks used in all these experiments were produced in rat 208F.JSRV21 cells as already described [46]. Briefly, 208F.JSRV21 derive from 208F cells [47] stably transfected with a plasmid expressing the JSRV21 infectious molecular clone [11]. 208F.JSRV21 cells were plated at 80% confluence and supernatants were collected after 24, 48 and 72 h. Virus was concentrated by ultracentrifugation [300×] as previously described [11] and resuspended in 1×TNE buffer (100mM NaCl, 10 mM Tris, 1 mM EDTA). The infectious titer for JSRV cannot be easily calculated in vitro, because of the lack of a convenient tissue culture system for this virus. In order to infect animals with the same amount of JSRV, pellets from various virus preparations were pooled into a single stock, divided into 1 ml aliquots and stored at −80°C until use. In all the experiments described below, each animal received the same amount of virus stock. In a related study, the same JSRV preparation used here, induced OPA in 4 of 4 experimentally infected lambs within 5 months after inoculation (Caporale and Palmarini, unpublished). Animal studies were performed at the Istituto G. Caporale (Teramo, Italy) and at the University of Glasgow. Prior to experimental infections all animals were anaesthetised with sodium pentobarbital anesthesia, and all efforts were made to minimize suffering. To facilitate the detection of infected cells, JSRV (1 ml) was inoculated directly into the accessory bronchus of the cranial lobe of the right lung by fiber-optic bronchoscopy. Sheep used in this study were females between 3 and 5 year old of either bergamasca cross-breed (study 1, 2 & 4) or blackface breed (study 3) unless otherwise indicated. Three independent studies were performed as follow. Formalin-fixed, paraffin-embedded OPA tumour samples from naturally occurring (n = 6) and experimentally induced (n = 2) cases were obtained from the Department of Veterinary Pathology, University of Zaragoza. All tumour samples were previously diagnosed as JSRV positive by immunohistochemistry as already described [23], [38], [48]. Four serial sections for each tumour were analysed by immunofluorescence as described below. Tissue sections were deparaffinised and hydrated using standard procedures. Antigen retrieval was performed using citrate buffer (pH6) and pressure cooker heating. To quench endogenous peroxidase, sections were incubated in 3% H2O2 diluted in methanol or PBS for 30 minutes. Sections were incubated overnight at 4°C with the following primary antibodies: polyclonal rabbit anti pro-SP-C (Seven Hills Bioregagents or Chemicon, dilution 1∶4000), monoclonal mouse anti Ki67 (DAKO, 1∶2000), mouse monoclonal anti JSRV Env (1∶200, kindly provided by Dusty Miller) [24], [49]. For CC10 detection we used either a polyclonal rabbit (Proteintech) or mouse (Dundeecell products) antisera generated against full length recombinant bovine CC10. Mouse CC10 was detected using goat anti-mouse CC10 clone T18 (Santa Cruz; 1∶200). Immunofluorescence detection was performed using the following labelled secondary antibodies: goat anti-mouse Alexa488, donkey anti-rabbit Alexa-555, donkey anti-rabbit Alexa 488. SP-C was detected using horseradish peroxidase (HRP)-conjugated donkey anti-rabbit secondary antibody (1∶6000) by tyramide signal amplification (TSA; Perkin-Elmer Life Science Products) while Ki67 was detected using donkey anti-mouse Alexa488 or Alexa-555. Slides were mounted with medium containing DAPI (Vectashield; Vector Laboratories). Immunohistochemistry was performed with Dako supervision system (DAKO) and slides were counterstained with haematoxylin. Confocal images were analysed and merged using Image-pro analyser 7 software (MediaCybernetics). Histological images were captured using cell∧D software (Olympus). Proliferation analysis was performed by counting SP-C/Ki67 double positive cells in the entire 10 lung sections for each animal using a Leica TCS SP2 confocal microscope. Numbers of double positive cells were normalized to the sectioned area using Image-pro analyser 7 Software. Bronchiolar cell proliferation was determined by counting the number of CC10+/Ki67+ cells in 100 terminal bronchioli for each animal. The JSRV-based vector employed in this study was derived from the JSRV21 infectious molecular clone pCMV2JS21 [11] and was termed pCJS-EfGFP-mC. Most of the JSRV gag and pol have been deleted and replaced by a cassette containing the promoter of the human elongation factor 1 α (EF1α) driving the enhanced green fluorescent protein (eGFP). The EF1α-eGFP cassette was derived from pDRIVE5-GFP-3 (InvivoGen). In addition, pCJS-EfGFP-mC also contains the woodchuck hepatitis post-transcriptional regulatory element (WPRE; before the env splice acceptor) [50]–[51] derived from pCCLcPPTPGKEGFPLTRH1shSOD1 (Addgene Inc.). In pCJS-EfGFP-mC, the JSRV env has also been deleted and replaced with the cDNA expressing the mCherry fluorescent protein, followed by two copies of the Mason-Pfizer constitutive transport element (CTE) [52]–[53]. The packaging plasmid pCMVGPP-MX-4CTE expresses the JSRV Gag, Pro and Pol genes and derives from plasmid pGPP-MX by the addition of 3 additional CTE copies. pGPP-MX has been already described [23]. pCMV-SX2.JS-env expresses the JSRV Env under the control of the CMV immediate early promoter and was derived from the pSX2.Jenv (a gift by Dusty Miller) [33], [54]. pCDNA3-HA-Sam68 is an expression plasmid for the RNA binding protein Sam68 and was a gift from David Shalloway [55]. Plasmids pCSGW-GFP (HIV-based vector), p8.2 and pMD.G have been described previously [56]. 293T cells and sheep choroid plexus (SCP) cells were grown in Dulbecco's modified Eagle's medium and Iscove's modified Dulbecco's medium respectively supplemented with 10% fetal bovine serum at 37°C, with 5% CO2 and 95% humidity. Particles of a JSRV-based viral vector (JS-EeGFP-mCherry) were produced by co-transfecting 293T cells with pCJS-EfGFP-mC, pCMVGPP-MX-4CTE, pCMV-SX2.JS-env and pCDNA3-HA-Sam68 plasmids essentially as described previously [23]. Viral particles were collected from supernatants of transfected cells, 24 and 48 h post-transfection, filtered through 0.45 µm filters (Millipore) and concentrated [200×] by ultracentrifugation as described previously [23]. A lentiviral vector (HIV-GFP) was used as control and prepared exactly as above by co-transfecting 293T cells with pCSGW-GFP, p8.2 and pMD.G. Target cells synchronization was established by culturing SCP cells in the presence of 0.2% fetal bovine serum (FBS) for 72 h. Synchronized SCP cells were then seeded at 5×104 cells/well in 6 well plates and treated for 25 h with 5 µg aphidicolin (Sigma). Target cells were infected with serial dilutions of the JSRV or HIV-based vector in presence of polybrene [57], [58]. Transduction controls included infection with heat-treated vector preparations (65°C/30′). 12 h post-infection, cells were washed three times with phosphate-buffered saline and incubated for further 48 h in the presence or absence of aphidicolin. Viral titers were expressed as fluorescence forming foci/ml and were determined by counting foci of GFP positive cells 48 h post-infection. Cellular DNA content was determined by staining cells with 7-Aminoactinomycin D (7AAD, Invitrogen) and measuring fluorescence in a Beckman Coulter flow cytometer. SCP cells were harvested by trypsinization and incubated for 1 h with 25 µg/ml 7AAD, 0.03% saponin (Sigma) and 1% BSA (Sigma). Cells were then transferred in 500 µl of 1×PBS and the proportion of cells in G0/G1, S and G2/M phases was estimated using expo32 software (Beckman Coulter) and counting 20000 events. Ultrastructural, histological and immunophenotyping studies have shown that OPA tumours, similarly to some forms of human adenocarcinomas, are formed by type 2 pneumocytes and to a lesser extent by Clara cells [40], [42], [59]–[62]. No data are available in the literature on whether JSRV is expressed in both these cell types in the OPA tumours. Here, we analysed by immunofluorescence and confocal microscopy serial tumor sections collected from six sheep with late stages of naturally occurring OPA and two lambs with experimentally induced disease, in order to characterize both the phenotype of the cells forming the neoplasm and viral expression. Type 2 pneumocytes and Clara cells can be easily identified by the expression of surfactant protein-C (SP-C) and the Clara cell 10 protein (CC10) respectively [63]–[64]. As expected, our confocal microscopy analysis revealed that all the neoplastic foci were composed mainly by SP-C+ cells (Fig. 1). In all cases the SP-C+ cells co-expressed the JSRV Env that was localized mainly at the apical surface of the cell (Fig. 1A–C). Despite multiple optical serial section (z stacks images) were analysed for each section, we found that the majority of tumor lesions were formed by cells that did not express CC10 (Fig. 1D). Areas with CC10+ cells were detected in 2 of the 6 natural OPA tumours analyzed. However, in both of these cases CC10+ positive cells did not show clear expression of the JSRV Env (Fig. 1E–F). Experimentally, OPA can be easily induced in lambs but not in adults [28]–[29], [65]. The incubation period of experimentally induced OPA is directly related to the age of the infected animals [29]. These data can be explained by hypothesizing a differential abundance of the cell targets for viral infection in lambs compared to adult sheep. Alternatively, the target cells for JSRV infection may be present both in lambs and in adults but only in the former, infection is able to progress to neoplastic transformation. In order to begin to address this issue we experimentally infected four newborn lambs and four adult sheep with JSRV and analysed virus-infected cells 10 days post-infection. Virus was inoculated directly in the accessory bronchus via bronchoscopy in order to facilitate subsequent detection (Fig. 2A). Animals were euthanized 10 days post-infection and lung samples collected from either 8 (in lambs) or 16 (in adult sheep) regions of the cranial lobe of the lungs to maximise the chances of detecting a small number of virus infected cells and in order to compensate the differences in size between the lambs and adult lungs. We detected JSRV infected cells by immunohistochemistry using monoclonal antibodies against the viral Env [24], [49]. We were not able to detect any JSRV-infected cells in all the sections derived from the adult sheep used in this experiment (Fig. 2B–C). In contrast, all sections analyzed from each lamb showed JSRV-infected cells (Fig. 2B, D–E). On average, in each lamb we detected 32 clusters of JSRV infected cells ranging in size from 1 to 36 cells (mean 4.9±6.5) with some of them clearly displaying a neoplastic phenotype. Overall these data strongly suggest that the age related susceptibility to OPA is due to the ability of JSRV to infect cells that are much more abundant in the lungs of lambs compared to adult sheep. We then characterized the phenotype of viral infected cells in the lungs of experimentally infected lambs. We analyzed by immunofluorescence and confocal microscopy lung sections incubated with both antibodies towards SP-C or CC10 and the JSRV Env. In all cases, JSRV Env+ cells were also SP-C+ (Fig. 2F, G). We were not able to detect any JSRV Env+ cell that was also CC10+. Some early neoplastic lesions were observed in the respiratory bronchioli but in these cases they were always CC10 negative (Fig. 2H–I). Overall, the data obtained in experimentally infected lambs at the early stages of viral infection are in accordance with the observations made in naturally occurring OPA cases and indicate that cells of the type 2 pneumocytes lineage are infected and transformed by JSRV. So far, our results showed that lambs are more susceptible to experimentally induced OPA due to the ability of the virus to infect type 2 pneumocytes in lambs but not in adults. Obviously, mature type 2 pneumocytes are present abundantly both in lambs and adults. Therefore, we reasoned that JSRV was able to infect a sub-population of SP-C+ that was abundantly present in lambs but not in adult sheep. The normal developed lung is a relatively quiescent organ, with low levels of cell division in the bronchioalveolar epithelium [66]. For a variety of mammals, lungs are not yet mature at birth but continue to develop during a period (“alveolar” stage) where the number of alveoli increases dramatically [67]–[68]. Thus, we hypothesised that JSRV infected lung alveolar proliferating cells instead of post-mitotic type 2 pneumocytes. In order to test this hypothesis, we first analysed by immunofluorescence the mitotic status of type 2 pneumocytes and Clara cells in lambs and adults sheep lungs using antibodies towards the proliferation marker Ki67 [69] in conjunction with either antisera towards SP-C or CC10 (Fig. 3). We found that proliferating type 2 pneumocytes (SP-C+/Ki67+), addressed here as lung alveolar proliferating cells (LAPCs), were up to 50 times more abundant in newborn lambs compared to adult sheep (p<0.001) (Fig. 3A–B). Also proliferating Clara cell (CC10+/Ki67+) in the terminal bronchioli were more abundant in lambs compared to adult sheep. We detected 94.5±39 CC10+/Ki67+ per 100 terminal bronchioli in lambs while there were only 5.5±2.1 CC10+/Ki67+ per 100 terminal bronchioli in adult sheep (p = 0.004) (Fig. 3C–D). A subset of SP-C+/CC10+ putative pulmonary stem cells (known as bronchioalveolar stem cells or BASCs) was identified at the bronchioalveolar junction in mice [43]. We analysed the localization of the proliferating Clara cells in the terminal bronchioli of lambs and sheep and found that they were not localised in a specific area of the terminal bronchioli but randomly distributed. In addition, we could not detect SP-C+/CC-10+ double-positive cells by confocal microscopy in either lambs or adult sheep, while we were able to identify cells with this phenotype in mice (Fig. S1). So far our data suggested that the presence of LAPCs in lambs is the main factor determining the susceptibility of young animals to JSRV infection as opposed to the resistance observed by adult sheep. Indeed, in the adult lungs, the proliferation rate of the respiratory epithelium is very low [68]. However, the lung has a significant reparative capability and after an injury the LAPC proliferate and play an important role in the tissue regenerative process. We therefore reasoned that we would be able to render adult sheep susceptible to experimental JSRV infection by previous induction of a mild lung injury that would stimulate LAPCs. 3MI is an organ-selective pneumotoxicant that affects specifically type I pneumocytes and bronchiolar epithelial (Clara) cells and it is especially effective in ruminants [69], [70]. Here, to assess the ability of 3MI to induce lung injury and repair we exposed two sheep to this pneumotoxicant and we then assessed lung injury after 48 hours. Histological examination showed diffuse pulmonary edema with scattered hemorrhagic foci (Fig. 4A–B). Next, we assessed the proliferation status of type 2 pneumocytes and Clara cells by verifying co-expression of SP-C or CC10 with the proliferating marker Ki67 by immunofluorescence as described above (Fig. 4C–F). The number of SP-C+/Ki67+ cells was 90 fold higher in sheep after lung injury as opposed to normal control sheep (p<0.001) (Fig. 4H). The examination of the terminal bronchioli in sheep after 3MI administration revealed that almost 100% of terminal bronchioli contained CC10+/Ki67+ (Fig. 4G). The total number of CC10+/Ki67+ cells was more than 100 fold higher in adult sheep after lung injury compared to healthy controls (p = 0.009) (Fig. 4H). Also in adult sheep after lung injury we were not able to identify any SP-C+/CC10+ double-positive cells (data not shown). Overall, the data presented above indicate that the number of LAPCs, that we identified as target cells of JSRV infection, increase dramatically after mild lung injury. In order to determine whether lung injury may render adult sheep susceptible to JSRV infection, we treated five sheep with 3MI and after 48 h we infected them with JSRV (Group I). Five additional sheep were infected with JSRV without pre-treatment with 3MI (Group II). 10 days after infection animals were euthanized (Fig. 5A). As expected, post-mortem examination revealed no signs of lesions attributed to lung injury. In each animal, the presence of JSRV infection was assessed in 15 sections collected from the cranial lobe by immunohistochemistry. JSRV Env expression was only detected in lung cells of animals that were infected after treatment with 3MI (Fig. 5B, D–F). On average, 10 clusters of JSRV Env+ cells (ranging from 1 to 80 cells) were detected in each animal while no JSRV infected cell was detected in those animals that were infected without 3MI pre-treatment (Fig. 5C). By immunofluorescence and confocal microscopy we found that all JSRV infected cells were SP-C positive (Fig. 6A–C). None of the JSRV Env+ cells were CC10+ (Fig. 6D–F), despite the high number of proliferating Clara cells induced by 3MI and the presence of numerous infected cells localized in the terminal bronchioli. Our data have shown that JSRV infects LAPCs but not the overwhelming majority of type 2 pneumocytes which divide very slowly. These data could be explained mechanistically by the fact that the majority of retroviruses, with the exception of lentiviruses [71], infect more efficiently cells that are in mitosis [72]–[73]. The proliferation rate of type 2 pneumocytes is very low in adults under normal conditions. On the other hand the higher proliferative rate of LAPCs during post-natal development or tissue repair in the adult would facilitate JSRV infection. Experiments with JSRV in vitro are hindered by the lack of a convenient tissue culture system for the propagation of this virus [32]. Therefore, we constructed a convenient JSRV-derived viral vector (JS-EeGFP-mCherry) in order to easily quantify JSRV infection in proliferating and non-proliferating cells. JS-eGFP-mCherry was derived by transiently transfecting 293T cells with (i) a packaging plasmid (pGPP-MX-4CTE) devoid of the JSRV packaging signal (Ψ) and expressing the viral Gag, Pro and Pol; (ii) a plasmid providing the JSRV Env in trans (pC-ML-JSenv, also devoided of Ψ), and (iii) the packaged JSRV vector (pCJS-EFGFP-MC) that upon infection and integration expresses eGFP under the control of an internal promoter (Fig. 7A). JS-eGFP-mCherry viral particles were then used to infect synchronized SCP cells in the presence or absence of a drug that, at the concentration used in this study, arrests cells in the G1 phase (aphidicolin) (Fig. 7B). Consistently, JS-eGFP-mCherry was able to transduce actively dividing SCP cells approximately 200 times more efficiently (p = 0.002) than the same cells where mitosis was arrested with aphidicolin while only minor differences between treated and untreated cells were observed with the lentivirus vector HIV-GFP (Fig. 7C). In this study we have investigated the pathogenesis of a unique virus-induced lung adenocarcinoma and obtained data that have a broad significance in pulmonary biology, carcinogenesis and retroviral pathogenesis. Most adenocarcinomas in humans display cells expressing type 2 pneumocytes or Clara cell markers but it is not completely clear whether the neoplasm arises from a stem cell that is able to differentiate into both cell types, or from a committed progenitor or from the fully differentiated cell compartments [74]. In this study, we identified the target cells of JSRV infection and transformation in vivo as proliferating cells of the type 2 pneumocytes lineage (SP-C+/Ki67+, LAPC). In addition, we showed that the age-related susceptibility to experimental OPA induction is directly related to the abundance of LAPCs. Importantly, induction of mild injury to the respiratory epithelium increased dramatically the number of LAPCs in adult sheep and rendered these animals susceptible to JSRV infection and transformation. We have not found evidence that CC10+/Ki67+ cells are infected and transformed by JSRV. Furthermore, we found that the CC10+ cells that are found in a proportion of late stages OPA tumours are not expressing JSRV proteins and may therefore not be true tumour cells, at least in the cases we examined. Our data provide important consideration for pulmonary biology and carcinogenesis. We infer from our study that at least in sheep, type 2 pneumocytes and Clara cells have two distinct populations of proliferating progenitor cells committed to the alveolar and the bronchiolar lineages. From this study, we cannot determine whether the LAPCs are progenitor committed solely to type 2 or type 1 pneumocytes. We showed that lung adenocarcinoma can originate from an alveolar proliferating cell of the alveolar lineage, rather than from a bronchioalveolar stem cell postulated to originate both type 2 pneumocytes and Clara cells. Studies in mice have identified a population of putative stem cells that are both SP-C+ and CC10+ (bronchioalveolar stem cells, BASCs) located at the bronchioalveolar duct junction [43]. Based on in vitro analysis, BASCs were hypothesised to give rise to Clara cells, alveolar type 2 cells and be the cell originating lung adenocarcinoma [43]. On the other hand, studies using genetic lineage-labelling experiments in mice, supported a model where bronchioli and alveoli are maintained and repaired distinctively by Clara cells and LAPCs respectively [44], [75]. The presence of BASCs in humans has not been confirmed and in general the biological relevance of BASCs is object of debate [44]–[45]. In our study, by confocal microscopy, we have not been able to detect SP-C+/CC10+ in sheep while we were able to detect cells with this phenotype in mice (Fig. S1). We cannot rule out the presence of a rare bronchioalveolar stem cell (SP-C+/CC10+) able to differentiate in both type 2 pneumocytes and Clara cell progenitors in sheep. We also cannot rule out the presence in sheep of phenotypically uncharacterized pulmonary stem cells. However, if these cells exist in the sheep, they are very rare and unlike LAPCs they do not appear to play a major role in OPA. Interestingly, from the anatomical and histological point of view the human lungs are more comparable to the sheep lungs as opposed to the mice lungs [76]–[77]. We showed with experiments in vitro that JSRV, similarly to other retroviruses, infects preferentially cells in active mitosis. These experiments provide a mechanistic explanation to the observation that JSRV infects readily LAPCs but not mature type 2 pneumocytes. As mentioned before, JSRV is a unique oncogenic virus as it possesses the viral Env (a structural protein) that behaves as a functional dominant oncoprotein both in vitro and in vivo. In general, viral oncoproteins are non structural proteins whose expression is not linked to productive infection. It would be detrimental from an evolutionary point of view of the virus, to have productive viral infection and carcinogenesis as strictly mutually dependent events (viral replication would in this case lead to the death of the infected host). Onset of lung adenocarcinoma in JSRV-infected animals could therefore be viewed as either “accidental” (similarly to other retrovirus-induced tumors) or “essential” in order to allow virus spread among susceptible hosts. Although these two alternative hypotheses are not necessarily mutually exclusive, the data obtained in this study and accumulated over the years on JSRV/OPA, strongly suggest that tumor induction plays an important part in the evolutionary strategies used by the virus to persist in the sheep population. In previous studies we have shown that development of OPA in the field occurs only in a minority of the JSRV-infected sheep [39]. On the other hand, animals with OPA produce lung secretions containing abundant amounts of infectious JSRV particles that pour freely from the nostrils of the affected sheep [41], [78]–[79]. The data from this paper strongly suggests that clinical OPA develops in natural conditions as a result of viral infection only when LAPCs are available to the virus: in young lambs during post-natal development or in the presence of an injury to the bronchioalveolar epithelium. Importantly, as mentioned in the introduction, JSRV proteins are detected readily only in the tumour cells of OPA affected animals (and in the LAPCs as shown in this study) [38] although low levels of virus infection and protein expression are detectable in cells of the lymphoreticular system of animals with or without clinical OPA. We and others have shown that the JSRV LTRs are the main determinants regulating the tight cell-specific expression pattern displayed by this virus. The JSRV LTRs contains lung-specific enhancer binding motifs that are preferentially active in cell lines derived from transformed type 2 pneumocytes [80]–[83]. In addition, in transgenic mice, reporter gene expression driven by the JSRV LTR has been detected specifically in type 2 pneumocytes [84]. Thus, JSRV-host equilibrium has been reached by a combination of factors. JSRV has evolved a structural protein that is a powerful oncoprotein but only when expressed at high levels in the LAPCs, which are relatively rare cells in the adult healthy sheep. Therefore, JSRV has a limited window of opportunity to infect the target cells of the host that allow high level of viral expression (and that can be consequently transformed). At the same time, onset of lung adenocarcinoma in a minority of the infected animals allows an amplification of the cells that can produce infectious virus and therefore it is a likely evolutionary mechanism that helps JSRV to persist in the population. It is important to note that in natural conditions, sheep with OPA present consistently a variety of other parasitic, bacterial or viral infections [9]. Classically, these infections were considered as “secondary” to JSRV infection. We suggest instead that in the adult, the induction of an injury to the respiratory epithelium by various pathogens substantially increases the number of LAPCs and renders adult sheep susceptible to JSRV-induced transformation, similarly to what we have shown experimentally in this study with the pneumotoxicant 3MI. Thus, inflammation induced by different pathogens is the “primary” event for OPA induction. It is feasible that in animals already infected with JSRV the virus present in lymphoreticular cells is able to spread to injured tissues where it can infect and transform alveolar progenitor cells actively involved in repairing the epithelium. In conclusion, this work provided unique insights into pulmonary physiology, lung cancer, and retrovirus pathogenesis and is another telling example where viruses have helped us to understand fundamental aspects of host biology.

10.1371/journal.ppat.1006542

Human cytomegalovirus IE1 downregulates Hes1 in neural progenitor cells as a potential E3 ubiquitin ligase

Congenital human cytomegalovirus (HCMV) infection is the leading cause of neurological disabilities in children worldwide, but the mechanisms underlying these disorders are far from well-defined. HCMV infection has been shown to dysregulate the Notch signaling pathway in human neural progenitor cells (NPCs). As an important downstream effector of Notch signaling, the transcriptional regulator Hairy and Enhancer of Split 1 (Hes1) is essential for governing NPC fate and fetal brain development. In the present study, we report that HCMV infection downregulates Hes1 protein levels in infected NPCs. The HCMV 72-kDa immediate-early 1 protein (IE1) is involved in Hes1 degradation by assembling a ubiquitination complex and promoting Hes1 ubiquitination as a potential E3 ubiquitin ligase, followed by proteasomal degradation of Hes1. Sp100A, an important component of PML nuclear bodies, is identified to be another target of IE1-mediated ubiquitination. A C-terminal acidic region in IE1, spanning amino acids 451 to 475, is required for IE1/Hes1 physical interaction and IE1-mediated Hes1 ubiquitination, but is dispensable for IE1/Sp100A interaction and ubiquitination. Our study suggests a novel mechanism linking downregulation of Hes1 protein to neurodevelopmental disorders caused by HCMV infection. Our findings also complement the current knowledge of herpesviruses by identifying IE1 as the first potential HCMV-encoded E3 ubiquitin ligase.

Congenital human cytomegalovirus (HCMV) infection is the leading cause of neurological disabilities in children, but the underlying pathogenesis of this infection remains unclear. Hes1, an important effector of Notch signaling, governs the fate of neural progenitor cells (NPCs) and fetal brain development. Here we demonstrate that: (1) HCMV infection results in loss of Hes1 protein in NPCs; (2) the HCMV immediate-early 1 protein (IE1) mediates Hes1 protein downregulation through direct interaction, which requires amino acids 451–475; (3) IE1 assembles a Hes1 ubiquitination complex and mediates Hes1 ubiquitination; and (4) IE1 also assembles an Sp100A ubiquitination complex and mediates Sp100A ubiquitination, but does not require amino acids 451–475. These results suggest that HCMV IE1 is a potential E3 ubiquitin ligase. Downregulation of Hes1 by HCMV infection and IE1 implies a novel mechanism linking Hes1 depletion to virus-induced neuropathogenesis.

As a leading cause of birth defects, congenital human cytomegalovirus (HCMV) infection causes irreversible maldevelopment of the central nervous system (CNS) in newborns and children [1–4]. To understand how HCMV interferes with neurodevelopment, neural progenitor cells (NPCs) have been utilized as a clinically relevant model for investigation of the underlying mechanisms [5–10]. Proper self-renewal and differentiation of NPCs are fundamental to normal fetal brain development. Notch signaling is one of the best-characterized pathways governing NPC maintenance, proliferation and differentiation [11–13]. This regulatory role is achieved, at least partially, through essential downstream effectors such as the Hairy and Enhancer of Split (Hes) proteins, which belong to the repressor-type basic helix-loop-helix family [14, 15]. Hes1 is one of seven members in the Hes family, which play a crucial role in maintaining the undifferentiated and proliferative status of NPCs [16–18]. The auto-negative feedback regulation at the transcription level, the instability of the mRNA, and the rapid ubiquitination-dependent proteasomal degradation of the protein together result in the well-regulated Hes1 oscillation, which in turn fine-tunes the timing of NPC proliferation and differentiation and further controls the shape, size and integrity of brain structures [13, 19–21]. Studies in mice have shown that Hes1-deficient murine NPC neurospheres fail to expand, and Hes1 knockout accelerates neurogenesis from radial glial cells representing NPCs in mice [21–24]. These evidences imply that the dysregulation of Hes1 expression leads to abnormal NPC differentiation and proliferation, potentially contributing to fetal brain developmental disorders. We have previously reported that HCMV dysregulates Notch signaling by targeting Notch1, including its intracellular active domain (NICD), and the ligand Jag1 [5]. Moreover, as an important downstream effector in Notch signaling, the regulation of Hes1 expression in NPCs is disrupted by HCMV infection [25]. During HCMV infection of permissive cells, immediate early (IE) genes are the first to be expressed from the viral genome, and IE proteins can be detected as early as 2h post infection (hpi). IE proteins trigger viral early gene expression and, subsequently, viral genome replication and late gene expression. In addition, IE proteins also interact with multiple host factors to tune the cellular environment for initiation of viral replication. Thus, the synthesis of IE gene products is necessary for a full viral replication cycle [26, 27]. The most abundant and arguably most important HCMV IE gene products, termed IE1 and IE2, are encoded in the major IE transcription unit. The 72-kDa IE1 is a nuclear phosphoprotein and has been subject to extensive study [28]. The IE1 protein promotes the accumulation of IE2 gene products and synergizes with IE2 to activate viral early promoters [29–33], in part by antagonizing histone deacetylation, to facilitate virus replication [34, 35]. IE1 also affects host gene expression by activating or repressing transcription. For example, IE1 up-regulates the transcription of interleukin 6 (IL-6) and the genes activated by signal transducer and activator of transcription 1 (STAT1) [36]. Moreover, IE1 inhibits transactivation of p53-dependent downstream genes, disrupts transcription of STAT3-activated genes, and downregulates certain essential NPC markers such as the glial fibrillary acidic protein (GFAP) [37–40]. The recent study on the crystal structure of the IE1 ortholog from Macacine herpesvirus 3 (Rhesus cytomegalovirus) revealed striking similarities between IE1 and tripartite motif (TRIM) family proteins [41]. Many TRIM proteins possess E3 ubiquitin ligase activities [42], and E3 ubiquitin ligase activities have been described in several α- and γ-herpesvirus proteins, including ICP0 of herpes simplex virus type 1 (HSV-1), ORF61p of varicella-zoster virus (VZV), and replication and transcription activator (RTA), K3 and K5 of Kaposi’s sarcoma-associated herpesvirus (KSHV) [43–48]. However, to our knowledge, no E3 ubiquitin ligase has been identified in HCMV or any other β-herpesvirus so far. In the present study, we demonstrated that HCMV infection downregulates Hes1 protein levels in infected human NPCs. Importantly, IE1 leads to Hes1 depletion by mediating Hes1 ubiquitination and proteasomal degradation by acting as a potential E3 ubiquitin ligase. IE1 physically interacts with Hes1 via amino acids (AA) 451–475, which are also essential for IE1-mediated Hes1 ubiquitination. In addition, Sp100A, an important component of PML nuclear bodies (PML-NBs), is identified as an additional ubiquitination substrate of IE1 ubiquitination. This study reveals a novel potential mechanism of HCMV induced neuropathogenesis and represents the first description of the potential E3 ubiquitin ligase activity of HCMV IE1. Human NPCs are fully permissive for HCMV replication [6–9], and our previous work has demonstrated that HCMV infection dysregulates NICD1 and Jag1 in NPCs, which are two essential components upstream of Hes1 in the Notch signaling pathway [5]. To investigate the effect of HCMV infection on Hes1, the HCMV Towne strain was used to infect NPCs at a multiplicity of infection (MOI) of 3. Protein levels of Hes1 were then examined by immunoblotting (IB) from 4hpi to 96hpi. In comparison to mock-infected cells, protein levels of Hes1 in infected NPCs were clearly decreased by 12hpi and were nearly undetectable from 24hpi to 96hpi (Fig 1A). UV-inactivated HCMV had no effect on Hes1 protein levels (Fig 1B), suggesting that the downregulation of Hes1 may depend on viral transcription and/or de novo synthesized viral gene products rather than input viral components. We have previously shown that Jag1 and NICD1 protein levels did not decrease before 16-24hpi (S1 Fig) [5], which was after Hes1 downregulation started (12hpi), indicating the Hes1 downregulation is not a consequence of Jag1 and NICD1 dysregulation in the IE phase of infection. To confirm that Hes1 downregulation requires newly synthesized viral proteins, HCMV-infected NPCs were treated with the protein synthesis inhibitor cycloheximide (CHX). Due to the low levels of endogenous Hes1 in NPCs, 0.25μg of Hes1 expressing construct (pCDH-Hes1) was nucleofected into NPCs to enable Hes1 protein detection in the presence of CHX. As expected, the representative de novo synthesized viral protein IE1 was present in untreated HCMV-infected NPCs, but undetectable in CHX-treated cells. Similar to the data shown in Fig 1A, the Hes1 protein levels substantially decreased upon HCMV infection in the absence of CHX, but remained at similar levels between HCMV- and mock-infected NPCs upon CHX treatment (Fig 1C). Thus, de novo synthesized viral proteins, but not the input virus components, are necessary for Hes1 downregulation during HCMV infection. Taken together, these data demonstrate that HCMV infection of NPCs results in depletion of the Hes1 protein via a mechanism that requires de novo viral protein synthesis. The fact that Hes1 downregulation occurs as early as 12hpi and the requirement for de novo viral protein synthesis suggest the potential involvement of HCMV proteins present during the IE phase of infection. Therefore, candidate viral gene products, including the major IE proteins (IE1 and IE2) and the most abundant input virus component pp65, were introduced into NPCs via nucleofection with plasmids pCDH-IE1, pCDH-IE2 and pCDH-pp65, respectively. Expression of pp65 displayed no significant effect on Hes1 protein level, consistent with the notion that the input components of HCMV virions are not involved in Hes1 downregulation induced by virus infection. In contrast, expression of IE1 or IE2 lead to markedly lower Hes1 protein levels (Fig 2A). Considering that IE1 showed a stronger down-regulating effect on Hes1 than IE2, we focused our subsequent work on IE1. The effect of IE1 on protein level of Hes1 in the context of HCMV infection was further tested by comparing a recombinant IE1-deficient virus (TNdlIE1) to the parental wild-type (TNwt) and a “revertant” virus (TNrvIE1), respectively [38, 49, 50]. To overcome the replication defect of TNdlIE1 exhibited at low MOIs, the infection was performed at an MOI of 10, and the protein levels of Hes1 and IE1/2 were analyzed at 12hpi. As expected, IE1 was only present following the infection by TNwt and TNrvIE1 but not TNdlIE1, while IE2 levels were similar in all three infections. In TNwt- and TNrvIE1-infected NPCs, protein levels of Hes1 were downregulated to 32.5±0.04% and 32.0±0.08% of that in mock-infection (Fig 2B, right panel), respectively. In contrast, no obvious difference in protein levels of Hes1 between TNdlIE1- and mock-infected cells was observed (Fig 2B). Although IE2 expressed by itself seemed to also downregulate the protein level of Hes1 as shown in Fig 2A, it displayed virtually no capacity to alter Hes1 protein level in TNdlIE1 infected NPCs, perhaps due to its low abundance during infection and the relatively weak downregulating effect on Hes1 protein. Consistent with our findings using the TNdlIE1 virus, shRNA-mediated IE1 knock-down in HCMV-infected NPCs restored the diminished Hes1 protein amount to normal level (Fig 2C). To exclude potential interference of endogenous Hes1 at the transcription level, a different cellular environment was applied to examine the downregulating effect of IE1 on exogenous Hes1. 293T cells were co-transfected with the constructs expressing IE1 (pEYFP-IE1) and Hes1 (pCDH-Hes1), and the protein levels were analyzed at 48 h post transfection (hpt). As shown in Fig 2D, the expressed IE1 reduced the exogenous Hes1 protein amount in a dose-dependent manner. Taken together, IE1, either produced during HCMV infection or expressed in the absence of virus, downregulates the protein level of Hes1, and IE1 knock-down restores Hes1 protein level during virus infection. Notably, the protein level of Hes1 was minimally affected by IE2 in the absence of IE1 during the IE phase of infection. These results support that IE1 is sufficient and necessary for downregulation of the Hes1 protein during HCMV infection. Having shown that IE1 expression was associated with downregulation of Hes1, we next investigated the underlying mechanism. Based on high resolution images obtained by two-photon microscopy, the Hes1 protein was distributed evenly across mock-infected NPC nuclei. However, upon HCMV infection, Hes1 was observed to relocate and concentrate at sub-nuclear areas where IE1 was present at 4hpi (Fig 3A). The co-localization between IE1 and Hes1 suggests that these two proteins could be physically associated. To investigate a potential interaction between IE1 and Hes1 (IE1/Hes1 interaction), we carried out an immunoprecipitation assay (IP) in NPCs infected with TNwt, TNdlIE1 or TNrvIE1 for 12h. An IE1/Hes1 interaction along with Hes1 downregulation was only observed in TNwt- and TNrvIE1-, but not in TNdlIE1-infected cells (Fig 3B). To exclude the possibility that other viral proteins are mediating the IE1/Hes1 interaction, IP analysis was performed in the context of transduced NPCs expressing IE1 (Fig 3C) and transfected 293T cells co-expressing the two proteins (Fig 3D). Co-precipitated IE1 and Hes1 were detected by IB following IPs using Hes1- or IE1-specific antibodies, respectively. A normal IgG control was used for IP to rule out the possibility of non-specific binding (Fig 3C and 3D). Both IE1 and Hes1 are transcription factors: IE1 activates many different promoters and binds nucleosomes [51]; Hes1 binds its own promoter for auto-suppression [52]. To exclude the possibility that the IE1/Hes1 interaction is mediated via viral or cellular nucleic acids, cell lysates of transfected 293T cells were treated with DNase and RNase prior to IP analysis. As shown in Fig 3E, removal of DNA/RNA by DNase/RNase treatment (S2 Fig) did not alter the IE1/Hes1 interaction. This result indicates that IE1/Hes1 interaction is independent of nucleic acids and likely resulted from a direct protein-protein interaction. The HCMV (Towne) IE1 protein comprises a total of 491 amino acids and can be roughly divided into an N-terminal (AA1-85, shared with IE2), a central/core (AA86-372) and a C-terminal (AA373-491) domain. Many interactions of IE1 and other proteins have been mapped to a region proximal to the C-terminus, which contains four short low complexity motifs including three acidic “domains” (AD1, AD2 and AD3) enriched in aspartic and glutamic acid, and a stretch of amino acids enriched in serine and proline (SP) [49]. The AD1 and SP motifs have been implicated in IE1/STAT2 binding [49, 53]. The terminal 16 amino acids following these four motifs are responsible for nucleosome targeting, and are thus named the chromatin tethering domain (CTD) [51] (Fig 4A). To examine whether the IE1/Hes1 interaction depends on residues in the C-terminal domain of IE1, we tested a number of IE1 deletion mutants (Fig 4A). To this end, 293T cells were co-transfected with constructs expressing Hes1 and wild-type (IE1) or mutant IE1. The IE1/Hes1 interaction was observed at ratios of pCDH-Hes1: pEYFP-IE1 (Hes1:IE1) plasmid DNA of 1:1 and 1:5 (S3A Fig). A clearer result was obtained at the ratio of 1:1, and therefore a 1:1 ratio was used for further interaction analysis. The steady-state levels of all tested IE1 mutants were comparable to that of wild-type IE1 in cell lysates (Fig 4B and 4C). Subsequently, the interaction of IE1 mutants with Hes1 was investigated by IP analysis. To rule out any nonspecific binding, the plasmid pEYFP and normal IgG were used as the vector control and the nonspecific IP antibody control, respectively (S3B Fig). Due to the lack of commercial antibodies from different species for Hes1 and IE1, both the IP and the subsequent IB were performed with mouse monoclonal antibodies. This resulted in the detection of IgG heavy chains (Fig 4B, 4C and 4E, indicated by arrows). IE1 mutants Δ373–420, Δ476–491 and Δ86–404 were found to be co-precipitated with Hes1, but not Δ421–475 (Fig 4B). To exclude the potential interference of heavy chain bands, which were similar to the size of Δ421–475, and narrow down the interaction site of IE1, AA421-475 was further subdivided along the boundary of AD2 and AD3 into AA421-445 and AA451-475. One of the resulting IE1 mutants (Δ421–445) was still capable of binding Hes1, while the other one (Δ451–475) failed to interact, as determined by IP analysis (Fig 4C). Both mutants (Δ421–445 and Δ451–475) were clearly distinguishable from the heavy chain. To confirm that AA451-475 of IE1 is required for binding to Hes1 and to test whether the interaction occurs in the absence of other viral or host factors, His-tagged versions of Hes1, IE1, and IE1Δ451–475 were expressed in E. coli, purified using metal affinity chromatography, and used in an in vitro pull-down assay. As shown in Fig 4D, His-IE1 but not His-Δ451–475 pulled down His-Hes1. Furthermore, the IE1/Hes1 interaction, along with Hes1 downregulation, was not observed in NPCs infected with AA451-475 deleted virus (TN-IE1(Δ451–475)) (Fig 4E). Taken together, our results indicate that IE1-mediated Hes1 downregulation is linked to direct interaction between the two proteins, which requires the AA451-475 region (comprising the AD3 motif) within the C-terminal domain of IE1. Recently published findings have revealed that the IE1 protein of Rhesus cytomegalovirus, a close homolog of HCMV IE1, contains a secondary protein structure similar to the coiled-coiled domain of TRIM-25 [41]. Many TRIM proteins have E3 ubiquitin ligase activity [42], and several IE proteins of herpesviruses other than HCMV also function as E3 ubiquitin ligases, some of which share functional similarities with HCMV IE1 [43, 44, 47, 54, 55]. These data lead us to investigate whether HCMV IE1 mediates Hes1 downregulation through prompting its ubiquitination and proteasomal degradation. When levels of ubiquitinated Hes1 were examined in NPCs, Hes1 ubiquitination was observed in both mock- and HCMV-infected cells, indicating that the ubiquitin-proteasome pathway is involved in Hes1 protein regulation, which is concordant with previous observations in other cell types [16, 52]. Treatment with the proteasome inhibitor MG132 led to a substantial increase of Hes1 steady-state levels in cell lysates, as well as of ubiquitinated Hes1 in HCMV-infected NPCs, compared to DMSO controls and mock-infected NPCs (Fig 5A). Notably, MG132 treatment did not completely restore the Hes1 protein levels in the infected samples to levels found in mock control, indicating proteasomal degradation is likely not the only mechanism involved in Hes1 regulation. To examine the effect of IE1 on prompting Hes1 ubiquitination, NPCs were transduced with a lentivirus expressing IE1. After confirmation of IE1 expression, cells were treated with MG132 or DMSO, and then subjected to Hes1-directed IP followed by IB against ubiquitin. The levels of ubiquitinated Hes1 were increased in MG132-treated IE1-expressing NPCs compared to the MG132-treated IE1-negative cells (Fig 5B). Furthermore, the ubiquitination level of Hes1 protein in TN-IE1(Δ451–475)-infected NPCs was significantly lower than that in TNwt- and TNrvIE1-infected NPCs, while similar to that in mock-infected NPCs (Fig 5C). Consistently, IE1 but not Δ451–475 promoted exogenous Hes1 ubiquitination in transfected 293T cells (Fig 5D). These data suggested that IE1 downregulates Hes1 protein level by enhancing its ubiquitination and proteasomal degradation, and that these activities require AA451-475 of IE1. Consistent with ubiquitination as a mechanism for Hes1 downregulation, expression of IE1 reduced the half-life of exogenous Hes1 from 19.1±2.9 min to 9.3±0.2 min in transfected 293T cells, while similar expression levels of IE1 mutant Δ451–475 did not alter Hes1 half-life (20.7±2.7 min) (S4 Fig). To ubiquitinate a substrate, it is necessary that an E3 ubiquitin ligase (E3) assembles the ubiquitination complex composed of E2 conjugating enzyme (E2), E3 and the substrate [56]. Therefore, we next investigated whether IE1 interacts with an E2 and Hes1 (substrate) and mediates the complex formation. Ubc5a serves as an E2 in HSV-1 ICP0 ubiquitinating Sp100A and KSHV RTA ubiquitinating IRF-7 and MyD88 [48, 55, 57], and an interaction between HCMV IE1 and Ubc5a was also observed in HCMV infected NPCs (Fig 6A). Thus, Ubc5a was used in the following assays as a potential E2 for IE1-mediated Hes1 ubiquitination. To exclude the possibility of influence from other viral or cellular components, the prokaryotically expressed proteins of His-Hes1, His-IE1 and His-Δ451–475 were purified. The purity of the purified recombinant proteins was examined by SDS-PAGE and Coomassie Brilliant Blue staining (S5A Fig). The purified proteins were incubated in vitro together with commercial Hes1- or Ubc5a-specific antibodies, and subjected to IP assay. The results showed that both Ubc5a (E2) and His-Hes1 (substrate) interact with IE1, and the His-Hes1/Ubc5a interaction was observed only in the presence of IE1, indicating that IE1 mediated the His-Hes1/Ubc5a interaction (Fig 6B). These data suggested that IE1 interacts with both Ubc5a and His-Hes1 to assemble a ubiquitination complex. To confirm whether IE1 can mediate Hes1 ubiquitination as an E3 ligase, an in vitro ubiquitination reaction was also performed. The in vitro ubiquitination reaction included His-Hes1, His-IE1 or His-IE1Δ451–475 (4μg), ubiquitin activating enzyme (E1), Ubc5a, ubiquitin and the reaction buffer containing ATP. Following a 2 hours’ incubation at the indicated temperature, the reaction products were subjected to Hes1-directed IP followed by IB for ubiquitin. The addition of His-IE1 led to ubiquitinated His-Hes1, whereas ubiquitinated Hes1 was undetectable when the IE1 mutant His-Δ451–475 rather than the wild-type protein was present in the reaction. No Hes1 ubiquitination was observed in the absence of E1, Ubc5a or ubiquitin, which ruled out non-specific reactions (Fig 6C). To further confirm this result, we doubled the amount of the purified proteins (8μg) used in the in vitro ubiquitination assay and directly assayed the results by IB with anti-Hes1 antibody but without a prior IP. Ubiquitinated Hes1 was clearly detectable only in the presence of His-IE1 together with all other necessary components, but His-Δ451–475 failed to ubiquitinate His-Hes1 (Fig 6D). These data indicate that IE1 mediates Hes1 ubiquitination in vitro as a potential E3 ubiquitin ligase, and AA451-475 is required for IE1-mediated Hes1 ubiquitination. Taken together, HCMV IE1 assembles Ubc5a and His-Hes1 to form a ubiquitination complex and mediates Hes1 ubiquitination as a potential E3 ubiquitin ligase. These processes require AA451-475 of the IE1 protein. Sp100A is one of the major components of PML-NBs. Previously published data show that Sp100A is downregulated upon HCMV infection via proteasomal degradation, and IE1, which interacts with Sp100A, is potentially involved in this HCMV induced Sp100A downregulation [58, 59]. Sp100A is also a ubiquitination substrate of HSV-1 ICP0, an E3 ubiquitin ligase [55]. Moreover, both HCMV IE1 and HSV-1 ICP0 are IE proteins of herpesviruses and share certain similarities in their functions. Based on this evidence we next considered whether HCMV IE1 downregulates Sp100A via ubiquitination. Both the endogenous and exogenous Sp100A were downregulated by IE1 in a dose-dependent manner in 293T cells (Fig 7A and 7B), which is supported by previous observations [58, 59]. Since downregulation of exogenous Sp100A was more efficient than endogenous Sp100A, the ubiquitination level of exogenous Sp100A was subsequently examined by in vivo ubiquitination assay in 293T cells transfected with pCMV-Myc-Sp100A and pEYFP-IE1 or vector control. Cells were treated with MG132 or DMSO, and then subjected to Myc-directed IP followed by IB against ubiquitin. Following treatment with MG132, the expressed IE1 increased the ubiquitination level of Myc-Sp100A compared to the vector controls, suggesting that IE1 prompts Sp100A ubiquitination in vivo (Fig 7C). Δ451–475 and wild-type IE1 showed a similar capacity to ubiquitinate Sp100A (Fig 7D), which was different from that observed with the ubiquitination of Hes1. To exclude interference from other cellular components in the in vivo system, ubiquitination complex formation and ubiquitination capacity were examined in vitro using purified His-Sp100A and His-IE1 (S5B Fig). His-IE1 bound to both Ubc5a and His-Sp100A (substrate), and mediated the interaction between them (Fig 7E), indicating IE1 plays an essential role in formation of the Sp100A ubiquitination complex. Moreover, ubiquitinated Sp100A was detected in the presence of His-IE1, as well as His-Δ451–475, together with all other necessary components (Fig 7F). Importantly, deletion of AA451-475 abolished the ubiquitination capacity of IE1 on Hes1 (Figs 5C, 5D and 6C, 6D), but had no influence on IE1-induced Sp100A ubiquitination (Fig 7D and 7F). The differential effect of IE1 AA451-475 deletion on Hes1 and Sp100A prompted us to analyze whether AA451-475 is also required for downregulation of Sp100A by IE1. Consistent with the ubiquitination capacity, IE1 mutant Δ451–475 failed to reduce the protein levels of exogenous Hes1 in 293T cells (Fig 8A). However, Δ451–475 still downregulated exogenous Sp100A (Fig 8B), and accordingly, retained a capacity for binding to Sp100A (Fig 8C). These data indicate AA451-475 is not required for either Sp100A downregulation or the IE1/Sp100A interaction. Taken together, consistent with regulation of Hes1 by IE1, IE1 initiates Sp100A ubiquitination and assembles an Sp100A ubiquitination complex. Moreover, while AA451-475 of IE1 are necessary for the IE1/Hes1 interaction, this region is dispensable for IE1-mediated Sp100A downregulation and IE1/Sp100A interaction. As a critical downstream effector in Notch signaling, the Hes1 protein regulates the fate of NPCs by repressing the transcription of pro-neural genes and thus plays an essential role in maintaining the stem cell status of NPCs, governing their differentiation towards neurons or glia, and controlling fetal brain development [16, 17]. Congenital HCMV infection is one of the most common causes of neurological disabilities in children. However, the relationship between Hes1 regulation and virus infection, in particular whether and how HCMV regulates Hes1, remained unknown. In the present study, we report for the first time that HCMV infection downregulates Hes1 protein at the level in human NPCs through IE1 via a newly identified function, which may be a key mechanism that contributes to fetal brain maldevelopment caused by congenital HCMV infection. Hes1 protein synthesis is highly regulated at the transcription level and by rapid degradation through the ubiquitin-proteasome pathway [52]. To exclude the possibility of potential interference of endogenous Hes1 transcription in NPCs, exogenous Hes1 was constitutively expressed in 293T cells using construct pCDH-Hes1. Under the control of the HCMV major IE promoter, the transcription level of Hes1 was elevated in the presence of HCMV IE1, but the protein level of Hes1 was still decreased, indicating the strong downregulating effect of IE1 on Hes1 at the post-translational level. In addition, IE1 physically interacts with Hes1, thus leading us to ask whether IE1 downregulates Hes1 through the ubiquitin-proteasome pathway. The ubiquitin-proteasome pathway has been implicated in numerous physiological processes and disease outcomes, including regulation of genes controlling fetal brain development, and has been associated with neurodegenerative diseases [60, 61]. Additionally, the proteasome pathway is also one of the cellular machineries commonly hijacked by viruses to degrade cellular or viral proteins to favor virus replication [62–64]. For example, HSV-1 utilizes proteasome-mediated ubiquitination-independent proteolysis for successful target cell entry [65], and KSHV relies on ubiquitin-dependent proteasome degradation for viral entry and intracellular trafficking in endothelial cells [66]. The ubiquitination-proteasome pathway is also utilized by HCMV to promote viral gene expression [67]. Ubiquitination relies on three essential steps, each depending on particular enzyme(s): (1) activation of ubiquitin by ubiquitin-activating enzyme (E1); (2) transfer of the activated ubiquitin to ubiquitin-conjugating enzyme (E2) and formation of an E2-Ub thioester; and (3) formation of an isopeptide bond between ubiquitin and the substrate protein by ubiquitin protein ligase (E3) [68]. To accomplish step 3, E3 ubiquitin ligase assembles a ubiquitination complex by interacting with both the E2-Ub thioester and the specific substrate, and then mediates the transfer of ubiquitin from the E2-Ub thioester to the substrate [69]. This specific substrate recognition and interaction of E3 ubiquitin ligase also confers the specificity of ubiquitination [70]. Many viral proteins have been identified to possess E3 ubiquitin ligase activity, and several members of the herpesvirus also encode their own E3 ubiquitin ligases, such as ICP0 of HSV-1, ORF61P of VZV, K3/K5 and RTA of KSHV, as well as mK3 and ORF75c of murine gamma herpesvirus 68 (MHVγ68) [54, 71]. Although no HCMV-, or even β-herpesvirus-, encoded E3 ubiquitin ligase has been identified, the following data imply that HCMV IE1 might be an E3 ubiquitin ligase: (1) all identified E3 ubiquitin ligases of herpesviruses are IE proteins [54]; (2) HCMV IE1 involves in HCMV infection induced Sp100A downregulation through proteasomal pathway, and interacts with Sp100A, which is also a target of HSV-1 encoded E3 ubiquitin ligase ICP0 [55, 58, 59, 72]; (3) the central core domain of the Rhesus cytomegalovirus IE1, the HCMV IE1 homolog, shares certain secondary structure similarity with the TRIM proteins, many of which function as E3 ubiquitin ligases [41]; and (4) most viral E3 ubiquitin ligases belong to the RING E3 family, whose activity depends on the RING domain—two zinc atoms complexed with the cysteine/histidine residues in a ‘cross-brace’ manner [73]. Interestingly, IE1 contains a zinc finger motif (HX2HXFX3LX2CX4C, AA267-284) with two cysteines and two histidines (underlined) capable of binding Zn2+ [37]. Our results and these known features of HCMV IE1 structure lead us to speculate that this viral protein may function as an E3 ubiquitin ligase and downregulate Hes1 and Sp100A via ubiquitination. IE1 does not contain any known canonical E3 ubiquitin ligase motifs, including HECT domain, RING finger and U-box [68]. However, proteins without these structures may also act as E3 ubiquitin ligases [73]. For example, RTA of KSHV functions as an E3 ubiquitin ligase regulating IRF7, K-RBP, MyD88, LANA-1 and KbZIP, but lacks any of the known E3 motifs [48, 57, 71]. The different families of E3 ubiquitin ligases regulate versatile E3 substrates via various E3 functional mechanisms, and their substrate recognition specificity makes E3 ubiquitin ligases share very few unified sequences or structure similarities. Thus, the most convincing and efficient method to identify a specific protein as an E3 ubiquitin ligase is through biochemical methods, including the in vitro ubiquitination assay [68]. Our studies have provided several findings to support the role of IE1 as a potential E3 ubiquitin ligase: (1) the purified IE1 ubiquitinates substrates Hes1 and Sp100A in vitro in the presence of E1, Ubc5a, ubiquitin and ATP, and the absence of any of these components resulted in a failure of IE1 to function in the ubiquitination assay; (2) IE1 interacts with Ubc5a and the substrates, which is one of the common features of E3 ubiquitin ligases [68, 73]; and (3) the ubiquitination reaction requires the formation of the E2/E3/substrate complex[56], and our data demonstrate that IE1 assembles Ubc5a and substrate (Hes1 and Sp100A) to form a tripartite complex. Ubc5a was used as the E2 in the in vitro ubiquitination reaction because it is recruited as an E2 enzyme when HSV-1 ICP0 or KSHV RTA ubiquitinate Sp100A or IRF-7 and MyD88, respectively [48, 55, 57]. Taken together, these results demonstrate that HCMV IE1 assembles the Ubc5a and Hes1 or Sp100A to form a ubiquitination complex, and mediates the ubiquitination of the substrates, thus functioning as a potential E3 ubiquitin ligase. AD3 (AA 451–475) in the C-terminal domain of IE1 is essential for the interaction of IE1/Hes1 but not IE1/Sp100A. IE1 Δ451–475 failed to interact with and downregulate the Hes1 protein, but retained its effect on Sp100A. The E3 ubiquitin ligase requires the assembly of the ubiquitination complex by binding E2 and substrate, thus AA451-475 of IE1 is a Hes1 binding site rather than the active site of E3 ubiquitin ligase activity. According to the structure of IE1, AA280 and AA284 cysteines within the putative zinc finger motif may be the potential catalytic sites of E3 activity. Alternatively, the IE1 zinc finger possibly mediates the interaction of IE1 and Ubc5a, similar to cellular E3 RNF125 (also known as TRAC-1) [74]. However, the E3 catalytic center of IE1 may be located elsewhere and thus requires further investigation. Multiple host cellular factors restrict viral replication. For instance, PML-NBs, consisting of PML, Daxx, Sp100A and other proteins, suppress HCMV replication as intrinsic anti-viral structures [75, 76]. Depletion of Sp100A, an important PML-NB component, enhances IE1/2 expression and HCMV replication [58]. Correspondingly, viruses target the intrinsic cellular defense proteins through viral encoded E3 ubiquitin ligases to counteract host restriction and favor viral replication [54, 55, 62]. HCMV has been reported to downregulate Sp100A via the proteasomal pathway, and IE1 has been implicated in this process [58, 59], which supports our finding that IE1 prompts Sp100A ubiquitination and proteasomal degradation as an E3 ligase. Downregulation of Sp100A protein occurred in both Towne and AD169 infected fibroblasts, but at different timing. Towne infection started to decrease Sp100A protein as early as 12 hpi, but obvious Sp100A protein downregulation was observed till the late stage (72 hpi) of AD169 infection [58, 59]. IE1 proteins of Towne and AD169 differ at only AA68 and AA394, Arginine and Alanine for HCMV-IE1, and Histidine and Valine for AD169-IE1, respectively. Both variations employ amino acids with similar characters, and locate outside of the potential E3 catalytic regions. Therefore, IE1 of both strains are presumed to similarly boost Sp100A degradation as E3 to facilitate viral replication, which is partially supported by the above mentioned studies [58, 59]. On the contrary, host cells also upregulate Sp100A to activate the intrinsic anti-viral response upon viral infection, and different viruses possibly trigger various levels of cellular response, as well as Sp100A upregulation. Taken together, the variation of Sp100A downregulation timing observed upon Towne and AD169 infection is resulted by the different cellular and viral counteraction profiles induced by different viruses. But this hypothesis requires further confirmation. Our preliminary data also showed that the expression levels of IE1/2 were decreased in Hes1 overexpressing human embryonic lung fibroblast cells (HELs) (S6 Fig). The downregulation of Hes1 through HCMV IE1 is a probable mechanism to counteract the suppression of Hes1. However, the mechanism by which Hes1 regulates HCMV replication remains unknown and requires further investigation. In summary, (i) HCMV infection downregulates Hes1 protein through IE1; (ii) IE1 directly interacts with Hes1, prompting Hes1 ubiquitination and proteasomal degradation, and thus reduces the half-life and the steady-state level of the Hes1 protein; and (iii) Sp100A is identified to be another target of HCMV IE1-mediated ubiquitination. In addition, IE1 assembles E2 (Ubc5a) and substrate (Hes1 and Sp100A) to form a ubiquitination complex, and further mediates their ubiquitination. The regulation of Hes1 by IE1 requires AA451-475, which mediate binding to Hes1, but this region is not required for regulation of Sp100A. Our study not only suggests an important mechanism for fetal brain development disorders induced by congenital HCMV infection, but also reveals a novel unanticipated function of HCMV IE1 as a potential E3 ubiquitin ligase. Human neural progenitor cells (NPCs) and human embryonic lung fibroblast cells (HELs) were isolated from postmortem fetal embryo tissues of brain and lung, respectively. The original source of the anonymized tissues was Zhongnan Hospital of Wuhan University (China). The cell isolation procedures and research plans were approved by the Institutional Review Board (IRB) (WIVH10201202) according to the Guidelines for Biomedical Research Involving Human Subjects at Wuhan Institute of Virology, Chinese Academy of Sciences. The need for written or oral consents was waived by IRB. NPCs were isolated and maintained in the laboratory as described previously [10]. Briefly, hippocampus, bilateral ventricular and subventricular zone tissues were isolated from the brain, washed with Hank’s Balanced Salt Solution (HBSS) supplemented with antibiotics (1,000U/ml of penicillin, 1mg/ml streptomycin, Life Technologies), diced with scalpel blades and eye scissors, digested with DMEM-F12 media (Life Technologies) containing 2.5U/ml papain (Worthington), 40U/ml dispase II (StemCell Technologies) and 1U/ml DNase I (Qiagen) at 37°C for 20 min, mixed with 90% Percoll in phosphate-buffered saline (PBS), and centrifugated at 1,500g for 15 min. The NPCs were obtained from the top Percoll layer. The growth medium (GM) used for NPCs was DMEM-F12 supplemented with 2mM GlutaMax, penicillin-streptomycin (100U/ml and 100mg/ml) 50mg/ml gentamycin, 1.5mg/ml amphotericin B (all from Life Technologies), 10% BIT9500 (StemCell Technologies), 20ng/ml epidermal growth factor and 20ng/ml basic fibroblast growth factor (Prospec). For maintenance, half of the NPCs culture medium was replaced with fresh GM every two days, and the replaced medium was liberated from cell debris and used as conditioned medium (CM) [10]. For HELs isolation, lung tissues were prepared under sterile conditions, rinsed with PBS containing 1,000U/ml of penicillin and 1mg/ml streptomycin, diced into 1mm3 fragments, digested with 0.25% trypsin (Life Technologies) at 37°C for total 15 min with gentle shaking for three times, then mixed with equal volume of MEM media containing 10% fetal bovine serum (FBS, Life Technologies), and centrifuged at 500g for 10 min. The HELs were obtained from the pellet. HELF-IE1, the cell line stably expressing HCMV IE1, was obtained by lentivirus (pCDH-puro-IE1) transduction and puromycin (8μg/ml, Sigma) selection on HELFs, which are immortalized human embryonic lung fibroblast cells (kindly provided by Dr. Jason J. Chen at Columbia University). Expression of IE1 was confirmed by immunoblotting and the resulting HELF-IE1 cell line was maintained in the presence of puromycin (4μg/ml). Human embryonic kidney (HEK) 293T cells were purchased from ATCC (CRL-11268). The cells were cultured in appropriate medium (MEM for HELs and HELF-IE1, and DMEM for 293T) supplemented with 10% FBS, penicillin-streptomycin (100U/ml and 100mg/ml) at 37°C in a humidified atmosphere containing 5% CO2. For exogenous protein expression in NPCs and 293T cells, pCDH-CMV-MCS-EF1-copGFP (referred to as pCDH, System Biosciences), pCMV-Myc and pEYFP-N1 (referred to as pEYFP, Clontech Laboratories) were used. The coding sequences of human Hes1, Sp100A, HCMV IE1 (UL123), IE2 (UL122) and pp65 (UL83) were cloned into the vectors to generate pCDH-Hes1, pCMV-Myc-Sp100A, pCDH-IE1, pCDH-IE2, and pCDH-pp65, respectively. pEYFP constructs expressing wild-type IE1 and IE1 mutants lacking specific segments of amino acids comprised pEYFP-IE1, pEYFP-IE1(Δ86–404), pEYFP-IE1(Δ373–420), pEYFP-IE1(Δ421–445), pEYFP-IE1(Δ421–475), pEYFP-IE1(Δ451–475) and pEYFP-IE1(Δ476–491). Hes1, Sp100A, wild-type IE1 and IE1(Δ451–475) were also cloned into pET28a with a hexa-histidine- (His-) tag fused to their N-termini for prokaryotic expression and purification of the target proteins, including pET-Hes1 (His-Hes1), pET-Sp100A (His-Sp100A), pET-IE1 (His-IE1) and pET-IE1(Δ451–475) (His-Δ451–475). For IE1 knock-down, lentiviruses expressing a short hairpin RNA (shRNA) targeting IE1 (sh-IE1) and a non-specific scrambled shRNA control (Scram) were applied. The sequences of shRNAs are as follows: sh-IE1, GCT GTG CTG CTA TGT CTT AGA CTC GAG TCT AAG ACA TAG CAG CAC AGC TTT TTG; Scram, CCT AAG GTT AAG TCG CCC TCG CTC GAG CGA GGG CGA CTT AAC CTT AGG TTT TTG. The viruses applied for infection comprised wild-type HCMV Towne strain (referred to as HCMV, ATCC VR-977), and its bacterial artificial chromosome (BAC)-derived recombinant viruses including TNwt (wild-type), TNdlIE1 (IE1 deleted), TNrvIE1 (IE1 revertant) and TN-IE1(Δ451–475) (IE1-AA451-475 deleted). TNwt, TNdlIE1 and TNrvIE1 were described previously [49, 50]. TN-IE1(Δ451–475) was generated by deleting the sequence encoding AA451-475 of IE1 from TNwt BAC via homologous recombination in E.coli DY380, followed by reconstitution in HELs as described previously [77, 78]. Viruses were propagated in HELs (HCMV, TNwt and TNrvIE1) or HELF-IE1 cells (TNdlIE1 and TN-IE1(Δ451–475)), titrated by standard plaque forming assay on monolayers of HELs, and stored as aliquots at -80°C for later use [79]. For infection, NPCs were infected with the indicated viruses at the specified multiplicity of infection (MOI), and the inoculum was replaced with culture medium consisting of half fresh GM and half CM at 3hpi [10]. Infected cells were collected at the indicated time points for further analysis. Exogenous gene expression in 293T (1.5×106 cells in 100-mm dishes) was accomplished using the specified amounts of the indicated plasmid DNA for transfection by Ca3(PO4)2 precipitation following a standard protocol as described previously [80]. For NPCs, nucleofection with plasmid or transduction using lentivirus were performed as described previously [5]. Nucleofection was performed using the Amaxa Mouse NSC Nucleofector Kit (Lonza) following the manufacturer’s instructions. Briefly, NPCs (5×106) were mixed with 100μl Mesenchymal Stem Cell (MSC) Nucleofector Solution (82μl Nucleofector Solution with 18μl Supplement 1) containing specified amount of indicated plasmid DNA, transferred to a certified cuvette, and transfected using Nucleofector Program A-033. After nucleofection, NPCs were gently transferred to 100-mm dishes coated with poly-D-lysine, cultured in GM/CM mixture (1:1) for further culture and experiments. Lentiviruses were prepared by transfecting 293T cells (1.5×106 cells in 100-mm dishes) using 15μg pCDH empty vector (serving as the vector control), pCDH-IE1, pLKO.1-shRNA-IE1 or pLKO.1-shRNA-scram together with plasmids pML-Δ8.9 (12μg) and pVSV-G (8μg) by Ca3(PO4)2 precipitation following a standard protocol as described previously [80, 81]. Following a medium change at 12hpt, the supernatant containing the lentiviruses were harvested at 72hpt, titrated on 293T cells by quantifying GFP-positive cells, and stored as aliquots at -80°C for further application. NPCs were transduced with the corresponding lentiviruses at an MOI of 1, and the expression of the transgene(s) was determined by immunoblotting. Cells were collected at the indicated time points with cell scrapers, rinsed with ice-cold PBS, pelleted by centrifugation, snap-frozen in liquid nitrogen and stored at -80°C until use. Cell lysates were prepared by lysing cell pellets in Cell Lysis Buffer (Beyotime). Protein concentration of cell lysates was measured using the BCA Protein Assay Kit (Beyotime) according to the manufacturer’s protocol as described previously [80]. Equal amounts of protein were separated by SDS-PAGE and transferred to PVDF membranes (Millipore). After sequential incubation with the primary antibodies against the indicated target proteins and corresponding horseradish peroxidase (HRP)-conjugated secondary antibodies, the membranes were incubated with SuperSignal West Femto Maximum Sensitivity Substrate or SuperSignal West Pico Chemiluminescent Substrate (Life Technologies). The chemiluminescent signals were detected using a FluorChem HD2 System (Alpha Innotech) and analyzed densitometrically using ImageJ (National Institutes of Health). For analyzing the cell lysates by IB, β-actin served as a loading control. Primary mouse monoclonal antibodies for the following proteins were used: HCMV IE1 (IgG1, Abcam, Cat# ab30924), HCMV IE2 (IgG1, Santa Cruse, Cat#SC69835), IE1/2 (IgG1, Virusys, Cat# p1215), pp65 (IgG1, Virusys, Cat# p1205), Hes1 (IgG1, Abcam, Cat# ab119776) and β-actin (IgG1, Sungene Biotech, Cat# KM9001). HRP-conjugated goat anti-mouse IgG (IgG, Abbkine, Cat# A21010) was used as secondary antibody. To determine whether de novo viral protein synthesis is required for Hes1 downregulation, NPCs nucleofected with 0.25μg pCDH-Hes1 were pre-treated with 0.1mg/ml cycloheximide (CHX, Sigma-Aldrich) for 1 hour prior to HCMV or mock infection, and CHX treatment was maintained throughout the infection and post infection culture. Cells were collected at 12hpi, and protein levels of IE1 and Hes1 were determined by IB. The half-life of exogenous Hes1 protein in 293T cells were measured and calculated as previously described [82]. Briefly, 293T cells were transfected with 5μg pEYFP, pEYFP-IE1, or pEYFP-Δ451–475 together with 1μg pCDH-Hes1. CHX (0.1mg/ml) treatment started at 24hpt, and cells were collected at the indicated time points, Hes1 protein levels were determined by IB. NPCs grown on coverslips coated with poly-D-lysine were infected with HCMV at an MOI of 0.5, and mock infection was performed as a control. Coverslips were collected at the indicated time points post infection and processed for IFA as described previously [83]. Mouse monoclonal antibodies for HCMV IE1 (IgG2a, clone p63-27, made in Prof. Britt’s lab) and Hes1 (IgG1, Abcam, Cat# ab119776) were used as primary antibodies, and FITC-anti-mouse-IgG2a (Invitrogen, Cat# 11-4210-82) and TRITC-anti-mouse-IgG1 (Southern Biotechnology, Cat# 1070–03) were applied as secondary antibodies. Nuclei were counterstained with Hoechst 33342 (Life Technologies). Images were obtained using a two-photon microscopy with the UltraVIEW VoX 3D Live Cell Imaging System (Perkin Elmer). E. coli BL21 containing pET-Hes1, pET-Sp100A, pET-IE1 or pET-IE1(Δ451–475) respectively was grown in Luria-Bertani medium containing kanamycin (50μg/ml, Amresco) at 37°C with vigorous shaking (220rpm). Isopropyl-β-D-1-thiogalactopyranoside (IPTG, 1.0mM, Sigma-Aldrich) was applied to induce target protein expression when the OD600 reached 0.6. Cells were collected after induction at the optimized temperature for the specific times: His-Hes1 at 37°C for 8h, His-IE1 and His-IE1(Δ451–475) at 25°C for 8h, and His-Sp100A at 16°C for overnight. These His-tagged proteins were affinity-purified using the cOmplete His-Tag Purification Resin (Roche) following the manufacturer’s instruction. Briefly, the centrifugation pelleted cells were resuspended in Buffer A (50mM NaH2PO4, 300mM NaCl, pH 8.0), lysed by ultrasonication, centrifuged at 4°C to remove the cell debris, and loaded to Buffer A equilibrated resin columns (Bio-Rad). Following a 3 hours’ rotation at 4°C, the column was rinsed with Buffer A containing 20mM imidazole. Then, the protein was eluted from the resin using Buffer A containing the following concentration of imidazole: 400mM for His-Hes1, 600mM for His-Sp100A, 60mM for His-IE1 and 120mM for His-Δ451–475. For in vitro ubiquitination analysis without IP (described below), the eluted proteins were concentrated in Amicon Ultra-15 Centrifugal Filter Unit with Ultracel-10 membrane (Millipore) following the manufacturer’s protocol. The purity of the purified or concentrated proteins was examined by SDS-PAGE followed by Coomassie Brilliant Blue staining (Sigma-Aldrich). For the protein complex immunoprecipitation assay (IP), the centrifugation pelleted cells were re-suspended in Cell Lysis Buffer (Beyotime) and spun at 12,000g for 5min to remove debris. For the in vitro pull-down assay, 1μg of each purified protein or GST-Ubc5a protein (BostonBiochem) was mixed and incubated in a total volume of 300μl Buffer A. Cell lysate containing 2μg protein or the whole protein mixture was incubated with 1μg of the indicated antibody in a total volume of 300μl Buffer A. After thoroughly mixing by rotation overnight at 4°C, 60μl of protein A+G agarose beads (Beyotime) were added to the reaction and incubated with rotation for additional 3 hours at 4°C. The IP complex was pelleted by centrifugation at 2,500g for 5min, washed by Cell Lysis Buffer (Beyotime), re-suspended in protein loading buffer, subjected to SDS-PAGE, and immunoblotted (IB) for the indicated targets. For DNase and RNase treatment, 10,000 units of DNase I (Qiagen) and 10μg RNase (TianGEN) were applied to pre-treat the cell lysate for 1 hour prior to IP [84]. The applied antibodies included mouse monoclonal antibodies against Hes1 (IgG1, Abcam, Cat# ab119776), IE1 (IgG1, Abcam, Cat# ab30924), β-actin (IgG1, Sungene Biotech, Cat# KM9001), polyclonal rabbit antibodies against Sp100A (IgG, Abcam, Cat# ab167605). Secondary antibodies included HRP-conjugated goat anti-mouse IgG (IgG, Abbkine, Cat# A21010) and goat anti-rabbit IgG (IgG, Abbkine, Cat# A21020). Normal mouse IgG (Boster, Cat# BA1046) or normal rabbit IgG (cell signaling, Cat# 2729) were used as non-specific antibody controls. For in vivo ubiquitination analysis, 12.5μM MG132 (Sigma-Aldrich) was applied to the cells. HCMV-infected-NPCs were treated with MG132 starting at 3hpi, and harvested at 12hpi. For transduced NPCs or transfected 293T cells, MG132 was administered when the expressions of the target proteins were confirmed, and cells were treated for 12h. DMSO was applied as the solvent control. Cell lysates were subjected to Hes1- or Myc-directed IP followed by IB using a polyclonal rabbit antibody for ubiquitin. For the in vitro ubiquitination assay, 4μg of each indicated purified protein was mixed with 2μg ubiquitin, 0.1μg E1 enzyme and 0.5μg Ubc5a (all from BostonBiochem) in a total volume of 1ml reaction buffer containing 50mM Tris-HCl (pH 7.4), 5mM MgCl2, 2mM DTT, 4mM ATP (Fermentas) and 20μM MG132 (Sigma-Aldrich) [85]. The mixture was incubated at 37°C or 4°C for 2 hours, and then quenched by protease inhibitor cocktail (Roche) and 1 mM N-ethylmaleimide (Sigma) on ice [86]. The cell lysate or reaction mixture was then subjected to the Hes1-, Sp100A-directed IP, and followed by IB against ubiquitin. In addition, an in vitro ubiquitination assay was also performed without IP, with using double amount of each indicated purified protein, and the reaction mixture was directly subjected to IB with anti-Hes1. The ubiquitination reaction included 8μg of purified and concentrated proteins of His-IE1 or His-Δ451–475, and His-Hes1, 4μg ubiquitin, 0.5μg E1 enzyme and 1μg Ubc5a in 60μl reaction buffer as described above. The primary antibodies included polyclonal rabbit anti-ubiquitin (IgG, ABclonal, Cat# A0162), -Sp100A (IgG, Abcam, Cat# ab167605) and -Ubc5a (IgG, Abcam, Cat# ab176561), monoclonal mouse anti-Hes1 (IgG1, Abcam, Cat# ab119776), -Myc (IgG1, Sigma, Cat# M4439), -β-actin (IgG1, Sungene Biotech, Cat# KM9001). The secondary antibodies included HRP-conjugated goat anti-mouse IgG (IgG, Abbkine, Cat# A21010) and goat anti-rabbit IgG (IgG, Abbkine, Cat# A21020). The immunoblotting and immunofluorescence images shown are representative of at least three independent experiments. The chemiluminescent signals of the immunoblotting images were densitometrically quantitated using ImageJ (National Institutes of Health). The relative levels of the target proteins were calculated as ratios to the corresponding control following β-actin normalization. The values are presented below the corresponding blot. The bar graphs were generated by using the GraphPad Prism software package based on the values from three independent experiments, and relative protein levels are represented as average ± standard deviation (SD). Differences were considered statistically significant when P≤0.05 based on a Student’s t-tests analysis. *, P ≤ 0.05; **, P ≤ 0.01.

10.1371/journal.pgen.1005754

Gene Regulatory Mechanisms Underlying the Spatial and Temporal Regulation of Target-Dependent Gene Expression in Drosophila Neurons

Neuronal differentiation often requires target-derived signals from the cells they innervate. These signals typically activate neural subtype-specific genes, but the gene regulatory mechanisms remain largely unknown. Highly restricted expression of the FMRFa neuropeptide in Drosophila Tv4 neurons requires target-derived BMP signaling and a transcription factor code that includes Apterous. Using integrase transgenesis of enhancer reporters, we functionally dissected the Tv4-enhancer of FMRFa within its native cellular context. We identified two essential but discrete cis-elements, a BMP-response element (BMP-RE) that binds BMP-activated pMad, and a homeodomain-response element (HD-RE) that binds Apterous. These cis-elements have low activity and must be combined for Tv4-enhancer activity. Such combinatorial activity is often a mechanism for restricting expression to the intersection of cis-element spatiotemporal activities. However, concatemers of the HD-RE and BMP-RE cis-elements were found to independently generate the same spatiotemporal expression as the Tv4-enhancer. Thus, the Tv4-enhancer atypically combines two low-activity cis-elements that confer the same output from distinct inputs. The activation of target-dependent genes is assumed to 'wait' for target contact. We tested this directly, and unexpectedly found that premature BMP activity could not induce early FMRFa expression; also, we show that the BMP-insensitive HD-RE cis-element is activated at the time of target contact. This led us to uncover a role for the nuclear receptor, seven up (svp), as a repressor of FMRFa induction prior to target contact. Svp is normally downregulated immediately prior to target contact, and we found that maintaining Svp expression prevents cis-element activation, whereas reducing svp gene dosage prematurely activates cis-element activity. We conclude that the target-dependent FMRFa gene is repressed prior to target contact, and that target-derived BMP signaling directly activates FMRFa gene expression through an atypical gene regulatory mechanism.

Nerve cells extend long processes that grow out to contact the target cells with which they communicate. When the nerve cell makes initial contact, the target cells send a retrograde signal back to the nerve cell. Such target-derived signals activate and maintain important genes that make the nerve cell functional, such as genes determining neurotransmitter type. This is a well-characterized phenomenon throughout the nervous systems of flies to mammals, but we still do not know how these signals actually activate gene expression. We now provide details regarding target-dependent signal regulation of nerve cell genes. We model this in Tv4 neurons of Drosophila melanogaster, which require target-derived BMP signaling to trigger FMRFa neuropeptide expression. Our study shows how DNA-binding transcription factors of the BMP signaling pathway integrate with other transcription factors at specific regulatory DNA sequences to activate FMRFa expression, and define the atypical logic by which this occurs. We also provide novel insight into how target-dependent genes are regulated before target contact. Instead of simply waiting for target-dependent activation, these genes seem to be blocked from being expressed prior to target contact. These findings have relevance to mammals because the role of target-derived BMP signaling in nerve cell gene regulation is conserved between vertebrates and invertebrates.

Nervous system development requires the differentiation of diverse neuronal subtypes under the direction of combinatorially acting transcription factors [1, 2]. However, target-derived signaling from axo-dendritic targets, in the form of retrograde bone morphogenetic protein (BMP), transforming growth factor β (TGFβ), neurotrophin, or cytokine signaling, is often required to terminally differentiate a neuron's identity, mature morphology or function [3–6]. Target-dependent genes are often neurotransmitter enzymes or neuropeptides that mediate intercellular communication [7–13], or ion channels that mediate mature physiological properties [14, 15]. In addition, target-derived signaling can induce subtype-specific transcription factor profiles that drive branching of axo-dendritic arbors or appropriate topographic mapping of projections [16–19]. Strong genetic and cellular data supports a role for target-derived signaling in triggering target-dependent and neuronal subtype-specific gene transcription, yet our current view is not well informed by an understanding of the underlying gene regulatory mechanisms. Two broad possibilities have been discussed regarding the role of pleiotropic target-derived signals in triggering subtype-specific gene expression [3, 4]. First, they may contribute by promoting the activity of established transcriptional complexes that pre-determine gene expression. Alternatively, dedicated signaling pathway transcription factors might bind cis-regulatory sequences and contribute alongside cell-specific transcription factors to combinatorially specify gene expression. Here, we examined the gene regulatory mechanisms of target-derived signaling by examining how target-derived BMP signaling triggers FMRFa gene expression selectively in Drosophila Tv4 neurons. In Drosophila, target-derived BMP signaling positively regulates neuromuscular synaptic morphology, transmission and plasticity [20–23], as well as subtype-specific neuropeptide gene expression [12, 13, 24]. Drosophila neuronal BMP signaling is induced by the postsynaptic-secreted Glass Bottom Boat (Gbb) ligand that acts at presynaptic BMP receptors Wishful thinking (Wit), Thickveins (Tkv) and Saxophone (Sax) [13, 20–22]. The type I BMP-receptors, Tkv and Sax, phosphorylate the receptor Smad, Mad (pMad; vertebrate Smad 1/5/8), which then couples with its co-Smad, Medea (vertebrate Smad 4) that together can act as sequence-specific transcription factors, or as transcriptional co-regulators [25–28]. The activities of the BMP and the closely-related TGFβ pathways can diverge from all levels of this linear pathway and feed into other signal transduction or miRNA pathways, providing multiple avenues by which BMP signaling could influence gene regulation [29–32]. The Drosophila ventral nerve cord (VNC) has one Tv4 neuron in each of the six thoracic hemisegments. These six Tv4 neurons express the neuropeptide gene FMRFa that encodes a prepropeptide (FMRFa). The FMRFa prepropeptide is processed to multiple amidated FMRFamide neuropeptides (FMRFamide), which facilitate neurotransmission at the neuromuscular junction, a mechanism required for behaviours such as escape responses [33–36]. Tv4 neurons are born at embryonic stage (Stg.) 14, and their axons innervate the ipsisegmental dorsal neurohaemal organ (DNH) in mid to late Stg. 17 embryos (Fig 1A). Tv4 axons gain access to Gbb at their target. Gbb activates a retrograde BMP signaling that is absolutely essential for FMRFa gene initiation and maintenance throughout the organism's life [13, 37]. A logical genetic explanation for the extreme specificity of FMRFa expression is provided by genetic analysis showing that FMRFa expression requires BMP signaling and a Tv4-specific combination of transcription factors (TFs); the sequence-specific TFs Apterous (Ap), Squeeze (Sqz), Dimmed (Dimm) and Grainy head (Grh), and the transcriptional co-regulators Eyes absent (Eya) and Dachshund (Dac). In gain-of-function studies, a combination of Ap, Dac and BMP-signaling is sufficient to induce strong ectopic FMRFa gene expression in other neurons [13, 38–42] (Fig 1B). We now address how BMP-signaling acts in relation to these known transcriptional regulators to initiate FMRFa gene expression. We identified necessary cis-elements within a 445 bp Tv4-specific FMRFa enhancer (including the homeodomain response element, HD-RE and the BMP-response element, BMP-RE), characterized transcriptional inputs that act at these two cis-elements, and provide an understanding of the developmental information that these two cis-elements contribute to shape FMRFa spatiotemporal expression [42, 43]. We show that induction of the FMRFa gene requires activation of the discrete HD-RE and BMP-RE cis-elements. Ap binds and trans-activates from the HD-RE, while BMP-activated Smads bind and trans-activates from the BMP-RE. Ap coordinates both cis-elements by virtue of its additional indirect regulation of the BMP-RE. Both cis-elements independently generate proper spatial expression, but because both cis-elements have low activity they must be simultaneously activated to generate Tv4-enhancer activity. Finally, we find that proper temporal initiation of FMRFa is produced by an unanticipated bipartite mechanism. Prior to target contact, the nuclear receptor Svp represses both cis-elements. Svp is downregulated immediately prior to target contact, which de-represses the HD-RE and permits the subsequent BMP-dependent activation of the BMP-RE upon target contact. The coordinate de-repression and activation of the HD-RE and BMP-RE in the late embryo then leads to Tv4-enhancer activation and FMRFa expression. A 445 bp cis-regulatory region upstream from the FMRFa gene, that we term the Tv4-enhancer, is sufficient to drive reporter expression exclusively in Tv4 neurons [42, 43]. Tv4-enhancer reporter activity requires apterous, and three candidate Apterous binding sites were postulated to mediate this function [42, 43]. We PCR-amplified the Tv4-enhancer from Oregon R and placed it into a phiC31-integrase-compatible transgenic nEYFP reporter vector, to generate a TvWT-nEYFP reporter transgene integrated into attP2 (Fig 1C, 1D and 1E). We found that TvWT-nEYFP expression faithfully reported FMRFa neuropeptide expression in Tv4 neurons (Fig 1D). We examined TvWT-nEYFP activity in early larval stage 1 (L1) larvae that were mutant for regulators known to affect FMRFa gene expression. We quantified the number (per VNC) of Tv4 neurons expressing nEYFP, as well as the relative intensity of nEYFP in individual Tv4 neurons (normalized to the mean of the control) (Fig 2A, 2B and 2C). Loss of BMP signaling in wishful thinking (wit) type II BMP receptor nulls eliminated FMRFa immunoreactivity and TvWT-nEYFP expression (Fig 2C and 2D). In strong ap hypomorphs, TvWT-nEYFP was expressed in ~2.5 Tv4 neurons per VNC at 58% of control intensity; comparable to the reduction in FMRFa immunoreactivity (Fig 2B and 2D). The co-regulator dac is only modestly required for FMRFa expression in embryos, but its overexpression upregulates FMRFa, and it acts combinatorially with apterous to trigger ectopic FMRFa in BMP-activated motoneurons [41]. In correspondence, in dac nulls, TvWT-nEYFP was expressed in ~5.5 Tv4 neurons per VNC at 72% of control intensity (Fig 2D). Overexpression of UAS-dac in Tv4 neurons (by apGAL4) upregulated TvWT-nEYFP to 144±10% of control levels (p<0.01 two-tailed t-test, n = 48 and n = 56 Tv4 neurons for control and overexpression, respectively). Also, ectopic TvWT-nEYFP expression was activated in motoneurons by OK6-GAL4-driven misexpression of UAS-dac alone, or UAS-dac and UAS-ap together (Figs 2G, 2I and S1). The co-regulator eya is essential for FMRFa expression [41], and TvWT-nEYFP was entirely eliminated in strong eya hypomorphs in late Stg. 17 embryos (Figs 2D and S1). The temporal transcription factor, grh, is required for generation of Tv4 neurons but its expression is reduced by the time of FMRFa expression [38, 44]. Predictably, TvWT-nEYFP expression was eliminated in grh nulls (Figs 2D and S1). Thus, BMP-signaling, ap, dac and eya are regulators of the Tv4-enhancer in postmitotic Tv4 neurons. Previous evidence suggested that FMRFa is sqz and dimm-dependent [45]. Here, our data suggest that this regulation is not directly at the transcriptional level. In sqz nulls, we verified that Tv4 neurons are often not generated in the T1 segment, and that supernumerary Nplp1-expressing Tv1 neurons are generated in Tv clusters (S2 Fig) [38–40]. We quantified TvWT-nEYFP expression in segments T2 and T3, but found no effect on FMRFa immunoreactivity or TvWT-nEYFP in sqz mutants (Fig 2E). In dimm mutants, FMRFamide immunoreactivity was 47% of control levels (p<0.001 Two-tailed t-test, n = 48 and n = 36 Tv4 neurons for dimmRev4/dimmP1 and dimmRev4/+ controls, respectively). In contrast, there was no reduction in the FMRFa prepropeptide or in TvWT-nEYFP (Fig 2D). Also, TvWT-nEYFP was not ectopically activated when we co-misexpressed UAS-ap and UAS-dimm in all motoneurons, using OK6-GAL4 (Fig 2H and 2I). This corresponds to our previous findings in adults showing that dimm knockdown eliminated FMRFamide but not FMRFa prepropeptide or FMRFa transcript [37]. Thus, we eliminate grh, sqz and dimm as direct regulators of the Tv4-enhancer. Our genetic analyses found that Ap, BMP signaling, Dac and Eya regulate Tv4-enhancer activity. As only Ap and BMP-activated Smads are sequence-specific transcription factors, we looked for potential binding motifs in the Tv4 enhancer. DNA sequence motifs for binding of Apterous and the Smads, Mad and Medea have been determined [42, 46–48]. Candidate sequences matching these motifs were identified within the Tv4-enhancer; and using phastCONS in the UCSC Genome Browser and EvoprinterHD [49] across 12 Drosophila species (Figs 1C, 1E, S3 and S4). These include three previously-described, putative Apterous motifs (HD-A,B,C) [42] (Figs 1C, 1E, S3 and S4), and six GC-rich sequences with some similarity to Mad motifs (Mad-A-F) [50, 51] (Figs 1C, 1E, S3 and S4). Of these, only Mad-D is perfectly conserved and precisely matches a characterized Drosophila Mad sequence [GGCGCCA] [47] (Figs 1C,1E and S3). Drosophila Mad and Medea typically act at a bipartite motif, such as [GGCGCCA(N4)GNCV] [47] or [GRCGNC(N5)GTCT] [48]. Only one region approximates either of these motifs; Mad-A is 6 bp from a GTCT sequence (Med-A), but this is inverted with respect to its typical orientation, and is 15 bp from a poorly conserved GTCT sequence, (Med-B) [GGGCCGTAATTACAGACTTCCGTCT] (Figs 1E, S3 and S4). The juxtaposition of the Mad, Ap, and Med bindings sites in this sequence was suggestive of an Ap/Smad integration site. Homeodomain TFs can act cooperatively or collaboratively with Smads at coupled motifs [52–56]. Such a model could account for restriction of FMRFa expression in Tv4 neurons, as Ap and BMP signaling in the VNC only coincide in Tv4 neurons. To identify essential sequences in the Tv4-enhancer and to also directly test putative Ap and Smad binding motifs, we performed deletion and substitution studies of the Tv4-enhancer. We placed each mutant Tv4-enhancer reporter transgene into the genomic attP2 site, to allow for quantitative comparison of all wildtype and mutant Tv4-enhancers, in vivo in their appropriate cellular context, the Tv4 neurons. Exact details of deletions and substitution mutation can be found in S1 Table. We quantified the number of Tv4 neurons that express nEYFP (in early L1 larvae), as well as fluorescence intensity normalized to the mean of the control (Fig 3A, 3B and 3C). A reporter-only empty vector control was used as the relative zero; Tv4 neurons with nEYFP reporter intensity above the upper 99% confidence interval for the empty vector control (9.7% of TvWT-nEYFP) were counted as expressing nEYFP. Reporter expression was reduced in most sequence deletions and mutants, but our combined analysis pinpointed two specific regions that are absolutely required for expression, and also contain putative Ap or Mad sequence motifs. Deletions that removed the short conservation islands containing either the HD-A or the Mad-D motifs eliminated expression (Fig 3B3, 3B4 and 3B5). Further, substitution mutants at the HD-A or Mad-D motif also eliminated reporter expression (Fig 3D). Thus, both HD-A and Mad-D cis-elements are required non-redundantly (Fig 3B4 and 3B5). Deletion of most regions outside these two cis-elements had partial or no effect on reporter expression. We also found that deletion of the low conservation region between HD-A and Mad-D abrogated reporter expression (Fig 3B7). To discriminate whether this region has informational content or acts as a simple spacer between HD-A and Mad-D, we mutated it in two ways; a complement sequence to maintain local GC/AT content, and also a non-canonical nucleotide transversion [57]. In both cases, reporter expression was abrogated (Fig 3B8 and 3B9). Thus, this intervening sequence is essential for expression and does not merely act to space HD-A and Mad-D to an appropriate distance. We conclude that the region spanning HD-A to Mad-D is absolutely critical for Tv4-enhancer activity. This region contains the critical HD-A homeodomain motif and the critical Mad-D Mad motif, that together flank a critical low conservation sequence with no predicted binding motif for known FMRFa regulators. In the nervous system, the coincidental expression of Ap with BMP activity appears to be unique to Tv4 neurons. Also, overlap of Dac and BMP-signaling in ventral nerve cord neurons is extremely rare [12]. Functionally, misexpression of Ap, Dac in BMP-activated motoneurons, or co-misexpression of Ap, Dac and BMP activation, is sufficient for widespread ectopic FMRFa expression in neurons [41]. Thus, given that Ap, Dac and BMP are combinatorially necessary and sufficient for FMRFa expression, we hereafter focused on the two cis-elements containing the HD-A and Mad-D motifs, as these are likely critical integration sites through which Ap and BMP-signaling generate the Tv4-enhancer’s spatiotemporal expression. What cis-regulatory information is encoded by the HD-A- and Mad-D-containing cis-elements? The HD-A sequence is flanked by Mad-A and Med-A, representing a prime candidate site for cooperativity of Ap and BMP-activated Smads. To explore mechanisms for Ap and Smad integration, we tested reporter activity generated by the cis-elements containing HD-A and Mad-D. We placed the conserved island containing HD-A (25 bp spanning Mad-A, HD-A, Med-A) and Mad-D (39 bp), separately, into integrase-compatible reporter vectors to generate attP2-integrated transgenic reporter flies. Monomers of either cis-element failed to generate reporter activity. However, concatemeric repeats of either cis-element generated reporter activity in Tv4 neurons, with robust expression occurring in 6xHD-A-nEYFP and 4xMad-D-nEYFP reporters. Remarkably, expression generated from either cis-element concatemer was highly Tv4-specific; ectopic expression was not observed in the VNC, and was found in only a few cells in the brain (for 6xHD-A-nEYFP) or late L3 eye imaginal disc (for 4xMad-D-nEYFP) (Figs 4B, 4D and S5). We also tested tetrameric concatemers of other regions from the Tv4-enhancer, but these failed to generate Tv4 neuron expression (Fig 4C, 4E and 4F). Thus, even though the short cis-elements containing HD-A and Mad-D have distinct sequences and are both required in the native Tv4 enhancer, each cis-element contains sufficient sequence information for Tv4 neuron expression. The Tv4-specific reporter expression of each concatemer provides us with the tools to determine which transcriptional regulators act at each cis-element. First, we examined BMP-dependence, as Mad motifs are present in both cis-elements. 6xHD-A-nEYFP was not affected in wit null mutants, or after blockade of retrograde BMP signaling, using apGAL4 to drive UAS-GluedΔ84, a truncated allele of p150Glued that blocks dynein-dependent retrograde transport, nuclear pMad accumulation and FMRFa expression [13] (Fig 5A). In contrast, 4xMad-D-nEYFP expression was severely reduced in wit nulls and UAS-GluedΔ84 (Fig 5C). Overexpression of Mad1 (UAS-Mad1) also eliminated 4xMad-D-nEYFP expression (Fig 5C); Mad1 cannot bind DNA but it is phosphorylated, couples to Medea and accumulates in the nucleus normally [58]. We tested the sequence-specificity of Mad binding to the Mad-D motif by electrophoretic mobility shift assay (EMSA) (Fig 5G). Purified GST-MH1-Mad (comprising the DNA-binding MH1 domain) band shifted a Mad-D region DNA probe in a GGCGCC sequence-specific manner (Fig 5G). Thus, the Mad-D sequence is necessary for activity of the Tv4-enhancer, exhibits BMP-dependent expression as a concatemer, and binds Mad in a sequence-specific manner. Henceforth, we termed this cis-element the BMP-Response Element (BMP-RE). In ap mutants, 6xHD-A-nEYFP was significantly down-regulated to 27% of controls (Fig 5A). This was expected due to the consensus Ap-binding sequence in HD-A, and the previous biochemical evidence for Ap binding to this motif [42], as well as the importance of the HD-A to Tv4-enhancer activity. EMSA analysis supported this; GST-CtermAp (the C terminal half of Ap that includes the homeodomain but excludes the LIM domains) band-shifted an HD-A sequence DNA probe in a TAATTA sequence specific manner (Fig 5E). Unexpectedly, 4xMad-D-nEYFP was also reduced to 23% of control intensity in ap mutants (Fig 5D), in spite of the lack of a putative Ap binding site. GST-CtermAp failed to band shift the BMP-RE sequence. Thus, we postulate that the genetic regulation of the BMP-RE by Apterous is likely mediated via Ap-dependent activation of another transcription factor (Fig 5F). We henceforth term the HD-A cis-element the Homeodomain-Response Element (HD-RE). We examined HD-RE and BMP-RE responsiveness to dac and eya. Both cis-elements are eliminated in eya mutants (Fig 5A and 5E) and upregulated by ~400% by dac overexpression in Tv4 neurons (Fig 5B and 5D). Thus, both co-regulators coordinately regulate trans-activation from both cis-elements. We conclude that Ap and Mad are recruited to the native Tv4-enhancer at distinct cis-elements; the Ap-recruiting HD-RE and the Mad-recruiting BMP-RE. While this may explain the combinatorial requirement for both cis-elements, our concatemer analysis shows that Tv4-specificity does not necessarily emerge from it being the point of intersection of the partially overlapping spatial activities of these two cis-elements. Instead, each cis-element independently encodes sufficient information for Tv4-specific expression. This is not explained solely by the activities of Ap or BMP, as these are present in other non-overlapping neuronal populations. Thus, Tv4-specificity presumably requires additional unknown inputs acting at each of these cis-elements. For the BMP-RE, this is perhaps an Ap-dependent transcription factor, as Ap is required for BMP-RE reporter activity but does not bind. The HD-RE and BMP-RE both encode the same spatial information. Ap acts as a central coordinator of the activity of both cis-elements. BMP-signaling acts via Smads at a BMP-RE cis-element, yet it also implicitly carries with it a temporal-encoding facet; the requirement for BMP-driven Smads ensures that FMRFa expression is inevitably tied to target contact in late Stg. 17 embryos. This led us to test the following model: The HD-RE is activated after Tv4 neuron birth, but its weak activity is not sufficient to initiate FMRFa expression at this early time. Once the target is contacted and BMP signaling is initiated, the BMP-RE becomes activated, and the combined activities of the HD-RE and the BMP-RE become sufficient to initiate FMRFa trans-activation. This model predicts that the HD-RE (6xHD-A-nEYFP) would initiate reporter expression prior to target contact because all known regulators of the HD-RE are present before target contact, and the cis-element's activity does not require target-derived BMP-signaling. In contrast, the BMP-RE (4xMad-D-nEYFP) would initiate only after target contact because it requires BMP-signaling for its activity. Upon testing this, we unexpectedly found that the HD-RE initiates reporter activity at the same time as BMP-RE and FMRFa, at late Stg. 17 (Fig 6A–6D). This paradoxical observation cannot be explained by another retrograde signal acting at the HD-RE, as UAS-GluedΔ84 expression did not affect HD-RE reporter activity (Fig 5A). Thus, we postulated that the HD-RE must be responsive to a novel timer that is separate to, but coincidental with, target contact. We reasoned that the existence of a second timer would be further supported if precocious BMP activity in Tv4 neurons could not initiate early FMRFa expression, prior to the normal time of target contact. We tested this using apGAL4 to drive excess Mad (UAS-myc::mad) that was phosphorylated by co-expression of UAS-TkvAct and UAS-SaxAct [13]. This generated high pMad immunoreactivity in Tv neurons by Stg. 16, prior to target contact (Fig 6E–6H). Remarkably, this had no effect on the initiation time of FMRFa. It initiated expression at its normal time at late Stg. 17 (Fig 6F and 6H), and failed to activate precociously at late Stg. 16 (n = 42 Tv clusters each for control and experimental) or even mid Stg. 17 (n = 48 Tv clusters for control and experimental) (Fig 6G). Thus, FMRFa does not simply 'await' target contact and BMP-dependent activation, as is generally assumed for target-dependent gene expression, but its expression is somehow prevented prior to target contact. Towards identifying a second putative timer, we tested a possible role for the nuclear receptor seven up (svp). Svp is required early in the NB5-6T neuroblast lineage (that gives rise to Tv neurons) as a switching factor that triggers the hunchback to Kruppel temporal transcription factor transition, in part through downregulating hunchback expression by an unknown mechanism [59, 60]. A second pulse of Svp expression occurs later in this neuroblast at the time of Tv neuron generation. Its expression is initially retained in all newly born Tv neurons, but then becomes downregulated in Tv1 neurons by Stg. 16, and in Tv4 neurons by early Stg. 17 [60]. This second pulse is required for the appropriate diversification of Tv1-4 neuron subtypes. Lineage confusion amongst Tv neurons is observed in the few Tv neuronal clusters that are generated in svp nulls. For example, the Tv4 neuron is not generated, supernumerary Nplp1 Tv1 neurons are produced, and transcription factor expression profiles suggest that many Tv neurons have mixed identities. The authors concluded that svp acts during Tv neuron lineage progression to generate diversity amongst Tv1-Tv4 neurons [60]. The coincidental timing of Svp downregulation and FMRFa initiation was intriguing, and prompted us to test the hypothesis that Svp acts in postmitotic Tv4 neurons to repress FMRFa up to the time of target contact. First, we tested the effect of maintaining Svp expression beyond its normal time of downregulation (at Stg. 17), by driving UAS-svp from apGAL4. This was previously shown to prevent expression of Dimmed, Nplp1 and FMRF in Tv1 and/or Tv4 neurons [60]. In early L1 larvae, we found that this eliminated FMRFa, HD-RE and BMP-RE reporter expression (Fig 6K–6N), but did not block pMad accumulation in Tv4 nuclei (Fig 6I and 6J). This overlap of pMad and apGAL4, UAS-nGFP allowed us to uniquely identify Tv4 neurons when UAS-svp is overexpressed, in spite of loss of FMRFa. This allowed us to examine if the expression of essential transcriptional regulators of FMRFa were downregulated in Tv4 neurons. We found that expression of all the confirmed regulators of FMRFa were expressed normally, including pMad, Dac, Eya and ap (apGAL4,UAS-nGFP) (Figs 6 and S6). Thus, unique Tv4 neuron identity and retrograde BMP signaling were unaffected by persistent Svp. This shows that maintained Svp expression blocks terminal differentiation of FMRFa expression in Tv4 neurons, rather than prevents the generation, axon targeting or BMP activation of Tv4 neurons. The coincidence of Svp downregulation with Tv4 neuron target contact suggested that BMP activation might be the trigger for Svp downregulation. We tested this by examining Svp immunoreactivity in wit mutants. Using 6xHD-A-nEYFP to identify Tv4 neurons in wit mutants, we found that Svp immunoreactivity was downregulated at its normal time in the absence of BMP signaling (Fig 6O and 6P), indicating that BMP activation does not downregulate Svp. We next examined whether the effect of Svp on the HD-RE and BMP-RE is mediated by direct Svp binding. Previous characterization including high throughput studies had identified a core Svp bipartite motif of two GGTCA half-sites separated by a short spacer [61–63]. We first confirmed that full-length recombinant Svp can shift a labeled DNA probe containing the characterized DR1 bipartite Svp binding sites in a sequence specific manner, using established conditions [62] (S7A Fig). Next, we found a near-consensus Svp half-site in the HD-RE and an adjacent half-site 4 bp away and outside the HD-RE (S7B Fig) Using an extended probe that included both of these candidate half sites, we examined whether recombinant Svp could bind this DNA sequence probe, in a sequence specific manner. A weak band shift of this extended HD-RE region was observed in the presence of recombinant Svp, using the same conditions as used for the DR1 element (S7B Fig Lanes10-15). Cold competitor sequences with mutated putative Svp binding sites competed as efficiently as wild-type competitors. Thus, Svp does not bind HD-RE in a sequence-specific manner. Addition of poly-dI-dC strongly reduced the HD-RE band shift (S7B Fig Lanes 16–19). The BMP-RE lacked any putative Svp binding sites and did not show any appreciable band shift in the presence of recombinant Svp (S7C Fig) Collectively, these data fail to support a model in which the HD-RE or BMP-RE are regulated by Svp through direct binding. These results suggest that Svp gates FMRFa initiation by indirectly preventing HD-RE and BMP-RE activity prior to target contact. Immediately prior to target contact, Svp expression is downregulated, which de-represses the HD-RE and BMP-RE. At the time of target contact, BMP signaling can then activate expression. This model would predict that loss of svp should activate FMRFa prematurely. Testing this directly is complicated by two factors. Tv4 neurons are not generated in svp nulls, and FMRFa initiation requires BMP activation that is temporally coincident with Svp downregulation. To test our hypothesis in a way that avoids these confounding factors, we examined the timing of HD-RE (6xHD-A-EYFP) initiation in strong svp hypomorphs, because this cis-element is BMP-insensitive but is driven by Ap, Dac, and Eya that are all expressed from the birth of the Tv4 neuron. We found that svp1/svp2 generates a normal number of correctly specified Tv neurons, including a single Nplp1-expressing Tv1 neuron and a single FMRFa-expressing Tv4 neuron (S6E Fig), yet is a strong enough allelic combination to generate other svp lineage phenotypes [64]. We tested the initiation time of HD-RE (6xHD-A-nEYFP) at numerous developmental stages in control and svp mutant embryos: Early Stg. 17, when the gut is beginning to fold but is not yet showing great complexity; Mid Stg. 17, at a time before the trachea starts to air fill; Air-filled trachea (AFT) stage, when the trachea is filling or filled but before mouth hooks form; Air-filled trachea and mouth hooks (AFT/MH), when both structures are well developed immediately prior to hatching. In controls, heterozygotic 6xHD-A-nEYFP was expressed in 28% of Tv4 neurons by AFT/MH and not at any stage prior. In contrast, in svp1/svp2 mutants, we observed reporter expression in 61% of Tv4 neurons at AFT/MH, also in 35% of Tv4 neurons at AFT, and in 14% of Tv4 neurons at mid Stg17 (Table 1). This shows that the reporter is initiated at an earlier timepoint in svp hypomorphs, and that its expression is more robust that in controls by late Stg. 17. This premature initiation time was svp dose-dependent, since svp1/+ heterozygous reporter expression was observed in 36% of Tv4 neurons at AFT/MH and prematurely in 26% of Tv4 neurons at AFT. We conclude that two independent timers together regulate the timing of FMRFa initiation. The first timer is Svp that acts as an intrinsic repressor that prevents HD-RE and BMP-RE activity prior to target contact. Downregulation of Svp immediately prior to target contact de-represses HD-RE and BMP-RE activity. The second timer is target-activated BMP signaling that directly activates FMRFa via Mad binding to the BMP-RE cis-element. Interestingly, although these two timing events are temporally coincidental, we find no evidence for a cross-regulatory genetic hand-over from one timer to the other. Target-dependent gene expression in many neurons is initiated upon contact of axons and/or dendrites with their target cell(s), but the underlying gene regulatory mechanisms are largely unexplored [3, 4]. Here, we examined these gene regulatory mechanisms, using initiation of the FMRFa gene in Tv4 neurons by target-dependent BMP-signaling as a model. We uncover key cis-regulatory sequences in a Tv4-enhancer of the FMRFa gene that integrate the necessary and combinatorially sufficient inputs of Ap, Dac, Eya and BMP-activity to generate FMRFa expression in Tv4 neurons upon target contact. These studies show that BMP-signaling contributes through Smad binding at an essential cis-element, and reveals surprising complexity in the integration of intrinsic and extrinsic inputs at the FMRFa enhancer. In addition, we provide evidence to support an hypothesis that target-dependent genes are repressed prior to target contact (in this case by svp), rather than simply awaiting activation. These genes become de-repressed around the time of target contact in order for the target-derived signal to be able to directly activate the gene's expression. We aimed to identify the cis-regulatory sequences and core regulatory mechanisms through which a genetically identified set of regulatory inputs determines the Tv4-specific expression of FMRFa upon target-derived BMP-signaling. We found that activity of the Tv4-enhancer requires the sequence-specific regulators Ap and BMP-activated Smads and the co-regulators Dac and Eya. As these inputs are all genetically necessary and combinatorially sufficient for FMRFa expression [13, 41], we focused on the mechanisms through which these critical regulators specify FMRFa expression. We found that Ap and BMP-activated Smads bind directly at the Tv4-enhancer, but that their binding is parsed onto two distinct and essential cis-elements, the HD-RE and BMP-RE, respectively. Thus, the combinatorial requirement for Ap and BMP appears to be conferred by the integration of both essential cis-elements. This indicates that BMP-signaling acts directly at the FMRFa enhancer. We propose that BMP-signaling forms part of the combinatorial code of transcriptional inputs that together specify FMRFa gene expression, as opposed to triggering the transcriptional activity of a transcriptional complex that is pre-established at the FMRFa enhancer. We identified two levels of regulatory coordination between the two cis-elements. Ap binds the HD-RE and is necessary for its activity, but Ap is also required indirectly for BMP-RE activity without direct binding. We postulate that Ap likely regulates the expression of an unknown transcription factor that binds and activates the BMP-RE, but verification of this model awaits the identification of Ap-dependent transcription factors acting at the BMP-RE. Also, we found that Dac and Eya are both important co-regulators that mediate the activities of both cis-elements. Eya in part mediates this effect by regulation of BMP-activity in Tv4 neurons and does not contribute to ectopic FMRFa expression when Ap, Dac and BMP signaling are present [41]. In contrast, we show here that Dac is a potent amplifier of HD-RE, BMP-RE and FMRFa expression in late embryos and early larvae. Dac does not appear to be required for the native low-level FMRFa expression in the late embryo and early L1 larval stage. However, Dac becomes essential for the high level expression of FMRFa thereafter [37], as well as for generation of ectopic FMRFa expression induced by Ap and BMP-signaling in other neurons. Although the function of Dac in gene regulation is still ambiguous in most contexts, it is generally viewed as a co-regulator that recruits histone modifying complexes and the mediator complex [65–68]. Thus, we postulate that Dac may promote a chromatin state that facilitates high-level transcriptional activation downstream of Ap/Smad engagement of FMRFa cis-regulatory sequences. Such a model will require detailed analysis, and likely also identification of other transcription factors acting at the HD-RE and BMP-RE cis-elements that may be required for recruitment of Dac. In this light, it is interesting to note that DNA-bound vertebrate Smad4 has been shown to recruit Dach1, which acts in that context as a co-repressor that recruits the nuclear receptor co-repressor (N-CoR), that in turn recruits histone deacetylases [69, 70]. Towards identifying the information that each cis-element contributes to overall Tv4-enhancer activities, we generated concatemers of the HD-RE and BMP-RE cis-elements. Unexpectedly, both HD-RE and BMP-RE concatemers independently generated the same spatiotemporal pattern as the full Tv4-enhancer. Thus, taken together with our finding that both cis-elements are required in the native Tv4-enhancer context, we conclude that the HD-RE and BMP-RE are low activity cis-elements required in combination for FMRFa expression but that encode the same spatiotemporal information from distinct inputs. These results were not expected, and dispel the simplest prediction that the HD-RE receives cell-specific transcription factor input contributing spatial information, while the BMP-RE receives the extrinsic BMP input contributing temporal information. Such a model would have been in line with evidence from examination of other enhancers in which the correct spatiotemporal expression is generated by combining the activities of distinct spatial and temporal encoding cis-elements [69–76]. However, the Tv4-enhancer does not appear to act as an integrator of differential spatial and temporal inputs encoded via these two cis-elements, as both cis-elements encode full spatiotemporal information from their respective developmental inputs. It is unclear why two cis-elements encoding the same spatiotemporal information are utilized, when either one could conceivably function alone. One rationale could derive from the small amount of ectopic, non-overlapping expression that is generated by the HD-RE or BMP-RE concatemers. Such non-overlapping ectopic expression may indicate that these cis-elements have low-level activity so as to restrict FMRFa activation only to cells where both cis-elements are activated. Indeed, attenuation of cis-element activity to restrict target gene expression has been demonstrated, via reduced transcription factor affinity or by inclusion of repressive elements [77, 78]. Another mechanism may be related to the ability of multiple weak cis-elements to generate robust and specific gene expression. For example, shadow enhancers are cis-elements with similar spatiotemporal outputs that act redundantly (to varying degrees) in normal conditions, but are required together for robust output in adverse conditions [79, 80]. Also, the addition of increasing numbers of redundant but individually weak cis-elements was shown to increase the robustness of Sonic hedgehog gene expression in different mouse tissues [81]. Moreover, robust and specific expression can be achieved by the accumulation of multiple low activity cis-elements; multiple Ultrabithorax binding sites are required together for spatially-restricted repression of spalt in the Drosophila haltere [82], and multiple weak Ultrabithorax-Extradenticle binding sites drive shaven baby in Drosophila epidermal tricomes [83]. Thus, the use of two discrete low activity cis-elements that generate the same spatiotemporal output from different developmental inputs may offer a solution for integrating all the appropriate spatial and temporal inputs into robust, exquisitely specific activity in only 6 neurons of the nervous system. Our analysis raises some unresolved questions. First, the specificity of the HD-RE and BMP-RE cis-elements remains unexplained, as Ap and BMP activity cannot alone explain HD-RE and BMP-RE spatiotemporal expression. Both regulators are active in many other neurons, yet the HD-RE and BMP-RE concatemers are not expressed in these neurons. Thus, unknown regulators must act with Ap or Smads at these cis-elements. We aim to identify those transcription factors in ongoing screens, because models that account for the Tv4-specificity of either cis-element will require incorporation of those transcription factors' activities. Second, the low conservation region between the HD-RE and BMP-RE contains sequence-specific information that is critical for enhancer activity. At this time, no identified transcriptional regulator has been predicted or shown to act at this region. Future analysis of this region awaits the identification of transcription factors that may act within this region. Finally, deletion or point mutagenesis of sequences 3' of the BMP-RE identify other regions that contribute to overall expression level. However, because none of these regions were found to be absolutely critical for enhancer activity in our assays, we did not focus on these in this study, and their precise contribution remains untested. The developmental initiation of target-dependent genes in neurons requires target contact and target-derived signaling, making it reasonable to assume that these genes simply wait to be activated prior to target contact. However, our seemingly paradoxical results regarding the timing of FMRFa activation lead us to a novel model wherein target-dependent genes are repressed prior to target contact: First, the BMP/target-insensitive HD-RE cis-element initiated expression at the same time as the BMP-RE cis-element and FMRFa itself. Second, precocious BMP activation failed to initiate FMRFa at an earlier timepoint. These data suggested that the HD-RE cis-element responds to another timer that prevents BMP-dependent FMRFa activation prior to target contact. We considered two possibilities: First, an unknown and necessary regulator is not expressed until the time of target contact. Second, a repressor is active prior to target contact. Our evidence supports the second, novel model. Previous work had shown that Svp is downregulated immediately prior to target contact [60]. Here, we found that this downregulation is required to de-repress the Tv4-enhancer via both the HD-RE and BMP-RE, as maintained Svp expression blocks the induction of HD-RE, BMP-RE and FMRFa. Moreover, we show that HD-RE cis-element expression initiates at increasingly earlier time points as svp dosage is reduced. This demonstrates that Svp expression level gates the initiation time of this cis-element. The mechanism by which Svp represses FMRFa is unknown; it does not alter the expression of known FMRFa regulators, and EMSA analysis did not support a role for direct Svp-binding to the HD-RE or BMP-RE cis-elements. Possible mechanisms include regulation of the expression of unidentified essential transcription factors, or direct interference by Svp on the transcriptional activities of transcription factors or chromatin modifiers. The seven up gene is an intriguing factor to play a role in gating the timing of terminal differentiation. In both Drosophila and vertebrates, Svp (vertebrate COUP-TF I/II) is a temporal switching factor that mediates transitions in the developmental potential of neuroglial lineages (reviewed by [59, 84]). In Drosophila, a transient Svp pulse triggers the hunchback to Kruppel switch, by repressing hunchback, in the neuroblast temporal transcription factor cascade in multiple lineages [64, 85], including in the NB5-6T lineage [60]. Also, in late larvae, a transient pulse of Svp is required to switch neuroblasts from expressing Chinmo to expressing Broad-Complex, which switches the fate and size of neuronal progeny [86]. Svp also acts as a sub-temporal switch to increase the diversity of Tv1-4 neuronal fates generated through the Castor/Grainy head temporal window late in the NB5-6T lineage [60]. Such switching roles are well conserved in vertebrates. The svp orthologs COUP-TFI/II are transiently expressed and required to switch numerous progenitor lineages from generating neurons to generating glial cells [87]. In spite of these many characterized switching roles for Svp/COUP-TFI/II, neither the regulation of Svp/COUP-TFI/II pulses nor its downstream molecular actions are well understood in any system. In conclusion, our work reveals the complex cis-regulatory mechanisms of neuronal subtype-specific and target-dependent gene initiation in the context of the target/BMP-dependent induction of FMRFa in Tv4 neurons. Detailed functional analysis of the cis-regulatory architecture of other target-dependent neuronal genes will determine whether principles learned here are unique to the FMRFa gene, or generalizable to most target-dependent genes. The following strains were used: sqzie and UAS-sqz [13]; UAS-ap, apRK506 (apLacZ) [88]; apP44 and apmd544 (apGAL4) [89]; dac3 [90]; UAS-dac [91]; eyaCli-IID [92]; eyaD1 [93]; dimmrev4 and dimmP1 [45]; grhIM [94]; Df(2R)Pcl7B (grhDf) [95], OK6-GAL4, witA12 and witB11 [22]; svp1 and svp2 [96]; UAS-Mad1 [58]; UAS-GluedΔ84 (UAS-GluedDN) [97]; UAS-tkvAct and UAS-saxAct [98]; UAS-myc::Mad [99]; UAS-svp type I (UAS-svp) [100]; UAS-nls.EGFPP; Df(2L)Exel7066 (DacDf) (Bloomington, IN). Mutants were kept over CyO,Act-GFP TM3,Ser,Act-GFP or CyO, twi-GAL4,UAS-2xEGFP or TM3,Sb,Ser,twi-GAL4,UAS-2xEGFP. w1118 was used as the control genotype. Flies were maintained at 25°C, 70% humidity. The empty Tv-nEYFP vector was generated from pUASTattB [101] digested with NheI/SpeI and blunted with Klenow fragment. The LoxP and attB sequences from pUASTattB and the multiple cloning site (MCS), HSP70 promoter, Tra nuclear localization signal and SV40-polyA sequences from pHstinger [102] were joined with EYFP from pDUAL-YFH1c [103] using SOE PCR to produce an EcoRV-loxP-MluI-MCS-hsp70 pro-EYFP-tra.nls-SpeI-SV40polyA-AvrII-attB-ZraI cassette that was digested with EcoRV/ZraI and ligated into the blunted pUASTattB backbone. The Wild type Oregon R Tv4-enhancer was PCR-amplified with EcoRI / XbaI adaptors in the primers, restriction digested and ligated into XbaI/EcoRI digested empty Tv-nEYFP. Nucleotide substitution and deletion mutants were generated by SOE PCR and similarly inserted into EcoRI / XbaI sites. The XbaI and NheI sites were used for the concatemers. Fly transformations were performed by Genetic Services Inc. (Cambridge, MA.) All transgenes were integrated into attP2 [104]. The Oregon R Tv4-enhancer contains two single nucleotide polymorphisms (SNPs), compared to the reference genome (v4 to v6). Recently sequenced wild Drosophila species concur with the Oregon R sequence; thus it is the reference genome that contains atypical SNPs [105]. A summary of all mutations and concatemerization sequences can be found in S1 Table. Standard protocols were used throughout [106]. Primary antibodies: Rabbit and Chicken α-FMRFa C-terminal peptide (1:1000) [39], Rabbit α-FMRFamide (1:2000; T-4757 Peninsula Labs); Chicken α-ß-Gal (1:1000, ab9361, Abcam); Guinea Pig α-Dimm (1:1000) [39]; Mouse α-Eya (1:100; MAb clone 10H6) and Mouse α-Dac (1:2; MAb Dac 2-3clone) (DSHB; Iowa U., Iowa); Rabbit α-pMad (1:100, 41D10, Cell Signaling Technology); Mouse α-Svp (1:50) [60]. Secondary antibodies: Donkey anti-Mouse, anti-Chicken, anti-Rabbit, anti-Guinea Pig (H+L) conjugated to DyLight 488, Cy3, Cy5 (1:100, Jackson ImmunoResearch). More than 5 animals were examined for every genotype. Analysis on the 445 bp Tv-enhancers was performed on homozygous reporter lines. Concatemerized cis-elements were analyzed as heterozygous reporters. Images were acquired with an Olympus FV1000 confocal microscope with settings that avoided pixel intensity saturation. Fluorescent intensity of individual Tv4 neurons was measured (or from Eya-positive Tv cluster when no Tv4 marker was detectable) in Image J (US National Institutes of Health). Mean pixel intensity for each neuron was measured from summed Z-projection and background fluorescence for subtraction was measured from an adjacent location using the same size circular region of interest. Each datum point of resulting nEYFP intensity was used to calculate mean intensity for a genotype or enhancer variant; each datum point was then represented as a percentage of the mean of the control group. Representative images of Tv neurons being compared in Figs were linear contrast enhanced together in Adobe Photoshop CS5 (Adobe Systems, Mountain View, CA). All statistical analysis and graphing were performed using Prism 5 (GraphPad Software, San Diego, CA). All multiple comparisons were done with One-Way ANOVA and a Tukey post-hoc test or Student’s two-tailed t-test when there were only two groups. Differences between groups were considered statistically significant when p<0.05. Data are presented as mean ± Standard Error of the Mean (SEM). Recombinant GST-CtermAp (LIM domains removed) and GST-MadN (the MH1 domain), were fused to the GST in pGEX6p1 (GE Health), expressed in Rossetta bacteria cells (EMD Millipore, Billerica, MA), purified using Glutathione-Sepharose beads (GE Health), and dialyzed into 20 mM HEPES pH 7.8, 50 mM KCl, 1 mM DTT, and 10% glycerol. Aliquots were stored at -80°C. Full length svp Type I cDNA [96] was cloned from UAS-svp I [100]. An N-terminus His tag was added to the Svp Type I cDNA and inserted into pGEX6p1. The GST-tagged His::Svp was expressed in Rossetta bacteria cells and purified on Glutathione-Sepharose beads as above, but with the removal of the GST tag using the PreScission Protease kit according to manufacturer’s instructions (GE Health) before concentration and storage as above. Oligonucleotide probes were synthesized and labeled with IRDye 700 by (IDT Inc, Coralville, IA). Gel shift assay for HD-RE or BMP-RE and Apterous binding was performed by incubation (30 min at 21 0C) of 1 μg of GST-CtermAp with 1 μl of 100 nM probe in a 20 -μl reaction buffer (20 mM HEPES pH7.8, 50 mM KCl, 10% glycerol, 0.25 mM EDTA, 0.1 mg/ml BSA, 1 mM DTT). Gel shift assay for BMP-RE and Mad binding was performed by incubating 300 ng of GST-Mad with 1 μl of 50 nM probe in 20 μl reaction buffer (25 mM Tris pH 7.5, 35 mM KCl, 80 mM NaCl, 5 mM MgCl2, 3.5 mM DTT, 0.25% Tween 20, 1 μg poly(dI-dC), and 1x Protease Inhibitor cocktail (Roche)). Svp binding to all tested probes was performed by incubating 1 μg of His::Svp with 1 μl of 50 nM probe in 20 μl Svp binding buffer (100 mM KCl, 7.5% glycerol, 20 mM HEPES pH 7.5, 1 mM DTT and 0.1% Nonidet P-40) on ice for 15 min with or without 1 μg of poly(dI-dC) [62]. For competition assays, unlabeled DNA sequences or mutant DNA sequences were incubated with the labeled probes. See Figs for stoichiometric ratio of unlabeled to labeled probe for each EMSA. DNA-protein complexes were resolved on a 4% non-denaturing polyacrylamide gel, and imaged immediately on a Licor Odyssey Imager system (Lincoln, NE.)

10.1371/journal.pcbi.1005625

Identifying direct contacts between protein complex subunits from their conditional dependence in proteomics datasets

Determining the three dimensional arrangement of proteins in a complex is highly beneficial for uncovering mechanistic function and interpreting genetic variation in coding genes comprising protein complexes. There are several methods for determining co-complex interactions between proteins, among them co-fractionation / mass spectrometry (CF-MS), but it remains difficult to identify directly contacting subunits within a multi-protein complex. Correlation analysis of CF-MS profiles shows promise in detecting protein complexes as a whole but is limited in its ability to infer direct physical contacts among proteins in sub-complexes. To identify direct protein-protein contacts within human protein complexes we learn a sparse conditional dependency graph from approximately 3,000 CF-MS experiments on human cell lines. We show substantial performance gains in estimating direct interactions compared to correlation analysis on a benchmark of large protein complexes with solved three-dimensional structures. We demonstrate the method’s value in determining the three dimensional arrangement of proteins by making predictions for complexes without known structure (the exocyst and tRNA multi-synthetase complex) and by establishing evidence for the structural position of a recently discovered component of the core human EKC/KEOPS complex, GON7/C14ORF142, providing a more complete 3D model of the complex. Direct contact prediction provides easily calculable additional structural information for large-scale protein complex mapping studies and should be broadly applicable across organisms as more CF-MS datasets become available.

Proteins physically associate into complexes in order to carry out the essential functions of life. Knowing how proteins are physically arranged three dimensionally in these complexes provides clues towards how they work. In principle, the associations between proteins in large-scale proteomics datasets should often reflect direct physical contacts between proteins in each complex. Here, we describe a statistical method to discover which subunits within complexes directly contact each other based on their co-purification behavior in published co-fractionation mass spectrometry datasets. Within our predictions, we recover many known protein-protein contacts, serving to validate our method, as well as unknown contacts that can inform future studies of these complexes. Specifically, we observe confident contacts between subunits within the exocyst and tRNA multi-synthetase complexes, two complexes that have incomplete structural information. Using our method, we further provide structural information for a previously missing subunit of the EKC/KEOPS complex. We anticipate that this method and the associated predictions will help to better inform our understanding of the functions and structures of diverse protein complexes.

Many proteins assemble into large macromolecular complexes with essential cellular functions. The three dimensional arrangement of proteins in a complex is vital to the complex’s function and knowledge of this arrangement would be highly valuable in understanding the mechanism of function. Conserved protein complexes are estimated to number in the thousands but the vast majority of these are structurally elusive by traditional structural biology techniques. Advances in proteomics technologies have allowed for the high throughput identification of protein complexes across the tree of life including large-scale affinity purification mass spectrometry (AP-MS) datasets [1–3] as well as high-throughput co-fractionation mass spectrometry (CF-MS) datasets comprising thousands of experiments across human, metazoan and prokaryotes [4–7]. In the CF-MS approach, cellular lysate is biochemically fractionated by multiple, non-denaturing chromatographic methods and then complexes are inferred bioinformatically in a machine-learning framework using correlations of the resulting protein elution profiles as a prominent feature. Although this approach has primarily been used to identify component subunits of complexes, we previously observed that the correlation structure of the protein elution profiles also revealed structural information about the complexes [6]. This allowed for the identification of sub-complexes, which were accurate when compared to known structural models and when compared to known functions. However, correlation did not consistently reveal the directly bound protein pairs that other experiments such as yeast two-hybrid [8, 9] and chemical crosslinking [10–14] can reveal across large portions of the proteome. Other computational approaches have been proposed to identify direct contacts by analyzing co-occurrence of proteins in mass spectrometry experiments but they have only been applied to AP-MS datasets [15]. Protein sub-complexes are valuable in understanding the three dimensional arrangement of proteins in a complex but correlation often convolutes specific physical interactions between proteins with indirect interactions and non-physical relationships. Removal of these spurious interactions from the correlation network is crucial to identifying which specific proteins directly contact each other. A classical statistical approach to remove such interactions can be achieved with graphical models [16]. Graphical models represent the conditional dependence structure of a set of random variables as a graph. Unfortunately, classical statistical methods to estimate graphical models fail in scenarios where the number of variables (e.g., proteins) greatly exceeds the number of samples, such as the case with co-fractionation profiles. However, recent advances in the field of statistical analysis, specifically on the topic of sparse high-dimensional statistical inference, have led to new methods for addressing these underdetermined problems (see, e.g. [17] and references therein). In biology, these methods enabled a number of successful applications of graphical modeling, such as estimating interactions between genes from high-throughput expression profiles [18], predicting contacts between amino acid residues from multiple sequence alignments [19], and inferring associations of microbes from environmental sequencing data [20], respectively. Here, we apply a graphical model to identify direct protein interactions (Fig 1) from one of the largest proteomic interaction datasets to date consisting of approx. 3,000 published human CF-MS experiments [6]. We make the assumption that conditional dependence is a proxy for direct protein interactions, which is consistent with the biochemical chromatography methods used in CF-MS experiments due to their separation of native complexes and sub-complexes. We evaluated the performance of our predictions in a precision-recall framework on a benchmark of large protein complexes with known molecular structures and observe substantial improvement over correlation alone. We also observe that the ranking of the learned conditional dependencies is insensitive to particular choices of the regularization parameter λ which balances model complexity and model fit. We additionally characterize our method’s performance finding better predictions for well-observed complexes and validate our predictions with a whole cell lysate crosslinking dataset where we observe enriched overlap. We therefore believe, in principle, these measures of conditional dependence could also be applied to additional proteomic datasets such as AP-MS as well as used in conjunction with other features of direct protein-protein contacts in supervised machine learning frameworks to further improve predictive performance. We highlight predictions made for the 26S proteasome complex and demonstrate agreement with the true set of contacts. We show new predictions for complexes without known structures, specifically the exocyst and tRNA multi-synthetase complex, to illustrate the utility of our approach. Finally, in our predicted set of directly contacting proteins we show support for direct contact of a recently identified component of the human EKC/KEOPS complex. Our results suggest that our predicted direct protein interaction edges will be a valuable constraint that can be used in structurally modeling the thousands of stable protein complexes in the human proteome inaccessible to current structure determination techniques, as we demonstrate with an improved 3D model of the EKC/KEOPS complex. In order to identify direct physical interactions between proteins, we first organized a large, published dataset of human CF-MS experiments [6]. CF-MS experiments consist of two steps, the first being to biochemically separate native protein complexes and sub-complexes along a specified gradient (e.g., hydrodynamic radius, charge, etc.) using non-denaturing separation techniques that preserve intact complexes. The second step is to identify and quantify the proteins that elute at each time point, providing a characteristic elution profile for each protein observed. The aim of our approach is to use these elution profiles to reconstruct the physical interaction network of the proteins identified, and specifically find which proteins directly contact each other within complexes. The dataset comprises d = 15,964 protein elution profiles each consisting of a vector of n = 2,989 protein abundance values. Each protein abundance value is derived from 28 fractionation experiments using multiple, distinct biochemical separation techniques, including ion exchange chromatography, isoelectric focusing and sucrose gradients, analyzing native protein extracts isolated from HeLa cells (17 experiments), HEK293 cells (8), glioma stem cells (2) and neural stem cells (1). Fractionation experiments consist of a series of collected fractions along a biochemical gradient of the applied chromatography method. The number of fractions ranges between ~10 to ~200 per experiment depending on the method. Each fraction of extract is then subjected to proteomic analysis using mass spectrometry producing observed protein abundances. We use the pipeline described in Wan et al. 2015 [6], where the proteomic consensus identification tool, MSBlender [21] is used to identify proteins from mass spectra. For peptide identifications, we use a false discovery rate of < 1%. Missing values that arise when a protein is not identified in a given fraction are set to 0.0. This diverse set of experimental conditions allows for the analysis of a large fraction of the proteome and thorough separation of endogenous complexes. We denote the resulting CF-MS data matrix by X∈R0d×n. Each column Xi, i = 1,…,n represents relative protein abundance data (compositions) and is normalized to sum up to 1. We next introduce a sparse graphical model learning framework to infer direct (physical) protein interactions from CF-MS data from the covariation pattern of the protein abundances. Here, the nodes of the graph represent proteins and the edges approximate direct protein contacts. We first note that components of the compositions Xi are not independent due to the unit sum constraint. Thus, higher order statistics, such as covariance matrices of compositional data exhibit negative bias due to closure. To alleviate this shortcoming we borrow a transformation technique from compositional data analysis [22], the so-called centered log-ratio (CLR) transformation. The CLR transformation is defined as CLR(Xi)=log⁡Xig(Xi), where g(Xi) denotes the geometric mean. This transformation is particularly useful, as it is symmetric and isometric with respect to the original composition. The CLR maps compositional data from the d-dimensional simplex to a (d − 1)-hyperplane in d-dimensional Euclidean space. A pseudo-count of 1 is added to all entries in X to ensure applicability of the transformation. We denote the corresponding covariance matrix by Γ = cov(CLR(X)). Recent work [23] has shown that, in the sparse high-dimensional setting and under certain technical conditions, the covariance matrix Γ is a good estimator for the covariance matrix Σ ∈ ℝd×d of the unknown absolute abundances. This observation is the basis for the proposed graphical model inference framework. Following [20, 24], we propose to learn a sparse undirected graph G∈Rd×d representing node-node interactions via the following minimization problem: G^(λ)=argminG∈Rd×d,Gjj=012tr(G⊺ΓG)−tr(G⊺Γ)+λ‖G‖1 for all j = 1,…,d where tr denotes the trace operator, ‖∙‖1 denotes the element-wise L1 norm, and λ > 0 is the regularization parameter. Each of the d subproblems is equivalent to fitting a linear regression model with L1 penalization (Lasso) [25] to each protein profile, using the other profiles as predictors. To relax any distributional dependencies of the regression, we also apply a non-paranormal (copula) transform to the data before the linear regression step [12]. To symmetrize the graph, derived from the described node-wise regression (or neighborhood selection) algorithm, the OR rule is applied across all node neighborhoods, i.e., an edge in the protein-protein graph is present if either node i is associated with node j or vice versa. It has been shown in [24] that, under certain conditions, the non-zero entries G^ij≠0 of this symmetrized adjacency matrix are asymptotically identical to the non-zero elements Θij of the inverse covariance (or precision) matrix Θ = Σ−1. This allows a clear statistical interpretation of the edges in terms of partial correlation coefficients among the nodes [26]. Thus, the procedure is able to remove transitive correlations among nodes by approximately learning the full conditional dependence among all nodes. One of the key challenges in learning a sparse graphical model from data is the selection of the regularization parameter λ > 0. In the unsupervised setting, several methods have been proposed, including cross validation and information criteria [27, 28]. One state-of-the-art model selection scheme is the Stability Approach to Regularization Selection (StARS) [29]. StARS selects the minimum amount of regularization that results in a graph that is sparse and comprises a stable edge set under random subsampling of the data at a prescribed stability level 1 − β [30, 31]. StARS typically selects N = 20 sub-samples of size b(n)=⌊10n⌋ and learns a graphical model from each subsample across the entire λ-path (here, 30 values of λ are chosen between 0 and λmax). StARS records for each edge in G^(λ) the empirical frequency of edge presence Pij across the entire λ-path, stored in a list of matrices P(λ) ∈ [0,1]d×d. Standard StARS selects λ where the normalized sum of variances of the Pij in the corresponding P(λ) drops below β = 0.1. It has been shown in [31] that this selection can lead to sub-optimal regularization selection. In the present application, we thus opted for an alternative semi-supervised selection procedure. For all positive edges in the interaction graph, we interpreted the edge frequencies as (protein) contact probabilities and ranked edges in order of decreasing contact probability. We compared these ranked predictions to a benchmark of physically interacting proteins determined from multi-protein complexes with known three-dimensional structures and selected the λ that maximized the area under the precision recall (AUPR) curve. The selected λ corresponds to a more conservative StARS variability threshold of β = 0.005. We also note that, in our application, our introduced edge ranking based on the edge stability was insensitive to the precise selection of λ. Finally, we filtered our reported direct contact predictions by protein interactions that are present in 896 complexes larger than 4 subunits from the human protein complex map, hu.MAP [32]. This step was to ensure pairs of proteins are present in the same complex thereby increasing the likelihood of direct contact. All computation was performed in R using the Hotelling package [33] for CLR transformation and the Huge [34] package for graphical modeling. For comparison purposes, correlation analysis was applied to each pair of protein co-elution profiles in the human CF-MS dataset. Profiles were first normalized by the total number of theoretical tryptic peptides for each protein and then a z-score was calculated for each value in the matrix relative to its corresponding fraction (i.e., column-wise standardization). Pearson correlation coefficients were then calculated for each pair of proteins. In order to evaluate the predictive performance of our direct contact prediction method we assembled a benchmark of 29 large non-redundant protein complexes with known structure (S1 Table). Due to the ease at which direct contacts can be predicted at random for small complexes, we restrict our benchmark to complexes having > 4 unique subunits. Note, subunits from certain complexes may not be sampled in our data or have ambiguous ortholog mapping. We process the reported biological assembly of each complex using the PISA tool [35], which calculates macromolecular interface surface area. All pairs of proteins within each complex with interfacial areas (Å2) > 0.0 were considered physically contacting and marked a true contact, comprising benchmark positive examples. Protein pairs with no interface area were considered not contacting, comprising benchmark negative examples. Note that protein pairs that spanned two complexes (e.g., protein 1 in complex 1 and protein 2 in complex 2) were not considered. For complexes whose structure was determined in an organism other than human, InParanoid [36] was used to identify human orthologs of the structurally solved subunit. If no human ortholog could be found for a given subunit, interactions involving that subunit were not considered. We split the benchmark into two sets, the first (10 complexes) to evaluate λ selection and performance and the second (19 complexes) to evaluate generality of the method. The complete protein pair benchmark is provided in S2 Table. We evaluated the overlap of our direct contact predictions with a set of identified inter-protein crosslink interactions from Liu et al. [10]. Similar to the method described in [32] we collapsed all crosslink interactions to one interaction per pair of proteins. We first generate a random overlap distribution by selecting random pairs of proteins from the crosslinking dataset and calculate the overlap with the direct contact predictions for 1000 repeated trials. We then calculate a z-score for the overlap of the direct contact predictions and the reported crosslinking interactions with regards to random distribution. We repeat the process for determining the enrichment of complexes from hu.MAP and the crosslinking interactions. To construct a structural model of the human EKC/KEOPS complex, we built structural models of human EKC/KEOPS proteins based on available template structures in the Protein Data Bank (PDB) [37] and then aligned those models with existing co-complex structures. Specifically, we used HHPred [38] to build alignments of the query protein and PDB sequences and then used MODELLER [39] to build homology models. Homology models of human proteins were then structurally aligned to the homologous structures in yeast and archeal crystal structures [40–42] using DaliLite [43]. Fig 1 shows a workflow of our direct contact prediction framework. Native complexes represented by the true physical interaction network are biochemically fractionated and their proteins identified using mass spectrometry. In order to find pairwise relationships between proteins in a given CF-MS dataset, prior work has relied on correlation analysis, which effectively reconstructs the subunit composition of complexes (especially when used as features in a supervised machine learning framework, a case we do not consider here), but only partially indicates the direct binding relationships among those subunits [4, 6]. More specifically, using correlation to identify pairwise relationships results in a large fraction of indirect interactions. For example, consider proteins A, B and C, where A directly binds B, B directly binds C, but A does not directly bind C. In this scenario, a network based on correlation would produce a spurious edge between proteins A and C due to the indirect relationship mediated by protein B. To address this issue, the inverse covariance matrix can be calculated, which represents a network of undirected edges between conditionally dependent nodes. With respect to CF-MS data, the nodes represent proteins and the conditional dependence edges represent direct physical contacts. The construction of this network has many theoretical solutions due to the limited number of samples and vast number of possible interactions, but methods are available to infer the inverse covariance matrix when the resulting network is expected to be sparse. Sparsity is a safe assumption with respect to protein interactions, as estimates of the total number of expected human protein-protein interactions range between 150k – 650k, orders of magnitude less than the roughly 200–300 million possible interactions [44–46]. As described in detail in the Methods, we analyzed a dataset of approx. 3,000 co-fractionation / mass spectrometry experiments [4, 6], and restrict direct contact predictions to known co-complex interactions. Specifically, we use complexes with structures in the PDB for evaluation and a set of 896 protein complexes larger than 4 unique subunits derived from >9000 published mass spectrometry proteomics experiments [1, 3, 4, 6] in hu.MAP [32], for all other predictions. In all, we identified 2,434 potential interactions (S3 Table). To evaluate whether our direct contact prediction method accurately identifies true interactions, we compared our predictions to a benchmark of physically interacting proteins determined from multi-protein complexes with known three-dimensional structures (S1 Table), as described in the Methods. Fig 2A plots the precision recall curve of our direct contact prediction method relative to the set of 10 complexes used to select λ. We observed high precision for the most confidently predicted contacts. This performance is in contrast to correlation analysis, also plotted in Fig 2A, which has limited accuracy for high correlation coefficients. Plotting precision-recall curves for the 29 alternative λ values considered during λ selection (Fig 2A, gray curves) confirmed that all predictions made with alternative λ values substantially outperformed correlation alone, demonstrating that this parameter was highly stable with regard to its selected value. We further evaluated our direct contact predictions on an additional 19 complexes with known structure (Fig 2B) and observe consistent behavior of our method in terms of precision recall. Interestingly, while correlation performs poorly relative to our method including all λ values on the first set of complexes, it performs substantially better on the second benchmark almost equal to our method. The precision recall curve of the combined benchmark with both direct contact probability and correlation threshold markers can be found in Fig 2C. We next asked if the ability to predict direct contacts was consistent across all complexes or if certain complexes performed better than others. We therefore calculated the area under the precision recall curve (PR AUC) for each individual complex and plotted its distribution in Fig 2D. For our direct contact predictions, we observe a large variance of PR AUC suggesting our method performs well for certain complexes and is limited for others. We still find, however, direct contact predictions outperform correlation analysis and random predictions. To further understand what types of complexes for which our method is appropriate, we investigated how much of an impact the amount of experimental observation affected the degree to which high confident direct contact predictions were made. We first calculated the number of nonzero protein abundance measurements (i.e. count of fractions) for each observed protein and then computed the mean count for every complex in the structure benchmark. Fig 3A shows the distribution of the mean counts for complexes that had at least one prediction with a direct contact probability > = 0.5 and those complexes which did not. We observe a difference in the distributions suggesting that complexes that are well sampled in our dataset are more likely to have high confident predictions. It is important to note that several complexes in our benchmark are not well sampled and our method errs on the side of false negatives so as to limit making false predictions. We additionally plot all direct contact predictions in Fig 3B to better understand the relationship between the direct contact probability score and amount of experimental sampling. We see that pairs of proteins that have high confidence predictions are more likely to have been well sampled suggesting that repeated observations of the proteins across many experiments are important. This trend is likely due to our subsampling scoring procedure which is robust to spurious co-elutions from a single experiment. Fig 4 shows the relationship between correlation and direct contact probability for four examples of complexes in our structural benchmark spanning a range of well observed to poorly observed. Two of the complexes, the proteasome and spliceosome have high confidence predictions made by our method, while the other two, mitochondrial ribosome and mitochondrial super-complex have high-ranking correlation analysis predictions but lack high ranking predictions by our method. The proteasome (pdbid: 4CR2) is well observed with an average nonzero fraction count of ~356. Fig 4A shows our method makes many high confident true positive contact predictions for the proteasome (i.e. top 9/10 are correct) while protein pairs with high correlation coefficient have more of a mix of true positive and false positives. The spliceosome (pdbid: 5MQF, Fig 4B) is moderately observed in the dataset with an average nonzero fraction count of ~182 and still shows good relative discrimination between true and false positive contacts (i.e. top 5/10 are correct). Most of the co-fractionation experiments were focused on identifying soluble cytosolic complexes and therefore membrane bound complexes as well as complexes in subcellular compartments have limited coverage. For example, two mitochondrial complexes, the mitochondrial ribosome (pdbid: 4CE4, Fig 4C) and the mitochondrial super-complex (pdbid: 2YBB, Fig 4D) are identified in a limited number of fractions, on average ~42 and ~105 nonzero fractions respectively. Our method makes very few direct contact predictions for both complexes while correlation has a wide distribution of coefficients, many receiving high scores. Interestingly, high correlation coefficients for the mitochondrial ribosome have a high false positive rate (i.e. top 10 are all false positives) while the mitochondrial super-complex performs better with 7 out of the top 10 pairs being true positives. The poor performance on the mitochondrial ribosome by correlation analysis contributes to the substantial dip in performance seen in the precision recall curves (Fig 2A). These examples further demonstrate the ability of the direct contact prediction method to balance true and false positives and to accurately report contacts when sufficient data is available. As more CF-MS experimental datasets are published, we anticipate an improvement in the coverage of moderately to lowly observed complexes. To assess our direct contact predictions on an independent dataset different from protein structures, we compared to a human cell lysate mass spectrometry crosslinking dataset [10]. The maximum Cα-Cα distance between cross-linked residues for the DSSO cross-linker reagent used is 23.4 Å, making an identified cross-linked subunit pair a reasonable proxy for directly contacting proteins. Since our direct contact predictions are limited to co-complex subunits, we first compare the crosslinking dataset to the set of complexes with which we restricted our predictions. Fig 5 shows that the overlap of complex edges and cross-linked subunits as well as the overlap of our conditionally dependent interactions and cross-linked subunits are both enriched compared to random pairs. Further, we see a much larger enrichment in our conditionally dependent interactions as opposed to complex edges demonstrating the direct contact predictions are highly enriched for physically close and contacting proteins pairs in human cell lysate. We next highlight our method’s ability to identify direct physical contacts among proteins by focusing on a specific protein complex with known structure. The 26S proteasome makes for a clear example of the utility of conditional dependency inference over correlation analysis due to the availability of known three-dimensional structures of this complex [47–50] and the presence of well-defined sub-complexes (e.g., the 20S core and 19S cap). Fig 6A shows the contacts from the known proteasome structure in the upper right portion of the matrix. Interactions are observed amongst the PSMA1 through PSMA7 subunits and PSMB1 through PSMB7 subunits, representing the core, as well as PSMC1 through PSMC6 and PSMD1 through PSMD14 subunits, representing the cap. Notably, not all subunits of the core contact each other, and there are relatively few contacts made between core and cap subunits. These known contacts can be compared with the case shown in the lower left portion of the matrix in Fig 6A, which plots correlation scores from fractionation profiles. While the correlation data exhibit a clear block structure with respect to the core and cap, they do not exhibit the more detailed structure observed in the true contact matrix. The conditionally dependent interactions for these same data are plotted in the lower left portion of the matrix in Fig 6B, representing the method’s estimate of directly contacting subunits. In contrast to the full block structure exhibited by the raw correlations, the direct contact predictions capture finer details of the true contact matrix. Notably, many of the spurious indirect contacts predicted by the correlation matrix are successfully eliminated. For example, the core subunit PSMA6 does not directly contact PSMA1, PSMA7 or PSMB1-5, but does directly contact PSMA2-5 and PSMB6-7. This binding specificity is at least partly captured by the direct contact predictions, but is completely missed by the correlation analysis. Specifically, our method predicts no direct contacts between PSMA6 and PSMA7 or PSMB1-3 subunits, while correlation analysis produces high correlation coefficients for all core subunits. This example exposes the inability of correlation to identify specific direct physical contacts amongst indirect contacts and demonstrates the capacity to remove spurious contacts based on identification of conditional independence. We looked further into cases were we predicted high confidence direct contacts that were labeled as incorrect based on structure data. We noticed an incorrect but high confidence direct contact prediction between two subunits of the spliceosome, SNRPD2 and SNRPD3 (direct contact prob = 1.0). The electron microscopy structure of the spliceosome (pdbid: 5MQF) shows these two subunits within ~17 Å of each other and between the two subunits is an RNA molecule. CF-MS is primarily a proteomics technique and does not observe other molecules such as RNA. We therefore expect to have a degree of error with respect to complexes with structural RNA present, as CF-MS will not show co-elution profiles that discriminate RNA—protein sub-assemblies. We do believe that when these data do become available, the direct contact prediction method is robust enough to identify direct contacts between RNA and protein molecules. Thus, in this case, the high confidence prediction points to a close biological relationship between the two subunits. Additionally, we predict a high confidence direct contact (direct contact prob = 0.95) between two subunits of the eIF3 complex, specifically eIF3e and eIF3h. The C-termini of these subunits participate in an octameric helical bundle at the center of the complex but do not directly contact in the structure used for evaluation (pdbid 5A5T) [51]. In contrast, another structure of eIF3 (pdbid: 3J8B) [52] does have eIF3e and eIF3h directly contacting in the helical bundle. Both structures have limited resolution and are not considered atomic-models suggesting that our data can inform in this discrepancy between models. The prediction of direct contacts gives an opportunity to characterize the structural architecture of complexes that do not yet have a solved structure. The exocyst complex, for example, is a hetero-octamer involved in tethering vesicles to the plasma membrane and is not well understood at the molecular level [53]. Recent studies by Heider et al. [54] and Picco et al. [55] have attempted to resolve the yeast exocyst subunit connectivity map using co-purification and nanometer precision fluorescence microscopy, respectively. Interestingly, Heider and colleagues identified two sub-complexes, sub-complex I consisting of Sec3/EXOC1 (denoting yeast/human orthologs), Sec5/EXOC2, Sec6/EXOC3, Sec8/EXOC4 and sub-complex II consisting of Sec15/EXOC6, Sec10/EXOC5, Exo84/EXOC8 and Exo70/EXOC7. Our direct contact predictions, plotted in Fig 7A, support the presence of these two sub-complexes in addition to identifying inter-sub-complex contacts between EXOC4—EXOC7, EXOC4—EXOC5, and EXOC3—EXOC8. These contacts along with the highly confident direct contact predicted between EXOC3—EXOC4 (also supported by the Heider et al. data) suggests that EXOC3 and EXCO4 form the core subunits of sub-complex I and serve as a bridge to sub-complex II. Likewise, the direct contacts predicted between EXOC5, EXOC7 and EXOC8 suggest they form the core of sub-complex II and are reciprocally responsible for the bridge between sub-complexes. In comparison to the correlation network shown in Fig 7B we observe a much denser network with fewer discriminating edges that help to identify the sub-complexes. We also see a range of correlation coefficients that, empirically, have lower precision then their corresponding direct contact probabilities when evaluated on our combined structural benchmark (Fig 2C). For instance, the interaction EXOC3-EXOC4 has a direct contact probability of 0.85 which is estimated to have a physical contact precision of ~70% while the corresponding correlation coefficient of 0.8 has an empirical precision of ~45%. This example illustrates the ability of our method to predict high confident physical interactions that discriminate from other indirect interactions. A second large complex that has thus far eluded structural characterization is the multi-aminoacyl-tRNA synthetase (also known as MARS) complex, which is composed of 9 synthetases and 3 structural subunits (AIMP1/p43, AIMP2/p38, and EEF1E1/p18/AIMP3) and is estimated to be 1 to 1.5 MDa in size [56]. Individual synthetases within the MARS complex are responsible for covalently attaching specific amino acids to their respective tRNAs and are essential for life. However, the function of the conserved supra-molecular assembly remains unclear. Structural studies, although limited, have identified a few trends in terms of overall architecture of the MARS complex [57], including the presence of two sub-complexes mediated by a core AIMP2/p38 subunit. As shown in Fig 7C, the direct contact predictions clearly establish AIMP2 as central to the architecture of the MARS complex, and strongly link the two larger structural subunits, AIMP1 with AIMP2. Yeast two-hybrid further supports the AIMP1 and AIMP2 interaction as well as the AIMP2 –DARS interaction and AIMP2 –KARS interaction [8]. Additionally, we see strong interactions between the isoleucyl tRNA synthetase IARS and other members of the complex, including the LARS subunit which is supported by mass spectrometry crosslinking data [11]. This suggests that IARS, in addition to AIMP1 and AIMP2, plays a central role in the physical organization of the MARS complex. We further compare the direct contact network to the correlation network for the MARS complex (Fig 7D). Like the correlation network for the exocyst complex described above, the correlation network for the MARS complex is substantially denser, with many more edges of similar coefficients connecting subunits. Interestingly, we find high correlation edges between subunits DARS and MARS, which do not have an edge in the direct contact network (Fig 7C). Since our method attempts to remove spurious conditionally independent edges, this suggests that the correlation coefficient observed between the DARS and MARS subunits can be explained by their mutual interaction with IARS. We see a similar pattern of a high correlation edge absent in the direct contact network including MARS-RARS, MARS-QARS, QARS-RARS, LARS-DARS as well as others. Many of these subunits also interact with the IARS subunit, again suggesting it is the central organizing subunit of the complex. This example demonstrates the utility of direct contact predictions to potentially remove spurious edges from a physical interaction graph. We observed multiple conditionally dependent interactions among a conserved human multi-protein complex with a recently discovered missing subunit. The Endopeptidase-like and Kinase associated to transcribed Chromatin (EKC)/Kinase, Endopeptidase and Other Proteins of small Size (KEOPS) complex is a highly conserved protein complex known to introduce an essential modification to tRNAs across the tree of life [58–60]. The N6-threonylcarbamoyladenosine (t6A) modification is required for normal cell growth and accurate protein translation in bacteria, archaea, and eukaryotes. While the bacterial and lower eukaryotic components of the EKC/KEOPS complex are known, some of the human subunits are substantially diverged and have only recently been discovered [61, 62]. In yeast, the complex consists of five proteins, visualized in Fig 8A: the atypical TP53 receptor kinase/ATPase (Bud32), the Kinase-Associated Endopeptidase (Kae1), and three small proteins, Cgi121, Pcc1, and Gon7 [58, 60]]. Clear orthologs of four of these occur in humans and have previously been confirmed to participate in the EKC/KEOPS complex: TP53RK (the ortholog of Bud32, known to partially complement a Bud32 mutant [63]), TPRKB (the ortholog of Cgi121), LAGE3 (the ortholog of Pcc1), and OSGEP (the ortholog of Kae1). The yeast Gon7 has generally been thought to be fungi-specific [59, 61], and has no clear mammalian ortholog in major ortholog databases [36, 64]. We found that the conditionally dependent interactions (plotted in Fig 8B) strongly supported direct binding of human TP53RK with TPRKB, consistent with expectation from the yeast and archeal crystal structures [40, 41]. Direct binding was also indicated between LAGE3 and OSGEP, again consistent with structural data from archeal homologues [42]. We next observed strong evidence supporting direct binding of OSGEP and LAGE3 with human protein, C14ORF142. Using profile-profile matching, we observe distant but significant homology between C14ORF142 and Gon7 (16% sequence identity and probability score of 92.0), as measured by HHpred [38], which identified Gon7 as the top hit for C14ORF142 from the full non-redundant (reduced to 70% identity) PDB database. This distant sequence similarity strongly supported the observed conditionally dependent protein-protein interactions and suggested that C14ORF142 was indeed likely to substitute for Gon7 within the human complex. Recently, C14ORF142 has been identified as the likely Gon7 ortholog by co-purification with known EKC/KEOPS members [62]. Additionally, the EKC/KEOPS complex was reconstituted in vitro and GST-C14ORF142 was shown to bind directly to the OSGEP-LAGE3 sub-complex validating our prediction. Taking advantage of our predicted direct contacts of C14ORF142 with OSGEP and LAGE3, we constructed a 3D model of the human EKC/KEOPS complex by homology modeling the human proteins onto their yeast orthologs of known 3D structure, including modeling C14ORF142 on the known Gon7 structure (Fig 8C). The resulting 3D model accounts for most of the OSGEP, TP53RK, and TPRKB amino acid sequences, but leaves the C-terminal region of C14ORF142 and the N-terminal region of LAGE3 unmodeled, pointing to additional aspects of this complex still yet to be described. Importantly, the model faithfully recapitulates the known functional and interaction data from the literature, the direct contact predictions from the co-fractionation / mass spectrometry datasets, and the newly recognized C14ORF142/Gon7 structural homology, and thus serves to integrate a large body of data into a single model to help guide future mechanistic studies of this ancient human protein complex. Knowledge of the three dimensional architecture of a protein complex is highly beneficial to understanding its mechanistic function, but thousands of complexes have thus far proved elusive to traditional structural biology techniques. We present an orthogonal approach in determining aspects of the three dimensional architecture of complexes by analyzing large scale CF-MS datasets. Using our method, we predicted thousands of direct contacts between complex subunits. We expect this resource can be used as a valuable constraint for structurally modeling the many stable protein complexes in the human proteome using available modeling tools [65, 66]. The method should easily extend to new organisms as additional large-scale CF-MS datasets become available. Code and input elution profiles file can be found at https://github.com/marcottelab/direct_contact.

10.1371/journal.pcbi.1000131

Modeling the Violation of Reward Maximization and Invariance in Reinforcement Schedules

It is often assumed that animals and people adjust their behavior to maximize reward acquisition. In visually cued reinforcement schedules, monkeys make errors in trials that are not immediately rewarded, despite having to repeat error trials. Here we show that error rates are typically smaller in trials equally distant from reward but belonging to longer schedules (referred to as “schedule length effect”). This violates the principles of reward maximization and invariance and cannot be predicted by the standard methods of Reinforcement Learning, such as the method of temporal differences. We develop a heuristic model that accounts for all of the properties of the behavior in the reinforcement schedule task but whose predictions are not different from those of the standard temporal difference model in choice tasks. In the modification of temporal difference learning introduced here, the effect of schedule length emerges spontaneously from the sensitivity to the immediately preceding trial. We also introduce a policy for general Markov Decision Processes, where the decision made at each node is conditioned on the motivation to perform an instrumental action, and show that the application of our model to the reinforcement schedule task and the choice task are special cases of this general theoretical framework. Within this framework, Reinforcement Learning can approach contextual learning with the mixture of empirical findings and principled assumptions that seem to coexist in the best descriptions of animal behavior. As examples, we discuss two phenomena observed in humans that often derive from the violation of the principle of invariance: “framing,” wherein equivalent options are treated differently depending on the context in which they are presented, and the “sunk cost” effect, the greater tendency to continue an endeavor once an investment in money, effort, or time has been made. The schedule length effect might be a manifestation of these phenomena in monkeys.

Theories of rational behavior are built on a number of principles, including the assumption that subjects adjust their behavior to maximize their long-term returns and that they should work equally hard to obtain a reward in situations where the effort to obtain reward is the same (called the invariance principle). Humans, however, are sensitive to the manner in which equivalent choices are presented, or “framed,” and often have a greater tendency to continue an endeavor once an investment in money, effort, or time has been made, a phenomenon known as “sunk cost” effect. In a similar manner, when monkeys must perform different numbers of trials to obtain a reward, they work harder as the number of trials already performed increases, even though both the work remaining and the forthcoming reward are the same in all situations. Methods from the theory of Reinforcement Learning, which usually provide learning strategies aimed at maximizing returns, cannot model this violation of invariance. Here we generalize a prominent method of Reinforcement Learning so as to explain the violation of invariance, without losing the ability to model behaviors explained by standard Reinforcement Learning models. This generalization extends our understanding of how animals and humans learn and behave.

In studying reward-seeking behavior it is often assumed that animals attempt to maximize long term returns. This postulate often forms the basis of normative models of decision making [1], choice behavior [2]–[4], and motivation [5], and plays a prominent role in the field of Reinforcement Learning (RL; see, e.g., [6]). RL is a set of methods for learning to predict rewarding outcomes from their association with environmental cues, and to exploit these predictions to generate effective behavioral policies. These are policies that comply with principles of reward maximization [7],[8] and invariance [9],[10]. Applied to reward-seeking behavior, the principle of reward maximization states that subjects should maximize the reward/cost ratio, and the invariance principle that subjects should be equally motivated when facing situations with identical reward/cost ratios. The idea of maximizing reward over time or effort is general and has provided an effective basis for describing decision-making where the choice between available options is basically a matter of preference. RL methods such as the method of temporal differences (TD) constitute an efficient way of solving decision problems in tasks where a subject must choose between a larger vs. a smaller reward, or between a more probable vs. a less probable reward, and predict courses of actions comparable to the actual behavior observed in animals performing the same tasks [11]–[13]. RL methods have proven less successful, however, in situations where motivation, defined as the incentive to be engaged in a task at all, plays a strong role [14]–[16]. A case in point is the behavior of monkeys performing visually-cued reinforcement schedules [17], wherein a series of identical actions is required to obtain reward, and a visual cue indicates how many trials remain to be completed before a reward is delivered (“reward schedule task,” see Figure 1). In this task, the error rate of the monkeys is proportional to the number of unrewarded trials remaining before reward, indicating that the value of the trial is modified by knowing the number of remaining trials. This violates the principle of reward-maximization: monkeys make errors in unrewarded trials that will have to be repeated, thus preventing optimal reward-harvesting behavior. Here we show that in trials equally far from reward, monkeys make fewer errors in longer schedules, when more trials have already been performed (“schedule length effect”). Thus, the value of the current trial is also modified by the number of trials already completed. This behavior violates the principle of invariance: monkeys perform differently in trials equally far from reward, depending on the number of trials already completed in the current schedule. Taken together, these results suggest that the behavior in the reward schedule task does not develop under the principles of invariance and reward-optimization, as commonly assumed when applying RL methods to understanding reward-seeking behavior. We present a RL rule which predicts the monkeys' behavior in the reward schedule task. Such a rule is a heuristic generalization of TD learning. When applied to the reward schedule task, it predicts all aspects of monkeys' behavior, including the sensitivity to the contextual effect due to schedule length leading to the violation of the invariance principle. When applied to a task involving choice preference, the new method predicts the same behavior as does the standard TD model. Thus, the behaviors in the reward schedule and in choice tasks can be the consequence of the same learning rule. Building on the special cases of the reward schedule and choice tasks, we then provide a general theory for Markov Decision Processes, wherein the transition to the next state is governed in a manner similar to a choice task, but is conditioned on whether the agent is sufficiently motivated to act at all, like in the reward schedule task. Finally, we link the schedule length effect to instances of “framing” [18],[19] and “sunk cost” effects [20],[21], which also emerge in conjunction with the violation of the principle of invariance. In this work we collate the behavior of 24 monkeys tested in the reward schedule task [17], and analyze the entire set of data as a group (see Material and Methods). In this task, a series of trials had to be completed successfully to obtain reward at the end of the series. This series is defined to be a schedule, which is then characterized by its length measured in number of trials (Materials and Methods and Figure 1). The monkey starts each trial by holding a bar which causes a visual cue to appear on a computer screen, followed by the appearance of a red dot in the middle of the screen. The monkey must wait for the red dot to turn green (“GO” signal), at which point it must release the bar within a 200–1000 ms window. If the bar is released correctly, the monkey proceeds to the next trial of the schedule. Each trial must be repeated until performed correctly. In the presence of visual cues informing the monkey of the progress through the schedule (Valid Cue condition), the percentage of errors in all monkeys was directly related to the number of trials remaining to be completed in the schedule, i.e., the largest number of errors occurred in the trials that are furthest from the reward (χ2 test, p<0.05; Figure 2A and 2B, circles; each trial is labeled by the fraction τ/s, where τ stands for current trial and s stands for current schedule length). The performance in terminal trials was indistinguishable across schedules for each monkey, was above 94% correct in 14 out of 24 monkeys, and above 90% in 19 out of 24 monkeys. In the Random Cue condition the visual cues were selected at random and bore no relationship to schedule state. In such a condition, error rates were indistinguishable across all schedule states (or idiosyncratic; “x” in Figure 2A and 2B; χ2 test, p>0.05 in 10 out of 15 monkeys tested in the Random Cue condition), and close to the error rates in terminal trials in the Valid Cue condition. Thus, performance in unrewarded trials in the Valid Cue condition was well below the ability of the monkeys. Since the individual trials of each schedule have the same perceptual and motor demands, we interpret the different error rates as being related to the different levels of motivation. This interpretation is also supported by the observation that, in most monkeys, the reaction times become faster as the end of the schedule is approached [17], [22]–[25]. In the penultimate trials of each schedule (i.e., 1/2, 2/3, and 3/4 when available) 20 of 24 monkeys made progressively fewer errors as the schedule became longer (sign test, p<0.005). The error rate in state 1/2 was significantly larger than in state 2/3 in 12 out of 20 monkeys (Marascuilo procedure, p<0.05, see Materials and Methods and Figure 2C and 2D). In two of three monkeys tested with 4 schedules, the error rate in state 2/3 was also significantly larger than in state 3/4. The third monkey tested with 4 schedules showed small error rates, and multiple comparisons between penultimate trials were not significant (monkeys often will not perform the task with 4 schedules [17]). In many of these studies the cues were distinguished by their brightness, where their brightness had been set according to the number of trials remaining in the schedule (Material and Methods), raising the possibility that performance was related to judging the brightness. However, this seems unlikely because the behavioral sensitivity was also seen when unique stimuli, e.g., Walsh patterns, were used as cues (e.g., Figure 2 of [26]), where no feature of the visual stimulus is a graded function of reward proximity or progress through the schedule. In conclusion, in a population of monkeys there was a significant tendency for motivation to increase with the number of trials already performed, at parity of proximity to reward. We refer this phenomenon to as the “schedule length effect.” In the reward schedule task, all trials have the same cost because they all require the same action in response to the same trigger (the appearance of the green dot); trials differ only in their proximity to reward, which in turn does not depend on how many trials have already been performed. A standard reinforcement learning method can only learn to predict the proximity to reward correctly, and thus, unlike the behavior shown by the monkeys, is insensitive to the context introduced by the schedule length. We address this issue in detail in the remainder of this manuscript. In reward schedule tasks, monkeys make substantially more errors in validly cued unrewarded trials than in rewarded trials. The number of errors decreases with reward proximity. Also, the error rates are typically smaller in trials equally distant from reward, but belonging to longer schedules (schedule length effect; Figure 2). Both of these features disappear and monkeys make fewer errors in the absence of valid cues. The monkeys do not maximize the amount of reward over the smallest number of trials, violating a principle requiring maximization of reward over time, and also violate the principle of invariance in trials equally far from reward, especially penultimate trials (Figure 2). This behavioral pattern occurs in most monkeys, thus it is robust and reliable. It only occurs after the meanings of the cues are learned, and persists over long periods (months or years despite constant practice in the task). Therefore, it should not be construed as maladaptive simply because it violates the principles of reward-maximization and invariance. Since the monkeys were allowed to work until they stopped by themselves, it can be inferred that they would get a sufficient amount of liquid reward, and were simply not interested in maximizing the amount over time. We have argued that the monkeys' behavior is a direct consequence of learning the motivational values attached to each trial by using the cues. Either randomizing the cues or damaging the rhinal cortex prevents the formation of this typical error rates pattern [26],[32], and damaging orbitofrontal cortex blunts it [33]. In the model introduced here, the motivational values of the schedule states arise through trial-and-error learning and lead to suboptimal behavior. In its basic form, i.e., with σ = 0, this model can be described as TD-learning for solving the value prediction problem [6],[27]. The standard RL approach is usually concerned with the development of behavioral strategies that adapt towards optimality, and less often with the simpler value prediction problem, i.e., the problem of learning to predict the long term return obtainable starting from each behavioral state and following a given policy. Our interpretation of the RL method used in this work follows this thread, because the policy (the performance function Equation 1) is fixed and is not modified by the learning algorithm (of course its arguments, the values, are). This has been called “learning with an indirect actor” by Dayan and Abbott [34]. The particular policy used for the reward schedule departs from previous accounts because it depends on the value of the current state only. This is one of two core departures of our model from existing ones (e.g., [35],[36]). The second is the modification of the learning rule so as to capture the schedule length effect. In our model, a single algorithm explains the differential behavior with valid and random cues. Assuming that the average value of the schedule states is a measure of overall motivation, the model predicts that the overall motivation is similar in the valid and random cue conditions. The difference in performance in the two paradigms is a consequence of the non-linear (sigmoidal) shape of the performance function Equation 1 (cf. Figure 4C). The finding that the same overall level of motivation leads to different patterns of error rates with valid and random cues is not built into the model but is an emergent property of the learning process. The context-sensitive model also predicts that, although the behavior appears to be the same in all terminal trials, terminal trials may acquire different values (see, e.g., Equation 9). This difference is not reflected in the behavior since the latter depends on both the values (which might be different) and the performance function (Equation 1), which tends to remove value differences in the high value region (Figure 4C). In this region the performance function (or its complement) is almost flat and slight differences in value will be unlikely to produce observable differences in error rate. The context-sensitive behavior is also an emergent property of the model. The model does not change the definition of the schedule states to accommodate their contextual meaning. Valid cues come to “label” the schedule states via predictive learning. The basic model translates these labels into a pattern of motivational values and error rates which only depend on reward proximity, and thus are the same in penultimate trials. This symmetry is broken in the context-sensitive model as a consequence of generalizing the temporal difference so as to look backwards as well as forward, and not through a redefinition of the schedule states. It might seem at first that the model does not take into account the cost of performing a trial, i.e., the cost of releasing the bar at the GO signal. In fact, this cost could be interpreted as the origin of the residual, non-zero error rate given by the performance function (Equation 1) when the values are maximal (approximately, the error rate in validly-cued rewarded trials). It is also possible to implement this cost so as to affect the values of each state, V(S)→V(S)−c, where c stands for cost. However, since the cost of the action is the same for all trials, it could not account for the differential error rates in different schedule states. Our analysis unveils the inadequacy of standard TD learning for the reward schedule task. The general statement can be proved that it is not possible to capture the schedule length effect with RL methods inspired to TD learning, including TD(λ) [27], if these only take into account the values of trials remaining in the current schedule (cfr. Equation 4; see Materials and Methods for details). Thus, for a method based on temporal differences to capture the schedule length effect, its learning prescription must have access either to the value of a past trial in the current schedule, as proposed in this manuscript, or to the value of a trial belonging to a different schedule, a method that is not clear how to generalize beyond the reward schedule task. The predictions of the context-sensitive model are the same as standard TD learning in a wide class of other tasks involving choice, where the values of states at decision nodes apply equally to whatever outcome of the decision. In simple choice tasks (cf. Figure. 5), both models predict a preference for more probable rewards, either always—under a greedy policy—or with occasional, temporary reversals of preference when the policy allows exploratory behavior—like the softmax function Equation 11. In the choice-schedule task of Figure 6A, the context-sensitive model predicts the same preference as the standard model. With schedules comprising more than two trials, choice preference of one model can be mapped into the choice preference of the other by readjusting the value of the discount rate γ appropriately. Thus the context-sensitive model, although heuristic in its derivation, appears to be a generalization of standard TD learning: it predicts the same behavior in tasks where human and animal subjects tprefer the choice leading to more probable or larger rewards; but it also predicts the violation of the principle of invariance occurring in the reward schedule task, not captured by the standard model; and it predicts the “procrastination-like” behavior of monkeys in the same task. The latter is to be generally expected in tasks requiring a step-wise approach to reward, where the willingness to act in each single trial exerts a powerful influence on the behavior. More work is required to characterize fully the mathematical properties of the model, and explore its possible derivation from well-defined principles as is customary in the fields of Machine Learning and RL, which is beyond the scope of this work. The reward schedule and choice tasks represent two particular cases of general MDPs where the problem of making a decision can be factorized into two sub-problems, the motivation to perform at all, and the selection of one among alternative choices given the motivation to act. We have used the strategy of dividing this general problem into two parts: we have analyzed the behavior as driven by motivational value using the reward schedule task, and the behavior as driven by choice preference using choice tasks. In both cases, we have compared the standard and the novel TD model using the same policy for both. These two components are simply multiplied in general MDPs, where by definition both the motivation to act and choice selection can occur. Our results indicate that only in the choice selection problem does the actor-critic architecture of RL [6],[37] potentially have a significant role. In the actor-critic architecture, the RL problem is solved by two related “structures,” one responsible for performing the action (the actor), the other responsible for criticizing those actions based on evaluative feedback (the critic). Actor-critic architectures usually lead to policies that maximize the long-term return, and thus they seem to have only a small role in the reward schedule task. If an underlying actor-critic is present, its effectiveness in producing optimal control might be blunted by an opposing force deriving from the purely motivational nature of the problem encountered in this task, i.e., whether or not to comply with its demands. Indeed, we have shown that it is sufficient for the critic to assess the value of the current trial and use it to direct the level of engagement in the task, without the need for a more specialized actor structure as would be required for action selection [38]. Instead, the process of valuation of several alternatives, potentially leading to different courses of actions and rewards as it typically occurs in general decision problems, could benefit more from an actor-critic organization of the behavior. The extension of RL to capture the fundamental role of motivation in reinforcement schedules is currently a major challenge for the field, and other authors have also considered how to include motivation in RL [14],[15]. These authors focused on incorporating overall drive (e.g., such as degree of hunger or thirst) so as to describe how habitual responses can be modified by the current motivational level, which is, in turn, assumed to influence generalized drive through sensitivity to average reward levels [39]. In the reward schedule, however, we focused on how motivation orients behavior in a trial-specific, not generalized, manner. In such a case, an alternative solution to ascribing errors to a decreased level of motivation is Pavlovian-instrumental competition, which has been used to explain suboptimal behavior [16]. Applied to the reward schedule task, this solution would posit that error trials would result from the competition between the negative valence of the valid cue associated to an unrewarded trial (acquired through a Pavlovian-like mechanism), and the incentive to perform the same trial correctly to reach the end of the schedule and obtain reward. This interpretation is supported somewhat by the fact that the visual cues have no instrumental role in the reward schedule (they are neither triggers nor instructors of correct behavioral actions). The schedule length effect, however, escapes explanations in terms of Pavlovian-instrumental competition and would still have to be taken into account. Instead, the single motivational mechanism put forward in this work accounts for all the aspects of the behavior; has a natural interpretation in terms of learned motivation to act, however originated; and can be extended to general MDPs. A dependence on the value of the preceding state implemented in our learning rule suggests an explanation of the schedule length effect as a history effect. When environmental cues are not perfect predictors of the availability of resources, monkeys' decisions about where to forage depend on past information like the history of preceding reinforcements [40], or stored information about recent trends in weather [41]. Lau and Glimcher [28] have found that past choices, in addition to past reinforcements, must be taken into account to predict the trial-by-trial behavior of rhesus monkeys engaged in a choice task resulting in matching behavior. However, contrary to the statistical description of Lau and Glimcher [28], past information in our model bears an effect on the learning rule, not directly on the action selection process, and it does so through the value of the previous state, as opposed to past reinforcements or past choice history. Taken together, these findings point to some form of sensitivity to preceding actions and visited states (or their values) in primates' foraging behavior, and the schedule length effect might be a side effect of such a mechanism, perhaps also present in other forms of reinforcement learning. Current theories of reinforcement learning posit that dopaminergic neurons code for a prediction error signal analogous to δ in our model and in TD learning in general [30]. Data from dopaminergic neurons of monkeys performing a reward schedule task, however, are not in sufficient accord with the predictions of such theories [24]. For example, one prediction is that δ, and therefore dopamine neurons, after sufficient training should cease to respond to predicted reward, and this was not observed. Recent developments [42],[43] rule out that this could be the consequence of the small temporal jitter around reward delivery. Despite the incongruence with the assumed role of dopamine neurons as signaling some form of prediction error, there is clear evidence of the involvement of dopamine in learning. In the reward schedule, the importance of dopamine D2 receptors for learning the meaning of new valid cues has been demonstrated in perirhinal cortex [26], and Ravel and Richmond [24] have argued that salient events may drive dopaminergic neurons, whose activity may be required for enhancing the connection of the stimulus with its prediction in perirhinal cortex. The contextual impact of the organization of the task in schedules has been found in the event-related responses of neurons in all neural structures investigated thus far in the reward schedule task, except perhaps for neurons of the area TE [22]. The brain area where the neural modulation with schedule state is most apparent is the anterior cingulate cortex [44]. One third of the neurons recorded in this area keep track of the progress through the schedule in the Valid Cue condition, and could reflect the (motivational) value of the schedule states and their being linked to one another in a chain of states culminating in the rewarded trial. Another candidate structure for the representation of the schedule states is the perirhinal cortex, whose neurons become selective for the meaning of the visual cues, as opposed to, e.g., TE neurons' responses that are locked to their physical identity [22]. In some brain regions, neuronal responses are different in trials of different schedules that might be regarded as homologous, particularly last trials of different schedules. Dopamine neurons [24], perirhinal neurons [22] and ventral striatum neurons [23] respond differently to valid cues in last trials (predicting the same reward, but in different schedules). This is reminiscent of the phenomenon that the context-sensitive model assigns different values to terminal trials belonging to different schedules. Neurons of the basolateral complex of the amygdala often have differential post-cue activity in first trials [25]. In these neurons another, different effect related to the organization in schedules has also been observed: these neurons increase their activity in the pre-cue period before the beginning of each schedule. No pre-cue activity was observed in the Random Cue condition, supporting the hypothesis that pre-cue activity is related to the contextual imprint of the task's organization in schedules [25]. This activity could be related to a context-sensitive representation of the values of the states, either in the amygdala itself, or in areas connected to the amygdala like perirhinal cortex [22], anterior cingulate cortex [44] and ventral striatum [23], where the schedule state meaning of valid cues is strongly represented. The possibility of a more specific role of the amygdala for the emergence of the schedule length effect will be considered later when discussing the analogous phenomenon of “framing” in humans. Finally, there is evidence for the role of the primate striatum in learned action selection, with some authors [45] proposing for its ventral part coding for the values of states (reminiscent of the critic in actor-critic RL methods), and its dorsal part coding for the values of actions and for action selection (reminiscent of the actor [11],[13],[45]; but see [38]). In the reward schedule, the largest population of ventral striatum neurons which are responsive around the time of bar release, do so in rewarded trials, with the second larger population being responsive in all trials [23]. Comparison of latency and periods of peak activity between these neurons and neurons of the orbitofrontal cortex suggest that the latter are better positioned for representing the reward contingency and thus for guiding action, whereas the former are more related to executing the action [46]. This role is usually ascribed to more dorsal regions of the striatum, but the involvement of the ventral striatum is conceivable in the reward schedule, given the simple action selection required (it amounts to the timely execution of the bar release in all contingencies), and it is compatible with our model, where the probability of a correct bar release is based solely on the value of the current state and not on action values. In the context-sensitive model, the mechanism responsible for the schedule length effect leads to the violation of invariance. The violation of this principle was invoked by Tversky and Kahneman in their description of “framing” [18],[19]. Framing describes the process whereby the choice made is influenced by the manner or context in which the choice is presented. Thaler [47] and Tversky and Kahneman [18] showed that humans often act as if they kept separate accounts for gains and losses, rather than estimate the total value. A consequence of keeping separate accounts is that the manner in which a problem is cast, in terms of gains, of losses, or of total value influences choices. For example, people purchasing two items, costing respectively $15 and $125, are more willing to put an effort (for example by driving to another store) to save $5 when this is presented as a discount on the $15 item, than when presented as a discount on the $125 item, even though the total saving is the same [18],[47],[48]. Similarly, monkeys are willing to put more effort in a trial if the total effort to get there had been larger, even though this does not affect the upcoming reward. A “minimal account” would consider only the proximity to reward, whereas the behavior of the monkeys shows that a combination of minimal (reward proximity) and topical (workload) accounts affects their motivation when facing a reward schedule. From this point of view, reward proximity could be seen as a property defining the state (in accord with Equation 7), much like the $5 discount defines the saving in the example above, independently of the item to which it is nominally attached. In both cases, it is the comparison with some truly contextual attribute that assigns a different motivational value to the same action. Thus, especially on penultimate trials, the length of the schedule seems to exert a contextual effect on the monkeys' motivation analogous to framing. A more direct, preliminary example of framing in monkeys has been reported recently [49] using a task similar to one previously used with starlings [50]. The schedule length effect is also reminiscent of the so-called “sunk cost” effect [20],[21],[51],[52], “a maladaptive behavior that is manifested in a greater tendency to continue an endeavor once an investment in money, effort or time has been made” [21]. The sunk cost phenomenon comes in different varieties and with different interpretations (to the point of having different names, like “Concorde effect,” “cognitive dissonance,” “work ethics,” see [20] for a review), some of which come close to framing. In one interpretation, sunk cost derives from the violation of the principle that “a prior investment should not influence one's consideration of current options; only the incremental costs and benefits of the current options should influence one's decision” [20]. The similarity with the schedule length effect and with the previous discussion about its interpretation in terms of framing seems obvious. A relevant example is Experiment 2 of Arkes and Blumer [21]. In this experiment, three groups of patrons were sold season tickets for the Ohio University Theater at three different prices, and those who purchased tickets at either of the discounted prices attended fewer plays during the season. In this case, the money spent at the beginning of the season influenced the patrons' choice to attend the plays. It could be argued that, in the reward schedule task, the cost of performing trials is not strictly a “sunk” (wasted) cost, as it would be if the monkeys had to start the schedule anew after each error trial. However, this would only be a minor difference with other instantiations of sunk cost effects; and it could similarly be argued that the money spent in Experiment 2 of Arkes and Blumer [21] is not a wasted cost, since it is necessary to attend the plays. Various explanations of sunk cost and framing have been proposed. Arkes and Ayton [20] explain the sunk cost fallacy as an overgeneralization of the “don't waste” rule, since based on their review of the literature, the effect is not unambiguously present in lower animals, and is not found in children [20]. Even if the schedule length effect can legitimately be interpreted in terms of sunk cost or framing, we think that this is unlikely to be the correct explanation. A better explanation may be linked to emotional factors. A functional imaging study [53] points to an important emotional component in the susceptibility to frames in humans. This study found the susceptibility to framing linked to amygdala activations, with the ability to resist the frame linked to activation of the orbital and medial frontal cortex. Similarly, we believe that there is a strong emotional component responsible for the monkey's reaction to unrewarded cues (leading to larger error rates), and possibly for the schedule length effect. Thus, a connection between this emotional component and parameter σ, which quantifies the schedule length effect in our model, could be speculated on the basis that a larger σ implies a larger schedule length effect, in the same way as a larger emotional component would imply a stronger susceptibility to framing [53]. We do not reject this idea as a possibility, but our data are not sufficient evidence for it. Our model does make a clear prediction in one case where framing has been found, i.e., in the increase in preference due to training with a larger cost [51], a case of state-dependent learned valuation. In this experiment, starlings preferred to choose stimuli which had previously associated with a larger effort (16 1-m flights vs. four 1-m flights) to obtain an otherwise identical reward. Since this paradigm pitted two reinforcement schedules of different length against each other, there are obvious similarities with our reward schedule task. Indeed, it would be possible to run a similar test in monkeys by associating different cues to terminal trials in different schedules (e.g., cue H for the longer schedule and cue L for the shorter), and then test the monkeys' preference in a choice task where there is no cost (or equal cost) to obtain the same reward from two sources, one cued with H, the other with L. Would the monkey prefer the cue associated during training with the longer schedule, as found in starlings [51]? Our model predicts exactly this. Because of the accumulation of previous values, the values of terminal trials are larger in longer schedules in the model with σ>0. Assuming that in the choice task preference depends on the same learned values, the source of reward cued by H (previously associated with the longer schedule) would be preferred. This also means that our model implies state-dependent learned valuation when the state of the animal is defined by the cumulative effort expended to obtain the reward. We stress, however, that our learning model is not meant to be a general model of the effects that frames, or sunk costs, have on humans and animals. For example, Pompilio et al [52] offer additional evidence of state-dependent valuation in an invertebrate (the grasshopper), but in their case the state of the animals at the time of learning is defined by their nutritional state (e.g., more or less hungry) as opposed to their expended cost. They found that the grasshoppers, in a later choice task with equal cost, prefer the food experienced when in a lower nutritional state during learning. We do not see a connection between this finding and the schedule length effect, or the role of the parameter σ. This should not be surprising. As Pompilio et al. [52] point out, there may be more than a single mechanism responsible for state-dependent valuation, depending on the animal and, in the same animal, depending on the paradigm used for training. In the heuristic modification of TD learning introduced in this work, the schedule length effect emerges spontaneously from the sensitivity to the immediately preceding trial, leading to the violation of the invariance principle. Since this principle is violated in instances of framing and sunk cost effects, we have interpreted the monkeys' behavior using the framing and sunk cost analogies, even though monkeys might not be susceptible to framing or sunk cost the way humans are. We are not aware of alternative RL models predicting the violation of the principle of invariance. In this work we collate the behavioral data from earlier studies on monkeys (Macaca mulatta) tested in the reward schedule task [22]–[26],[32],[44]. In all of these studies, randomly interleaved schedules of one, two or three trials must be completed to obtain a reward. In n = 3 monkeys, schedules with 4 trials were also used. A trial begins when the monkey touches a bar (Figure 1A), causing the appearance of a visual cue. Four hundred milliseconds later a red dot (WAIT signal) appears in the center of the cue. After a random interval of 500–1500 ms the dot turns green (GO signal). The monkey is required to release the touch-bar between 200 and 800 ms after the green dot appeared, in which case the dot turns blue (OK signal), and a drop of liquid reward is delivered 250 to 350 ms later. If the monkey releases the bar outside the 200–800 ms interval after the GO signal, an error is registered, and no reward is delivered. To start, monkeys are trained on this simple color discrimination task, with or without the presence of a cue, and are rewarded for every correct trial. When performance reaches criterion (at least 75% correct), reward schedules start. Each reward schedule is a sequence of 1, or 2, or 3, …, or Ns trials, where Ns is the maximal schedule length for that session (3 or 4; see Figure 1B for a 2-trial schedule). All schedules are selected with equal probability, and within a schedule error trials must be repeated until performed correctly. Only correct terminal trials are rewarded. After a correct terminal trial, a new schedule is selected pseudo-randomly. Each schedule state is labeled by the pair {τ,s}, where τ = 1, 2, …, s stands for trial and s = 1, 2, …, Ns stands for schedule. Terminal trials have τ = s. Trials of different schedules representing the same schedule fraction (e.g., 1/2 and 2/4) are considered different schedule states, even though they might have been associated to the same visual cue (Valid Cue condition, see below). Different cue sets have been used in different studies [17], [22]–[26],[32],[44],[54], producing similar behavioral results. For the data shown in Figure 2, collected by Sugase-Miyamoto and Richmond [25] (panel A) and Shidara and Richmond [44] (panel B), horizontal bars with different brightness were used as cues, and the cues were brighter as the schedule progressed. Other cue sets have also been used. Some, still based on cue brightness, had the opposite relationship between brightness and proximity to reward, e.g., cues were darker towards the end of the schedule, as, e.g., in Figure 1 [22],[24],[32],[54], to ensure that the behavior of the monkeys was not biased by the direction of brightness. Other cue sets were based on bar length [26],[32]; still others consisted of unique stimuli like, e.g., Walsh patterns [26], to establish that the behavior was not a consequence of having a sensory attribute (like length or brightness) increasing or decreasing with proximity to reward. The typical behavioral patterns that are the main focus of this work were similar across individual experiments and cue sets. In the paradigm with random cues, the same visual stimuli are present, but each stimulus is selected pseudo-randomly with equal probability in each trial (Random Cue condition). In such a case, there is no relationship between cues and schedule states, although the schedules are still in effect. The monkeys were not taught the “rules” of the reward schedule task but were simply exposed to it. The behavior reported in Figure 2 emerges spontaneously, typically within a week of the first exposure, depending on the monkey (in some cases, it emerges on the very first day), and it generalizes rapidly (in less than 3 days) to different cue sets. For each monkey, the error rates were calculated as the ratio of the total number of incorrect trials (in all sessions) to the total number of trials for each schedule state. Differences in error rates across schedule states were tested with a χ2 test of the contingency table obtained from the numbers of correct and incorrect trials (confidence was taken at the 5% level). Pair-wise comparisons of the error rates in different schedule states were tested with the Marascuilo procedure after a significant χ2 test [55]. If the χ2 test is significant at the α level, the Marascuilo procedure [55],[56] provides a confidence interval of 100(1−α)% for each pair-wise difference of error rates |pi−pj|, where pi = ei/ni is the error rate in schedule state i, and ei, ni are, respectively, the number of error trials and total trials in schedule state i. The Marasquilo confidence interval on |pi−pj| is given by . In this formula, is the critical value of χ2 with N−1 degrees of freedom at α level of significance (the point of the distribution which leaves an area of α in the upper tail of the distribution). N is the number of different schedule states. Schedule states with |pi−pj|>pˆij are significantly different at the α level. A sign test [57] was run on the number (n+) of monkeys showing better performance in penultimate trials belonging to longer schedules, as compared to the number (n−) of monkeys where either the inverted pattern, or no difference, was observed. The “exact” binomial probability for n+ successes in n++n− trials was used. Reaction times were defined as the time elapsed since the appearance of the GO signal and the bar release, and, as reported previously, were generally shorter in trials more proximal to reward [17], [22]–[25]. Reaction times had a similar relationship to schedule states as did error rates. Since they provide no new qualitative interpretation, they were not analyzed further. For each monkey, the theoretical error rates (pth) were fitted to the experimental error rates (pex) by minimizing a weighted sum of squares, , where the sum goes over all schedule states in both the Valid and Random Cue conditions, and [58]. The reason for this choice is that the interval is approximately a 68% confidence interval around pex,i based on Wilson's “score” equation [59],[60], and (Li,+−Li,−)/2 = Δpi. The theoretical error rates were given by Equations 2, 9, and 10. The minimization of χ2 was accomplished with a full factorial search of the best-fit values for parameters β, γ, and σ of Equations 2, 9, and 10. The formula Equation 6 of the main text for the equilibrium values of the basic model is exact only in the absence of errors, otherwise the values are smaller and are given by the self-consistent, recursion formula:(13)Here, S′ is the next state in the schedule, Pc|V(S)≡P(c|V(S)) is the probability of correct performance in (current) state S, conditioned on the value of that state, V(S). V(S) appears also on the left hand side, and for this reason the formula defines V(S) only implicitly. If S is a terminal trial, γV(S′) must be replaced by r in Equation 13. By iteration, Equation 13 gives(14)where to simplify the notation. This set of equations must be solved self-consistently for Vτs as the Pτs depend on Vτs. Under the optimal policy of not making any errors, i.e., with each Pc|V≡1 independently of V instead of Equation 1, Equation 14 becomes the explicit solution given by Equation 6 reported in the main text. Equation 13 can be derived as follows: at equilibrium, V(S) is the average of the value obtained after an error (Ve, occurring with probability Pe|V≡1−Pc|V) and the value obtained after a correct trial (Vc, probability Pc|V), conditioned on current average value being V, i.e.(15)with Vc(V) = V+α(γV′−V) and Ve = V+α(γV−V). The last two equations are simply the update equation for V after a correct and an incorrect trial respectively; V′≡V(S′) is the value of the next schedule state after a correct trial (γV′ must be replaced by r in terminal trials). Solving Equation 15 for V gives Equation 13. The same procedure, though more involved algebraically, gives the values in the context-sensitive model:where Prs is defined as for the basic model. This system of equations must be solved self-consistently for the values Vrs. In the absence of errors, each Prs = 1 and Equations 9 of the main text follow. We have checked with simulations that the approximate solution given by Equations 9 gives a good approximation to the correct values on our dataset of monkeys' data. For this reason, Equations 2 and 9 were used to estimate the theoretical values when fitting the theoretical error rates to the experimental error rates. In the Random Cue condition, the cues define the states of the model. The model learns the values of the cues using the same algorithm specified by Equations 1, 3, and 8, with St≡cuet. The next cue is selected at random with equal probability for all cues if the trial is performed correctly, otherwise the current cue remains as the next. We set δ = rt+σV(cuet−1)−V(cuet) in terminal trials, in keeping with the rule adopted with valid cues. The average value of random cues can be obtained by averaging the update equation over all trial types that produce a different temporal difference δ, obtaining(16)i.e., ∑i fiδi(V) = 0, where V is the sought average value, fi is the average frequency with which trial i occurs, and δi is the temporal difference in trial i. In the basic model, it is sufficient to distinguish three trial types: correct terminal trials, incorrect terminal trials, and non terminal trials. The frequency (f) of correct terminal trials is N+Pc|V, where N+ is the average fraction of rewarded trials, equal to the number of schedules divided by the number of schedule states. In correct terminal trials the temporal difference is δ = r−V. Incorrect terminal trials occur with frequency N+(1−Pc|V) and have δ = −V; non-terminal trials occur with frequency 1−N+ and generate a temporal difference δ = (γ−1)V, whether the trial is correct or not. Replacing these values in Equation 16 and solving for V gives(17) Equation 17 defines V only implicitly and must be solved self-consistently to give the exact value of V. For the small error rates usually encountered with random cues, Equation 17 is well approximated by its version in the absence of errors (Pc|V = 1 for any V), i.e. . Note how V increases with γ and is constrained between the average collected reward N+r (for γ = 0) and r (for γ = 1). Setting γ = 0 (value at which V is minimal) is the same as assuming that the next cue is always unknown and its value is zero (cfr. Equation 4). This implies that having some expectation about the next state, even a random expectation as for the random cues, increases the values and hence the motivation to perform correctly. The context-sensitive model can be solved in a similar way, with in addition non-first trials to be taken into account. The final result is , from which Equation 10 of the main text follows under the approximation of small error rates, i.e., Pc|V≈1. Similar results are obtained in the case of post-reward expectation, where the value of the next state after a rewarded trial is not set to zero, as shown in a later subsection. Since it is required that V>0, this result requires (γ+σ)(1−N+)<1, or (γ+σ)<(Ns+1)(Ns−1)−1. This inequality is never violated in the basic model (where σ = 0), but it might be, and must be imposed, in the context-sensitive model, especially for long maximal schedule lengths. Similar restrictions coming from the values of valid cues also apply (e.g., σ<1/2γ from Equation 9). Here we show that it is not possible to obtain values dependent on schedule length (like in the context-sensitive model) by using a standard TD learning rule, which considers only future trials within the current schedule. The most general such rule can be written as , where the coefficients {ai}i = 1,2,…,T may depend on pre-reward number (i.e., the number of trials remaining before reward), but not on schedule length. t+T is the time at which the terminal trial is reached: when St is a terminal trial, the states St+i are not defined and their values are set to zero. It is more convenient to express the values as a function of the number, n, of trials remaining before reward (“0” being the terminal trial), conditioned on schedule length being s, V(n|s), as in Equation 7 of the main text. At equilibrium (δt = 0) one has V(1|s) = a1V(0|s). Since V(0|s) = r does not depend on s, V(1|s) does not depend on s, which in turn implies that V(2|s) = a1V(1|s)+a2V(0|s) does not depend on s, and so on. It follows by induction that V(n|s) does not depend on s for all pre-reward numbers n, and for any value of the coefficients an (some of which may vanish). This result holds also in the case of post-reward expectation, where the value of the next state after a rewarded trial is not set to zero and the forward terms in the series ∑ aiV(St+i) are taken to the (T+1)th term. As shown later, all values are re-scaled by a constant factor which does not depend on s, leaving the above argument unchanged. It follows from this argument that, to obtain the schedule-length effect, it is necessary either to look backwards at the values of previous trials in the same schedule (as in the context-sensitive model of the main text), or to take into account trials belonging to different schedules [61]. The notable TD(λ) rule (see, e.g., [6]), that has been suggested to be implemented by dopamine neurons of rats [62], considers only forward trials within the current schedule, and therefore cannot produce the contextual effect due to schedule length. In fact, here we show that for the reward schedule, the equilibrium values in the TD(λ) rule are the same as those obtained with the basic model. In TD(λ), all forward trials within a schedule are considered, weighted by imminence. Formally, when in state St at time t, the TD term is evaluated as(18)where the sum is over all states remaining until the terminal one (reached after T steps). λ is a parameter between zero and one; NT≡1+λ+λ2+…+λT−1 is a normalization factor; and is the i-steps-ahead prediction starting from St. (If λ = 0, Equation 18 reduces to δt = rt+γV(St+1)−V(St), the basic model of the main text.) The values are updated in the usual way: Vt+1 = αδt. In the reward schedule it is (only the terminal trial is rewarded), and Equation 18 reads(19)where t+T is the time at which the terminal trial is reached. The solution to δt = 0, with δt given by Equation 19, is the same as for the basic rule (Equations 3 and 4 of the main text), i.e., V(St+i) = γT−ir, or V(τ,s) = γs−τr if St≡{τ,s}, as can be proved, e.g., by direct substitution. This was confirmed in simulations of TD(λ)-learning of the reward schedule implemented through the use of eligibility traces, an alternative approach to TD(λ) (see [6] for details). So far, the value of the next state at the end of each schedule had been to set to zero. In other words, the learning rule following a rewarded trial is δt = rt+σV(St−1)−V(St), which written in this form applies to all cases, including the case of random cues and the basic model (where σ = 0). As said in the main text, this is the common choice in RL [6]. Here we show that behavior predicted by the model does not change if we assign a positive value to the next state (“post-reward expectation”). The reason is that, in a terminal trial, the next trial is not known and thus the same value must be assumed independently of current schedule. The actual value is immaterial, but for the sake of argument we shall make a choice. In the Random Cue condition, the current value of any cue chosen at random will do; in the Valid Cue condition, since the only available information is that the next state will be one of the initial trials {1,s}, the average value of all first trials will be taken, i.e., . It can be shown that the average value of each state is increased by a constant factor (1−γϑ)−1, where ϑ is the ratio of the value of the state post-reward to the value of rewarded trials. For the value chosen above (average of first trials), (note that this choice gives γϑ<1). Similarly, the value of random cues changes from to . Thus, there is no qualitative difference with respect to the case of no post-reward expectation of the main text. A similar argument also shows that the qualitative behavior does not change if σ>0 in first trials of each schedule in the context-sensitive model.

10.1371/journal.pcbi.1002649

Assessing the Relative Stability of Dimer Interfaces in G Protein-Coupled Receptors

Considerable evidence has accumulated in recent years suggesting that G protein-coupled receptors (GPCRs) associate in the plasma membrane to form homo- and/or heteromers. Nevertheless, the stoichiometry, fraction and lifetime of such receptor complexes in living cells remain topics of intense debate. Motivated by experimental data suggesting differing stabilities for homomers of the cognate human β1- and β2-adrenergic receptors, we have carried out approximately 160 microseconds of biased molecular dynamics simulations to calculate the dimerization free energy of crystal structure-based models of these receptors, interacting at two interfaces that have often been implicated in GPCR association under physiological conditions. Specifically, results are presented for simulations of coarse-grained (MARTINI-based) and atomistic representations of each receptor, in homodimeric configurations with either transmembrane helices TM1/H8 or TM4/3 at the interface, in an explicit lipid bilayer. Our results support a definite contribution to the relative stability of GPCR dimers from both interface sequence and configuration. We conclude that β1- and β2-adrenergic receptor homodimers with TM1/H8 at the interface are more stable than those involving TM4/3, and that this might be reconciled with experimental studies by considering a model of oligomerization in which more stable TM1 homodimers diffuse through the membrane, transiently interacting with other protomers at interfaces involving other TM helices.

G Protein-Coupled Receptors (GPCRs) are the largest family of membrane proteins targeted by drugs in clinical practice. Despite being at the forefront of biomedical research for many years, there is still considerable uncertainty about how GPCRs function at a molecular level. Although substantial evidence exists in support of their association in cell membranes, it is unclear how general and/or long-lasting this phenomenon is and whether it plays a significant role in GPCR function. This observation highlights the importance of understanding the rules that govern receptor-receptor interactions in living cells. Here, we report the results of computer simulations from which we estimated the relative stability of dimers formed by different, yet highly homologous, prototypic GPCRs. Our results suggest overall transiency in receptor-receptor interactions at the simulated different dimerization interfaces, but a variable strength of association depending on the specific residue composition or shape of the interface. The methodology we propose is expected to provide a level of molecular detail that is unattainable using current experimental techniques. Our ultimate goal is to generate unique hypotheses of receptor-receptor inter-helical interactions that can be tested experimentally to help elucidate the role of receptor association in GPCR function.

G Protein-Coupled Receptors (GPCRs) have been reported to associate in the cell membrane to form dimers/oligomers. While incontrovertible evidence exists for the constitutive dimerization of disulfide-linked family C GPCRs [1], the interpretation of oligomerization studies of members of the largest subfamily A of GPCRs [2] has often been difficult and controversial [3], [4], [5], [6], since the majority of the techniques used to infer GPCR association in living cells are unable to conclude unambiguously in favor of direct physical interaction between receptors. Most importantly, very few GPCR oligomerization studies have been able to provide any information about the fraction of receptors that are interacting at a given time or the corresponding dynamics of the interactions, rendering it impossible to determine, with any certainty, which molecular species (i.e. individual protomers, dimers, or higher-order oligomers) signal through interaction with intracellular proteins. These uncertainties have fueled an ongoing debate regarding the physiological role of GPCR oligomerization, exacerbated by the evidence that individual GPCR protomers, when reconstituted into nanodiscs, can signal to G proteins [6], [7], [8], [9]. Recent studies using single-molecule approaches have begun to address the details of the spatial and temporal organization of GPCR complexes in living cells. Single-molecule total internal reflection fluorescence microscopy (TIR-FM) was recently used to track the position of individual molecules of the M1 muscarinic acetylcholine receptor (M1R) labeled with fluorescent M1R antagonists in living cells [10]. Both single- and dual-color imaging experiments suggested a transient (∼0.5 seconds) formation of M1R dimers and a dimeric fraction of only ∼30% dimers at any given time. Although similar conclusions were reached by a single-molecule study of another family A GPCR, i.e the N-formyl peptide receptor [11], the possibility cannot be excluded that the fluorescent ligands used to image the single molecules in both studies might have altered the lifetime and preferred stoichiometry of the observed GPCR oligomers. It remains to be determined whether or not the features highlighted in these studies are the same for all GPCRs, or just specific subtypes. Recent fluorescence recovery after photobleaching (FRAP) studies of human β1 and β2-adrenergic receptors (B1AR and B2AR, respectively) [12] have raised the possibility that the strength of GPCR association may vary significantly, even among closely related receptor subtypes. Although the antibody-mediated capping approach used to immobilize receptors in these studies may have affected the interpretation of the results, B1AR was suggested to interact transiently (on a timescale of seconds) whereas B2AR appeared to form more stable complexes (on a timescale of minutes). Fung and colleagues [13] reported data in support of spontaneous B2AR oligomerization using Förster resonance energy transfer (FRET) between relatively small fluorescent probes attached to purified B2AR reconstituted into phospholipid vesicles. The authors hypothesized predominant tetrameric arrangement for the B2AR, although they did note the difficulty of unambiguously determining the stoichiometry of receptor oligomers from a reconstituted system. Additional FRET saturation studies showed greatest energy transfers for H8 and smallest for TM6, based on which, the authors proposed a preferential oligomeric arrangement of B2AR involving TM1 and H8 at the interface, similar to that previously suggested for the dopamine D2 receptor from a combination of molecular modeling and cysteine cross-linking experiments [14]. Further support for the simultaneous involvement of helices TM1 and H8 at an interface was recently provided by chemical cross-linking of endogenous cysteines in rhodopsin in disk membranes [15]. Several additional experimental studies support the direct primary involvement of TM1, as well as TM4 in GPCR oligomerization under physiological conditions [16]. Although an alternative dimerization interface involving both helices TM5 and TM6 was recently suggested by the crystal structures of the chemokine CXCR4 [17] and μ-opioid [18] receptors, its physiological relevance has not yet been demonstrated. Here, we sought insight into the dimerization free energy of models of human B1AR and B2AR based on high-resolution inactive crystal structures interacting at two putative interfaces involving TM1/H8 or TM4/3, using biased molecular dynamics (MD) simulations. Specifically, we combined umbrella sampling and metadynamics simulations to provide hypotheses of the role played by the interface sequence and configuration in imparting stability to the specific dimeric arrangements that we have simulated for both B1AR and B2AR. These studies provide a mechanistic insight into the association of GPCRs at putative dimerization interfaces at a level of molecular detail that is unattainable using current experimental techniques, yet crucial to guiding future experiments aimed at exploring the role of dimerization in receptor function. The results presented herein derive from approximately 160 microseconds (µs) of coarse-grained (CG), biased MD simulations. Two different dimeric arrangements were considered for the B1AR and B2AR. Specifically, these correspond to the interfaces illustrated schematically in the insets of Fig. 1, and are named TM4/3 and TM1/H8 after the helical domains involved in symmetric interactions. Notably, there can be different types of TM1/H8 interfaces depending on whether the TM1 residues involved in symmetric interactions are adjacent to TM7 or TM2. We focused on the interface with TM1 residues adjacent to TM7 based on inferences from recent cross-linking experiments of a family A GPCR [14]. We also note that this particular packing of TM1, simultaneously involving H8, is not possible using a rhodopsin structural template [14]. The corresponding MARTINI-based CG representations of these dimeric configurations were embedded in an explicit, CG, hydrated 1-palmitoyl-2-oleoyl-sn-glycero-3-phosphocholine (POPC)/10% cholesterol model membrane, resulting in approximately 30,000 particles. Following earlier studies [19], [20], we used three collective variables (CVs) to describe the dimeric arrangements recapitulated in the inset of Fig. 1A. Briefly, these CVs describe the separation (r) between the centers of mass (COMs) of two interacting protomers a and b, i.e., Ca and Cb, and their relative orientation, described by the relative rotational angles θa and θb. Thorough exploration of the TM4/3 and TM1/H8 interfaces of B1AR and B2AR homodimers was achieved through a pair of biasing forces applied to the CVs, i.e. a harmonic umbrella restraint applied to the separation (r) between the COMs of each of the protomers, and a history-dependent Gaussian bias applied to a pre-defined rotational angle range. POPC and cholesterol exchange at the interfaces was evaluated following the procedure described in Text S1 (see also Figs. S1 and S2) to exclude the possibility of desolvation problems at the protomer-protomer interface. Unbiased free energies were subsequently obtained as a function of the separation of the protomers, and of the angle exploration, as described in Materials and Methods. A theoretical framework, identical to that previously described by us in [19], was applied to derive the relative dimerization free energies and dimer lifetimes for each system, at the different interfaces. Results are presented in Table 1. A representative CG structure for each system was extracted from the energetic minimum in both r and (θa, θb) space and converted to an atomistic representation. Each structure was subsequently simulated for 1 ns in an explicitly represented, solvated POPC/10% cholesterol bilayer, and the details of the contacting residues at the interface were derived. Fig. 1A shows the reconstructed free energy surface (FES) as a function of the separation, r, between the COMs of the protomers for the B1AR (red) and B2AR (blue) homodimers. We note that the overall shapes of the corresponding curves are similar and their depths are equivalent, within the calculated error bars. Upon offsetting the curves to a zero value in the region beyond which the protomers were seen not to be interacting and were therefore designated as monomeric states (r = 4.5–4.8 nm), we observe that the depths of each of the two minima are similar between the B1AR and B2AR homodimers. To confirm the choice of the reference state, we sampled one of the systems, specifically the B2AR interacting at the TM4/3 interface, to larger separation distance (Fig. 1). As reported in Table 1, the primary minimum is at −4.8 kcal/mol and −5.8 kcal/mol for B1AR and B2AR, respectively. Using the same theory described in [19], and the equations 1–3 reported in Materials and Methods, dimerization free energies (i.e., mole-faction standard state free energy changes ΔGX°) of −2.3 kcal/mol and −3.7 kcal/mol were calculated for B1AR and B2AR, respectively. We proceeded to identify the relevant orientations of the protomers within each of these dimeric minima by comparing the FES calculated as a function of the angles (see Fig. S3) at r = 3.42–3.48 nm for B1AR and r = 3.42–3.46 nm for B2AR. The FES as a function of the angle for the TM4/3 interface indicates minima situated approximately at Θ1(θa,θb) = (0.2,0.4) (or the symmetric Θ1'(θa, θb) = (0.4,0.2) value) and Θ2(θa,θb) = (0.45,0.45) radians, respectively. Superpositions of energetically-optimized, all-atom reconstructions of representative structures of the Θ1 and Θ2 minima, obtained using the procedure described in the Materials and Methods section, are shown in Fig. 2A for both B1AR (red/pink respectively) and B2AR (blue/light blue). Symmetric inter-helical contacts, defined as average interaction distances between residue Cβ atoms less than or equal to the threshold distance of 11 Å during 1 ns unbiased all-atom MD simulations are listed for each of these representative structures in Table S1. Fig. S4A shows the location of the corresponding residues involved in these contacts. Fig. 1B shows the FES resulting from the calculation of B1AR and B2AR at interfaces involving TM1/H8 as a function of their separation, r, and calculated using the CVs illustrated in the inset of Fig. 1B. As for the TM4/3 interface, the overall shape and depth of the corresponding curves are similar, albeit not identical, between the B1AR and B2AR dimers, most likely due to slight structural divergences between the corresponding reference structures. For the TM1/H8 interface, the monomeric state was defined to be r = 5.5–5.9 nm. The primary minimum for each system at this interface indicates a separation between the protomers of ∼3.7 nm, corresponding to −12.0 kcal/mol and −12.9 kcal/mol for B1AR and B2AR, respectively. The ΔGX° of the TM1/H8 dimers is approximately −9.7 kcal/mol for B1AR and −10.0 kcal/mol for B2AR. The reconstructed atomistic TM1/H8 dimers of B1AR and B2AR obtained in the same way as previously described, for the TM4/3 interface, are shown in Fig. 2B. Once again, the relative orientations of the protomers in these specific configurations involving the TM1/H8 interfaces were determined from the FES shown in Fig. S3 as a function of the angles, θa and θb, at r = 3.72–3.77 nm for B1AR and r = 3.68–3.72 nm for B2AR. The minima are approximately situated at Θ1(θa,θb) = (0.45–0.50, 0.45–0.50). Contacting residues between the protomers during 1 ns of explicit simulation are listed in Table S1 and depicted in Fig. S4B. The spatial and temporal organization of GPCRs in living cells is currently the subject of lively discussion. Although recent applications of single-molecule approaches are beginning to address the preferred stoichiometry, lifetime, and ratio of GPCR dimeric/oligomeric complexes to individual protomers in living cells, they are unable to provide the molecular details of receptor association. Biased MD simulations of the type reported here ensure thorough exploration of the interface for GPCR complex systems in an explicit lipid bilayer, that can be used to draw conclusions about relative values of dimerization free energies and dimer lifetimes. We have carried out free energy simulations of different GPCR subtypes interacting at two different interfaces that have been suggested, by experiment, to form dimers under physiological conditions. The goal of this study was not to predict the most stable interfaces of dimerization for the GPCR systems studied, which would have required comparison of all possible interfaces, but rather to investigate the effect of different sequences and/or interface configurations on the strength of GPCR dimerization at two dimeric interfaces inferred to be relevant under physiological conditions. Our study indicates that interfaces involving TM1/H8 are the most stable and the most long-lived (minutes) of the two simulated interfaces for B1AR and B2AR homodimers, based on estimates derived from the calculated free energies. The orientation of the protomers in the TM1/H8 interfacial arrangement is consistent with the close interaction of R333 in H8 (at 13 Å in our dimeric model), which is the residue at which the greatest energy transfer was observed in the recent FRET study of B2AR [13] suggesting spontaneous B2AR oligomerization. The dimer corresponding to the TM4/3 interface appears to be significantly more transient (hundreds of microseconds to milliseconds) than the TM1/H8 interface for both receptors. The similar lifetimes estimated for B1AR and B2AR homodimers interacting at each of the interfaces tested suggest little difference in temporal organization between these two receptors, sharply contrasting with implications of the recent FRAP studies of human B1AR and B2AR [12] that motivated the present work. However, the possibility cannot be ruled out that the antibody-mediated capping approach used to immobilize receptors in the FRAP study might have caused the B2AR and B1AR to prefer interaction at different interfaces. While the estimated longer lifetime (minutes) of the B2AR dimer involving TM1/H8 at the interface may be considered in line with the views of the aforementioned FRAP study [12], the observation contrasts with inferences from other FRAP studies on dopamine D2 receptors [21], as well as conclusions of recent single molecule studies on muscarinic [10] or N-formyl peptide receptor [11] dimers, which suggest more transient interactions between GPCRs. Although we cannot set apart the contribution of the different membrane environments, we suggest that the results of our simulations may be reconciled with those experimental observations that imply only short-lived interactions by proposing a model of diffusion that features more stable receptor dimers, with TM1 at the interface, diffusing through the membrane and interacting transiently with one another at interfaces involving other TMs, to form short-lived tetrameric or higher-order arrangements. By superimposing the TM region of one of the two protomers of the simulated B2AR dimers on the active B2AR TM region of the recent crystal structure of the B2AR- Gs complex (PDB ID: 3SN6 [22]) we note interactions of the second protomer with the G-protein vary, depending on the specific dimeric arrangement of B2AR (Fig. 3). An interface involving TM4/3 would favor an exclusive interaction of the B2AR dimer with the alpha-helical domain of the nucleotide binding Gα subunit of the Gs protein (“GαAH” in Fig. 3A). In contrast, in a B2AR dimeric arrangement with TM1/H8 at the interface, the second protomer would not be involved in significant interactions with any of the G-protein subunits (Fig. 3B). It must be noted that these proposed conformations of the B2AR dimer in complex with the Gs were derived from simple superimposition, and thus would require additional simulations, beyond the scope of this study, to relieve the steric clashes arising from our use of the inactive conformation of the B2AR within the dimeric configurations. In summary, we have developed a protocol to assess the relative stability of GPCR dimers comprised of protomers interacting symmetrically at different TM regions that is robust within the standard caveats that apply to using a MARTINI-based CG model of proteins and membranes. We have recently published evidence for both small membrane dimeric systems [23] and GPCRs [20] that our CG simulations produce estimates of relative dimerization free energies that are in line with experimental data. Additional validations of the MARTINI model have been independently reported in the literature (e.g., see [24], [25], [26]). We herein demonstrate the dependence on the relative orientation of the protomers, in as much as the FES as a function of protomer separation is significantly different for B1AR and B2AR at interfaces involving either TM4 or TM1. Although the free energy and lifetime estimates reported herein are somewhat dependent on the nature of the starting crystal structures, our calculations appear to be consistent for the different systems we have reported, and within the limits of the theories we have employed. While this manuscript was under review, a paper reporting further elaboration of simulation protocols used to study the self-assembly of rhodopsin molecules [27], and now coupled with umbrella sampling calculations similar to those we published on delta opioid receptor [19], [20], has appeared in the literature [28]. Notably, the conclusions of these independent simulations about the relative stability of dimerization interfaces are in agreement with our calculations. We are confident that the protocol described herein can be generalized to begin deciphering the mechanistic details of dimerization of other GPCRs at a level of molecular detail that is unattainable using current experimental techniques. Although we do not expect to obtain an exact correspondence between the estimated lifetimes of GPCR dimers and those measured experimentally, our assessment of relative strength of association at different interfaces can be used constructively to predict specific interactions at the dimerization interface that might aid the design of experiments to assess the role of dimerization in receptor function. Initial molecular models of human B1AR and B2AR were built using available inactive crystal structures of the turkey B1AR and human B2AR (PDB identification codes: 2VT4 [29], chain B, and 2RH1 [30], respectively) as structural templates. First, missing segments in the B1AR and B2AR crystal structures were built using Rosetta [31]. Specifically, these segments corresponded to sequence fragments 256–260, 306–310, and 313–317 in the B1AR and the intracellular loop 3, sequence fragment 231–262, which had been replaced by a T4 lysozyme in the B2AR crystal structure. To restore the broken ionic lock between TM3 and TM6 in the B2AR crystal structure, a standard MD simulation of 100 ns was carried out after embedding the receptor into an explicit hydrated POPC/10% cholesterol bilayer, following the procedure described in [32]. The homology model of the human B1AR was built using Modeller v8 [33] after alignment of the human and turkey sequences. The collective variables used in these simulations were the same as previously described in our earlier publications [19], [20], and are illustrated here in insets of Fig. 1 for each simulated dimeric arrangement of protomers a and b. Briefly, these CVs correspond to (i) the distance, r, between the centers of mass Ca and Cb of the TM regions of protomers a and b; (ii) the rotational angle, θa, defined as the arccosine of the inner product of the normalized vectors connecting the projections onto the plane of the membrane of the centers of mass of the specific TM(s) at the interface (i.e., TM4/3 and TM1/H8) and of the two TM bundles (Ca and Cb), and (iii) the equivalent rotational angle, θb, for the second protomer. To aid simulation convergence, we restricted the exploration of θa and θb, using steep repulsive potentials, the details of which are in the SI, Table S2. All simulations were performed using GROMACS version 4.0.5 [34] enhanced with the PLUMED plugin [35], and the system components were represented using the MARTINI forcefield [36], [37], [38] (using the parameters from version 2.1 for the protein beads, and version 2.0 for POPC and cholesterol), as described in our previous publications [19], [20]. We focused on two dimeric interfaces that have received experimental validation under physiological conditions according to recent publications on GPCRs, specifically those involving TM4 or TM1. The resulting dimeric configurations are illustrated by cartoons in the insets of Fig. 1, and correspond to TM4/3 and TM1/H8 interfaces. Thus, initial configurations were built for homodimers of B1AR or B2AR, as described in our previous publications [19], [20]. Subsequently, the dimers were converted to CG representation and embedded in a pre-equilibrated CG POPC/10% cholesterol membrane; the system was then solvated and counterions were added to neutralize the charge and to generate a physiological salt concentration of 0.1 M. An elastic network was used to restrain the protein system according to the strategy described in our previous publication [19]. Briefly, standard secondary structure constraints were introduced as per the MARTINI prescription; in addition, following a protocol put forward by Periole and colleagues [38], we introduced elastic potentials between beads within a cutoff of 9 Å to maintain the integrity of the protein tertiary structure. In a modification of the original implementation, the force constants of the elastic network were weaker on loops (250 kJ/mol) and stronger on the helical residues (1000 kJ/mol), with values chosen by matching the Cα fluctuations to those of a 50 ns, all-atom simulation of the same system [19]. The mean and standard deviation of the RMSD of the TM regions, and the whole receptor, calculated over all the simulations and reported in Table S3 for each system, demonstrate that the proteins maintained reasonably native conformations in the TM regions. Metadynamics simulations [39] were carried out to generate the starting configurations for the umbrella sampling [40] simulations. During these metadynamics simulations, Gaussian bias was only applied to the CV describing the distance between protomers (r up to values representative of a protomeric system, see Table S2), and the CVs describing the relative rotation of the protomers (θa and θb) were restrained to ensure the starting structures all corresponded to interactions at the interface of interest only. Thus, we limited the sampling of the two rotational angles, θa and θb, to a predefined interval using upper and lower steep repulsive restraining potentials. Specifically, the upper and lower limits of this interval were set equidistant either side of the starting values from the initial dimeric configurations (see Table S1 for the range of θa and θb). Approximately 40 umbrella sampling windows, were prepared for each of the dimeric systems with a force constant of 2400 kcal/(mol⋅nm2), (see Table S2 for range of separation, r), and extra windows were included at values of r where the reweighted distribution of the distances was found to be insufficiently sampled. We combined umbrella sampling [40] with well-tempered metadynamics [39] simulations to ensure thorough exploration of both the distance and angle space available to the system, for each of the windows. Thus, in addition to the constant external harmonic bias of the umbrella sampling algorithm applied to CV1 (i.e., r), a time-dependent sum of Gaussian biases in the well-tempered metadynamics algorithm was applied to the angle CVs (i.e., θa and θb). In contrast to standard metadynamics, the bias potential in well-tempered metadynamics does not fully compensate the free energy surface, but rather depends on the underlying bias, decreasing to zero when a given energy threshold is reached [39]. Thus, not only does the computational effort remain focused on the physically relevant regions of the conformational space in these simulations, but also the convergence of the algorithm to a correct free energy profile can be proven rigorously. To ensure proper sampling, we checked that the chosen CVs could diffuse across one Gaussian size within the deposition time, so that local instantaneous equilibrium of the CVs would be satisfied. An improper choice of the bias update rate and Gaussian size would have resulted in trajectories failing to show multiple transitions across different minima and dependence on the initial starting configuration of the systems. None of the above was observed in our simulations, which showed instead increasingly fast diffusion of the CV dynamics due to the flattening of the underlying free energy surface, suggesting a proper sampling of the phase space of the systems. In all cases, the initial height of the biasing Gaussians was set to 0.12 kcal/mol, with a deposition stride of 10 ps, σM = 0.035 radians, and a bias factor of 15. Each window simulation was run for at least 1 µs (and up to 2 µs in regions of r thought to require additional sampling), resulting in a cumulative simulation time of ∼40 µs for each system, and a total of approximately 160 µs for the two receptor systems. Finally, using a model potential, we provide (in the SI section) validation that the accuracy of the method combining umbrella sampling and metadynamics is comparable to those of standard multidimensional umbrella sampling or metadynamics (see both corresponding SI text and Fig. S5). The well-tempered metadynamics bias acting on the angle CVs distorts the probability distribution of the distance CV, thus requiring reweighting before equilibrium Boltzmann distributions could be reconstructed with WHAM. To recover the unbiased probability distribution of the distance CV from well-tempered metadynamics, we used the reweighting algorithm originally derived in [35] and direct the reader to the SI section for a description of this algorithm, as well as for the parameters used and additional technical details. After recovering the unbiased probability distribution of the distance CV from well-tempered metadynamics, we used the well-documented WHAM technique [41], [42] to reconstruct, for each simulated system, the free energy surfaces as a function of the separation, r, between protomers (Fig. 1); the technical details are provided in the SI section. An error analysis of the reconstructed free energies was carried out combining recently proposed methods for the error estimation in umbrella sampling [43] and well-tempered metadynamics [44], [45] simulations. Equations used to estimate these errors are reported in the SI section. For each of the dimers at the primary minima (in Fig. 1), we have also reconstructed the FES as a function of the angles θa and θb. To reweight the angle distribution at a fixed protomer separation r, i.e., to remove the umbrella bias and obtain unbiased free energies as a function of the angles, we used the same algorithm as above [35], but we accounted for the umbrella potential as an external potential. Fig. S3 shows the FES as a function of (θa,θb) for the homodimers of the receptors at the two different interfaces. The minima marked Θ1, Θ1' and Θ2 are the principal minima from which representative minimum structures were extracted. Upon derivation of the FES as a function of r and the angles (θa,θb), for each dimeric system, we extracted a representative frame from the trajectory that corresponded to the minima therein. We then used Pulchra [46] to convert each of the CG models to an atomistic representation. These representations were solvated in an atomistic POPC/10% cholesterol membrane, energy minimized and simulated with harmonic restraints of decreasing strength applied to Cα atoms for a few picoseconds. Where necessary, we used an adiabatic biased MD simulation (of ∼5 ns) to improve the integrity of the helical structure of the re-converted receptors, by steering the Cα of the transmembrane regions toward the original atomistic structure of the B1AR or B2AR (i.e., that prior to coarse-graining), before 1 ns of unrestrained, all-atom simulation, during which the analyses to obtain the lists of contacts presented in Table S1 were conducted. We can compare the strength of dimerization at each of the interfaces by the relative values of ΔGX°. In equation 1, we remind the reader of the formulation for ΔGX° in equation 5 of [19]. Specifically, the mole fraction standard free energy change can be expressed as:(1)where R is the universal gas constant, T is the temperature in Kelvin and KX is the association equilibrium constant on the mole fraction concentration scale. Following our derivation in [19], KX is approximately equal to:(2)where NL is the number of lipids in the membrane of area A, and KD is the dimerization constant expressed as in surface concentration units. For our membrane patch, NL/A was ≈1.65×106 µm−2. Using the theory originally proposed by Roux and colleagues [47] and adapted by us to the case of GPCR dimers [19], the dimerization constant can be expressed as a function of the free energy F(r) of the system constrained to a predefined angular region Ω0, where r is the distance between the interacting protomers, Ω = (θa,θb) describes their relative orientation, and β = 1/kBT. This correction was necessary because the relative orientation between protomers had been constrained into a region Ω0. In such a case, by extending the integral up to the maximum distance rD allowed for dimeric states, KD is given by equation (3):(3)Here, ||Ω0||, i.e., the product of the allowed ranges for θa and θb in radians (see SI and details in Table S2), is (maxθ-minθ-σM)2≈0.512≈0.26 at TM4/3 interfaces, and 0.462 = 0.22 at TM1/H8 interfaces. The above theory can be used to estimate the kinetics of dimerization by approximating the diffusion-limited association and dissociation rates, kon and koff, at long timescales, following the Smoulchowski theory in two dimensions [48] (i.e., the membrane plane). This derivation is the same as that described in [19]. kon is given by equation 4:(4)where DC is the sum of the diffusion constants of the two protomers, R is the sum of the protomer ratios, and γ is the Euler-Mascheroni constant and t refers to the experimental timescale of diffusion. Subsequently, koff = kon/KD. An initial concentration of dimers of [D]0 evolves over time to give a concentration of:M(5)and thus the half-life of dimers can be estimated using:(6)

10.1371/journal.pcbi.1000455

Identifying Currents in the Gene Pool for Bacterial Populations Using an Integrative Approach

The evolution of bacterial populations has recently become considerably better understood due to large-scale sequencing of population samples. It has become clear that DNA sequences from a multitude of genes, as well as a broad sample coverage of a target population, are needed to obtain a relatively unbiased view of its genetic structure and the patterns of ancestry connected to the strains. However, the traditional statistical methods for evolutionary inference, such as phylogenetic analysis, are associated with several difficulties under such an extensive sampling scenario, in particular when a considerable amount of recombination is anticipated to have taken place. To meet the needs of large-scale analyses of population structure for bacteria, we introduce here several statistical tools for the detection and representation of recombination between populations. Also, we introduce a model-based description of the shape of a population in sequence space, in terms of its molecular variability and affinity towards other populations. Extensive real data from the genus Neisseria are utilized to demonstrate the potential of an approach where these population genetic tools are combined with an phylogenetic analysis. The statistical tools introduced here are freely available in BAPS 5.2 software, which can be downloaded from http://web.abo.fi/fak/mnf/mate/jc/software/baps.html.

The study of bacterial population biology is complicated by the fact that, although bacteria are largely asexual, they can also exchange genetic materials through homologous recombination. Unlike eukaryotes, recombination in bacteria is not an obligatory process. Furthermore, the recombination mechanisms are subject to many biological and ecological factors that can vary even within different populations of the same species. Although increasing evidence for homologous recombination has been found in many bacterial species, determining the frequency of recombination and understanding the influence that it exerts upon the evolution of bacterial populations remains a challenging work. In this article, we provide a dynamic picture of recombination within and between closely related bacteria species. Through an integration of several Bayesian statistical models, our method highlights the importance of a quantitative estimation of recombination. Our analyses of a challenging multi-locus sequence typing (MLST) database demonstrate that combined analyses using both traditional phylogenetic methods, explorative MLST tools and Bayesian population genetic models can together yield interesting biological insights that cannot easily be reached by any of the approaches alone.

It has become increasingly evident that recombination plays a major role in shaping the genetic structure of bacterial populations. Whether or not certain populations (as defined by allele frequencies) are more likely than others to undergo recombination, either as donors or recipients of DNA, is not well understood, though there are several biological reasons why this might be the case. Such preferential recombination, which we may intuitively describe as currents in the gene pool [1], should lead to a greater degree of admixture between the populations in question, and this should be detectable using DNA sequence data. Conceptually related investigation of highways of gene sharing among bacterial species at a general level was done by [2], who found evidence for uneven distribution of transfer intensity among groups of prokaryotes. Discovery of such gene flow currents is scientifically interesting in its own right, as a means for characterizing populations and reflecting upon accumulated taxonomic understanding of their heterogeneity. However, there are other potential uses for detailed knowledge concerning the genetic structure of a bacterial population, e.g. when it can be connected to patterns of virulence and antibiotic resistance. Statistical analysis of molecular variation and reproductive isolation in natural populations is in many cases far more challenging for bacteria than for eukaryotic organisms, due to difficulties in acquiring broad-coverage samples and the putatively complex admixture events [3]. Traditional population genetic tools for inferring genetic barriers within a population, such as measures [4], are not usually applicable to bacterial molecular data given the lack of relevant populations to condition the calculations on, albeit some exceptions exist (see, e.g. [5]). Standard phylogenetic analyses, on the other hand, may provide a distorted view of the ancestral relationships among bacteria when recombination events are sufficiently common in a population. Moreover, they do not yield a detailed and easily interpretable picture of the patterns of admixture and eventual genetic barriers, as such constructs are not present in the standard phylogenetic models that can be routinely applied to large data sets. However, an algorithmic approach to phylogenetic analysis which can build networks for hundreds of taxa and can be useful for data sets harbouring recombination was introduced by [6]. A model-based phylogenetic method (ClonalFrame) that deals explicitly with recombinations was introduced by [7], however, it does not easily scale up to the level of population complexity we are here interested in, due to the extreme computational intensity of the model fitting for large databases. With the above-mentioned difficulties, it is hardly surprising that a Bayesian statistical approach based on explicit admixture models has recently gained popularity in studies of bacterial populations [8],[9]. Such models are anchored in the general idea of a probabilistic partition, where an unknown origin of an arbitrary quantity (for example, the membership of an individual) is inferred through the conditional probability of the origin over the range of putative alternatives (commonly referred to as clusters), given the observed features of the quantity. Application of such partition models has been made possible by a class of generic Markov chain Monte Carlo (MCMC) algorithms [10], that can be used for fitting the models to molecular data. Despite the success of the standard MCMC approach in a variety of studies of bacterial populations (see e.g. [11]), it is clear, both theoretically and practically, that the performance of the standard MCMC computation decreases rapidly as the complexity of the estimated population structure and the size of the investigated data set increases [10],[12]. To address this, an array of methods has been introduced and implemented in the software BAPS [13]–[16]. Here we introduce a graphical characterization of recombination patterns from MLST data using a weighted network with statistically identified populations as cluster nodes and estimated average levels of DNA transition as relative gene flow weights. Also, we introduce a model-based representation of the molecular variability of populations and their affinities towards each other. We refer to this as the genetic shape of an identified population. However, it is important to notice that a population identified by BAPS may have a different interpretation in different evolutionary contexts. The BAPS models target for identifying molecular evidence that links a particular group of strains together in terms of sufficiently similar nucleotide frequencies. Thus, such a population may for instance arise in the analysis due to common ancestry within a clonal complex. In contrast, a population can also be identified from the traces left by recombination events which have imposed considerable gene flow between separate lineages of strains. Also, under certain circumstances a more heterogeneous population may arise analogously to long branch attraction in phylogenetics, in particular, when very limited numbers of strains from the corresponding lineages are present among the analyzed samples. All these three cases are illustrated in our analyses. As BAPS is capable of capturing a variety of distinct biological signals hidden in molecular data, interpretation of the identified populations must be done with care, using preferably both complementary phylogenetic methods and auxiliary knowledge about the strains under investigation. To illustrate the levels of complexity at which our methods can operate, we consider a population sample of 5086 strains that have been identified as Neisseria meningiditis and Neisseria lactamica species. We also present analyses of simulated data to demonstrate the potential of our Bayesian approach to handle large databases and complex genetic population structures. Our analyses illustrate that biological insights to complex data are best gained by combining several complementary methods of analysis. Assume that the target population consists of genetically distinct populations, among which the extent of gene flow is to be modeled. Usually, and the genetic population structure associated with it are a priori unknown. In our statistical approach presented later we consider in detail the inference of these from molecular data. Here we aim to estimate the strength of gene flow via a stochastic characterization of the rates of admixture between the identified populations. Let , index the populations and let represent for the population the probability of an strain acquiring DNA from bacteria present in the population . DNA acquisition could be understood as an aggregated result of the currently known mechanisms (conjugation, transduction and transformation). Conditional on the probability , it is possible to consider a sample of unrelated strains from population to represent Bernoulli trials, where the binary outcome , refers to the success/failure of DNA acquisition from this particular source. These are obviously considerably simplifying assumptions, but they allow us to characterize patterns of admixture. Were the outcomes , known, the relative admixture could simply be characterized by . However, we note that in reality represents intrinsically unobservable latent events during some interval of the evolutionary time scale under consideration. Assuming that a particular strain within population has acquired DNA from the population (i.e. ), we may attempt to quantify the intensity with which such events have occurred over the analysed sequence. A multitude of statistical break-point models designed to capture such recombination traces have been introduced in the literature, e.g. [17]–[19]. For such models the focus has typically been on a small number of short viral genomes, to identify the locations where putative recombination events have taken place. In the most basic form, recombination may be represented by a homogeneous spatial Poisson process , where the events correspond to the number of recombinations within the genome of an strain , such that the DNA is acquired from the population . It follows for such a process that the stochastic variable , with equal to the total length of the considered sequence, has the Poisson distribution(1)where represents the average rate of events in which DNA is imported from population to . Again, if the outcomes were observed, the average rate could be statistically quantified, e.g. as , by using the maximum likelihood estimate. To arrive at a statistical characterization of the rates of admixture among the populations under the above framework, let denote a matrix of probabilities, such that the element equals . Further, let the matrix , with the elements , represent collectively the Poisson intensities. Let be a directed graph with the populations as the node set , and as the arc set. Each arrow in can now be associated with a weight depicting the rate of admixture from to . For instance, a gene flow weight matrix can be defined in terms of the elementwise matrix product , with the convention that the diagonal elements are normalized by the other elements on th row of . When an element equals zero, it is natural to set , i.e. the corresponding arrow is absent in . It follows from the definition of that these parameters remain unidentifiable when the events are unobserved, as a suitable rescaling of the model configuration can yield identical likelihoods. The statistical challenge related to this context is further accentuated by the fact that the underlying genetic structure, i.e. the number of underlying populations as well as their molecular characteristics, is unknown a priori. Modern Bayesian statistical framework utilizing state-of-the-art MCMC computation can in principle be thought to provide a suitable setting for fitting such models to MLST sequence data. However, the computational complexity associated with the models suggests that formal posterior inferences would remain beyond the bounds of computational tractability even for only moderately sized population samples. This is crucial, as to study such problems we require large samples with a broad coverage of the genetic variation in the underlying population. Therefore, we consider here an approximate inference strategy to estimate , which is computationally manageable for large samples, while still providing a reasonable statistical characterization of parameters that can be interpreted in terms of and in the above model formulation. Assume we have a sampled set of aligned DNA sequences , from genomic regions in bacterial strains. A concatenated sequence for an strain is denoted by and refers jointly to all the DNA sequence data from the strains. For any subset of strains from , the notation will be used for the DNA data observed for these strains. Let be a partition of the strains representing an underlying genetic structure (i.e. a representation of a genetic mixture model), with the clusters corresponding to genetically distinct populations. Hereafter we will use the terms ‘cluster’ and ‘population’ interchangeably. Mathematically, () is a collection of subsets of , such that , for all . Symbol defines the space of all such partitions for a given . For any partition , cardinalities of the populations are denoted by . In a series of earlier works in [12],[13] various stochastic partition models have been introduced for Bayesian inference about genetic population structure based on different types of molecular information. The mathematical motivation of the stochastic partition approach was recently derived by [20]. Under these models, the biological hypothesis corresponding to any particular partition , states that the strains allocated in the same cluster represent a sample from a genetically distinct population, and thus, the partition provides a qualitative representation of the underlying genetic population structure. Let denote the a priori uncertainty about the underlying genetic structure in terms of a probability distribution over the space . Then, we may specify the probability measure(2)where is the marginal likelihood of the sequence data given the structure. The posterior distribution of given the sequence data is determined by Bayes' rule according to(3) Here we use a Bayesian estimate of the genetic structure provided by the posterior mode(4)or possibly separately for a range of such estimates identified by stochastic optimization, if the molecular data are not decisively supporting a single structure. Methods to numerically obtain such estimates have been introduced by [12],[20]. The marginal likelihood for the observed sequence data given any structure has under the stochastic partition framework the following product form(5)which encapsulates a symmetry among the underlying populations , as any specific labeling of them without further auxiliary information would not be possible. However, to explicitly specify the terms a number of assumptions are required. Here we exploit the genetic linkage model developed by [15] to provide an explicit characterization of the terms in (5) for MLST type sequence data. The linkage model captures dependencies in the sequence data in terms of a Markovian model for each gene, such that each population has its own nucleotide frequency parameters, the joint prior distribution of which factorizes according to the Markovian model. Utilizing the standard results for so called hyper-Markov probability laws for multinomial-Dirichlet distributions, it is possible to calculate the marginal likelihood analytically given any value of S. This result is of importance, because it enables the development of an efficient learning algorithm which avoids Monte Carlo errors associated with the nucleotide frequency parameters in the populations specified by the genetic structure model. However, it should be noted that because the genetic mixture model operates at the level of sequence data, it is vulnerable to misalignments of the sequences similar to other comparable statistical methods. Given a plausible representation of the underlying genetic population structure based on (4), our aim is to obtain a model-based characterization of the rates of admixture between the populations, such that an estimate may be derived for the gene flow weight matrix . This sequential estimation strategy is motivated by the observation, that joint estimation of and the extent of admixture leads to problems with statistical identifiability and over-fitting discussed by [14]. In particular, they described a property of the admixture models which enables an increase in the number of populations without necessarily increasing the effective number of parameters (allele frequencies) in the model. This is in contrast with genetic mixture models, where such an increase always occurs as a function of the number of populations, thus resolving the problem with weak identifiability and/or high dependence of the inferences on the particular prior distribution used in the analysis. The most recent version of the BAPS software (5.2) contains an implementation of the admixture estimation algorithm introduced by [16] under the linkage model of [15]. Here we use this procedure to estimate the extent of admixture among the populations. Let be an estimate of the genetic structure underlying the sample according to (4), and let , be a vector of admixture coefficients representing the proportion of the genome of strain having ancestry in the corresponding populations, respectively. Let be the joint probability of the data from the region for strain under population . Then, the admixture model likelihood for the data in is determined by(6)The marginal posterior mode estimates of are obtained by numerical maximization combined with a simulation, to account for the uncertainty about given the partition estimate . As illustrated by [14], the posterior distribution of does not entirely plausibly represent the statistical uncertainty about , as the strain coefficients may in some cases have a mode in the interval from ∼.1 to ∼.2, while still reflecting only random fluctuations in in the populations, in contrast to real ancestry in a particular population, say . The issue was solved in [14] by calculating simulation-based for under the null hypothesis of no admixed ancestry. In the sequel, let denote such a for an strain . We now combine the statistical tools from [14], and [15] to obtain an estimate of . Firstly, populations are estimated using the posterior mode partition . Then, for each identified population , the extent of admixture events is estimated for using a significance level , such that(7)where is an indicator function equal to one when the argument is true, and zero otherwise. The estimate (7) is for the population an average relative amount of (significant) DNA acquired from population , thus representing a combination of an average recombination intensity and the propensity of recombination events taking place between these populations. It should be noticed that the admixture model ignores possible contiguity of genes or genome regions. However, the genes present in MLST analyses tend to represent quite distant genome regions, which motivates the assumption of independence. If the genes are taken from a more contiguous region, it is possible to treat them as a single linked region in the model by concatenating the sequences prior to the genetic mixture analysis. Notice also that the above gene flow estimates can be complemented by investigating separately the rate and size of the exchange events. The rate of exchange events is represented by the proportion of strains in a population showing significant admixture from a particular source. In turn, the size of the exchange events is revealed by characterizing the values of the corresponding estimated admixture coefficients. The latest version of BAPS (5.2) contains an implementation of the estimation procedure leading to (7), such that high-resolution images of the directed graph with the associated weight matrix can be produced directly with the software. As illustrated in the RESULTS section, this facilitates the analysis of large data sets for which the numerical estimates of admixture can be very tedious to examine. The above presented statistical models and tools provide means for assessing the number of genetically isolated populations and the extent of recombination among them. However, this leaves open questions related to the underlying genetic population structure. In particular, the model summary estimates do not provide any information on the area occupied by the population in sequence space, which we term its genetic shape. By a genetic shape we refer both to the molecular heterogeneity present in a population, as well as the genetic affinities of its members towards other identified populations. We will illustrate that an investigation of the genetic shapes in this sense can yield useful characterizations of the population, pinpoint interesting subgroups of strains, and eventually provide clues to relate the genetic structure to some auxiliary information. Let be the estimated genetic population structure and the structure where the strain has been moved to the population . The relative genetic affinity of the strain towards population can be quantified in terms of the change in the log-predictive likelihoods(8)which is always non-negative given that we have identified the true posterior optimum (4). However, even when does not equal the global posterior optimum, our estimation algorithm is designed in such a way that negative values of (8) cannot be obtained, as any parameter configurations in the neighborhood of leading to an improvement of the posterior probability will be detected. The difference (8) can be interpreted as the amount of information we lose in the prediction of the molecular characteristics in when the strain is assigned into another population, given that the remaining population structure is kept fixed. Thus, at the boundary, when (8) is equal to zero, no information will be lost. From the genetics perspective, the distribution of the values , reflects the genetic shape of the population towards the population . It is clear that this shape does not necessarily have an easily interpretable geometric configuration in low enough dimensions (1–3), such that it can be visualized. However, the shape of the distribution of , can still be used to reveal patterns of interest, which is illustrated in the RESULTS section. To numerically characterize the genetic shape of a population using the values of (8) for , we use a kernel density estimate of the underlying distribution of the affinity measures. This is implemented in BAPS 5.2, which outputs graphical displays of the density curves. These are based on the standard Gaussian kernel with the Gaussian optimal bandwidth (see, e.g. [21]) according to(9)where , and further is the maximum likelihood estimate of the standard deviation of values. Such density curves will provide useful information concerning both the within and between population molecular variation as well as affinity. The simulated MLST data sets were generated by assuming a tentative gene flow graph with the weight matrix is changing randomly. The gene flow graph estimated by BAPS 5.2, denoted as , was then compared with for evaluating the prediction accuracy. The characteristic of genetic shapes for the identified populations was also investigated for a wide range of scenarios given by . The assumptions for the data generation are based on a simplified, yet reasonable evolutionary model of bacterial populations. We assumed that each population has a common local ancestor, and further back in time these local ancestors originated from a common ancestor of the whole population, termed as a global ancestor. This assumption enables a tree representation of the evolutionary relationships among the populations (Figure 1). It is important to note that we do not explicitly model the time at which these ancestral events occurred and therefore the edges in Figure 1 are in arbitrary length. When ignoring recombination, the strains in a population will differ from each other only through the accumulation of point mutations. The mutations may have accumulated in two consecutive stages. In what follows, we referred to a mutation that occurred prior to the local ancestors as a stage-1 mutation, and a mutation that occurred afterwards as a stage-2 mutation. We assumed the infinite-site model of mutation, which implies that at most one mutation per site can occur in the DNA sequences [22]. This would imply that stage-1 mutations provide heterogeneity that leads to population diversification, while stage-2 mutations generate variations within a population. We further assumed that these two types of mutations occur independently of each other and result in a number of segregating sites. Let denote the total sequence length, then the expected number of segregating sites , where and are the mutation rates for the two stages. To simplify the problem, we considered recombinations that always lead to changes of DNA, so that they can be detected as admixture events. This corresponds to assuming that the DNA introduced by admixture needs to be distinctive compared to the homologous sites that have been observed within the population. Given the tree representation of the population evolution, such recombinations are restricted to occur at stage-1 mutation sites only. Following these assumptions described above, the expected numbers of stage-1 and stage-2 mutation sites can be obtained byandwhere is the expected number of segregating sites; is the ratio of the two mutation rates. We considered the time length in stage-1 is longer than that in stage-2, so we did set . For data simulation we used and , and set an equal population size for all the populations. A simulated population data set thus contains sites and strains, where is the number of populations. We specified a putative gene flow graph that consists of populations and the arrow set is specified in Figure 2. The rates of admixture between populations are characterized in the matrix , which is by definition a product of and . Therefore by simulating and we can generate a parameter set in that conforms to the graph structure in Figure 2. We chose a consistent sampling scheme for such that the diagonal elements for , and the non-diagonal elements are uniformly distributed. is also sampled from the Uniform distribution , but with the row constraints , since refers to the fraction of DNA sequence acquired from a particular source population. Sampling a data set according to the putative population structure consists of three steps. First, a global ancestor of 500 segregating nucleotides was randomly simulated and for each of the six population a local ancestor was generated by randomly altering each nucleotide of the global ancestor with the probability 0.8. The sample strains for each population were generated by randomly mutating the local ancestor with the probability 0.2. The strains that have been recombined were randomly selected according to the parameter , and for each of the recombined strain the actual amount of recombinations was determined by . Using the procedure described above, a population data set can be simulated for each of the selections of and . The population structure analysis was done with our Bayesian framework implemented in BAPS 5.2. We reported the accuracy of BAPS partition as choosing different values of and . Once the true partition has been identified correctly, we assessed further the accuracy of the predicted gene flow structure, i.e. the similarity of graph topology between and . Note that the non-diagonal elements in determine the propensity of acquiring DNA through recombination from to , therefore a larger implies that the recombination would affect a higher proportion of strains in . The increasing admixture propensity would make the recombination unidentifiable, since our Bayesian framework tends to favor the alternative hypothesis that the allelic frequency at the recombination site is more likely an effect of the stage-2 mutations, rather than a consequence of admixture. We therefore expected a negative correlation between and the gene flow structure accuracy. We also expected that in order to obtain a reliable partition estimate, the non-diagonal elements in should be near zero, since a small rate of recombination along the DNA sequences might not perturb the population structure in a large scale. A large , however, implies frequent recombination that might blur the population boundaries, so that the original population structure could be no longer identified. To illustrate the presented methods with a real data set, we applied BAPS 5.2 to MLST bacterial data. MLST(multi-locus sequence typing) is an approach to the unambiguous characterization of bacterial strains. The internal sequence of seven housekeeping genes, which include the abc Z, adk, aro E, fum C, gdh, pdh C and pgm genes are obtained and unambiguously characterize the strain. The strain sequences are generally reported to the publicly accessible MLST strain databases (see, e.g. http://www.mlst.net), which are currently hosting a fast growing number of bacterial genera and also a few eukaryotic organisms. We chose the Neisseria species for validation as homologous recombination is known to be frequent in both N. meningitidis and N. lactamica species [23],[24]. Furthermore, occasional horizontal gene flow over the species boundary has also been observed [25]. However, it is unclear to what extent the gene flow occurs and its consequence for population structure. To investigate this we applied BAPS 5.2 to a MLST sample which contains 4823 strains of N. meningitidis and 263 strains of N. lactamica. The data was accessed for analysis from the Neisseria MLST database at 17/3/2006 [26]. For such MLST type sequence data, we utilized the genetic linkage model [15] to account for dependency within the neighboring nucleotide bases. To assess the ability of our methods to find correctly or nearly correctly the populations hiding in the data, we considered a simulation scenario for generating a bootstrap sample which contains the strains randomly selected from a sub-collection of the identified Neisseria populations, using the procedure as follows: By repeating the scenario multiple times (we use 5 repeats) for each , we can check how well the resulting partition agrees with the chosen and how close the partition is to the general setting. This approach allows us to investigate the statistical power to correctly detect populations when the number of available strains is quite limited per population. Conditional on the identified population structure, comparative rates of admixture between the populations can be further estimated and summarized in a gene flow graph. We plotted the genetic shapes of several populations in N. meningitidis which show significant gene flow towards the N. lactamica species, and also compared their similarity in terms of admixture tendency. To investigate whether the signals of admixture varied considerably over the seven genes, we performed a bootstrap analysis where a single gene at a time was excluded when inferring the rates of admixture. The analysis was performed conditional on the clusters identified using the original complete data set. The relative importance of each gene could then be tentatively summarized by calculating for each cluster the average changes in incoming and outgoing gene flow while treating the estimates from the complete data as a baseline. Phylogenetic analysis of the Neisseria data was performed using MEGA v.4.0.2 [27]. Neighbor-Joining (NJ) tree was constructed with the maximum composite likelihood model assuming rate uniformity and pattern homogeneity. eBURST analysis of the Neisseria data was performed using the default options in the online version 3 available at http://eburst.mlst.net [28]. We reported the partition accuracy with respect to different choices of and under a constant population size in one scenario and in another. The partition accuracy measured by the Rand Index (RI) (see e.g. [29]) is summarized as a grey-scale map (Figure 3). In the presence of a small amount of admixture, i.e. , the tentative population structure can be identified with high accuracy. As the recombination rate increases over a critical threshold, e.g. as for the current setting, the partition accuracy drops quickly. Therefore, a higher recombination rate, indicated by a lower , would imply a lower partition stability. Such an observation matches our expectation that excessive amount of admixture tends to obscure the putative population structure. We may look further into the gene flow graph prediction only if the genetic structure (i.e. the true partition) is correctly identified. We used Hamming distance as a measure of gene flow structure accuracy and the result is shown in Figure 4. The gene flow graph structure can be satisfactorily discovered when and . However, a negative correlation between and was also noticeable. This result suggests that if admixture affects a population through a small proportion of strains, then the chances of its correct estimation by BAPS 5.2 are high. In contrast, admixture that occurred at most of the strains is more likely to be ascribed to variation arising within the population by mutation. These observations are in harmony with the investigation of the effect recombination intensity on the emergence of distinct populations for a bacterial species in [3]. Extensive levels of recombination will act as a cohesive force keeping populations together as a large gene pool, which consequently prevents the statistical detection of the recombination in terms of such a population genetic model as investigated here. This is entirely reasonable, because any substantial genetic population boundaries will not exist under such circumstances, and consequently, recombinations over population boundaries are not meaningfully defined, let alone detectable by a statistical model. Moreover, if certain parts of the data are too weak for reliable admixture inferences due to very small population cardinalities in the genetic mixture estimate, it is possible to leave the admixture coefficients undetermined for them using the option available in BAPS, as discussed in [14]. The extensive simulation study performed by [30], showed that the BAPS inferences about the genetic structure were generally sensible from a phylogenetic perspective, even in the presence recombination events, provided that the data are at least reasonably informative. With very weakly informative molecular data, it cannot be expected that any detailed statistical population genetic model would provide highly accurate estimates of the population characteristics. We used a simulated data set for illustration of genetic shapes represented as the density estimation in (8). The data set was generated with and . Figure 5 shows the estimated genetic shapes using population 2 as the reference, as compared to the other five populations. It can be seen from Figure 5 that the influence of admixture between the populations is reflected also on the genetic shapes. For example, the density curves for population 1 (red) and for population 3 (blue) are more shifted towards zero than the other populations, and hence imply a closer relationship to population 2. This is not surprising since population 2 is a common donor of DNA to populations 1 and 3 (Figure 2). On the other hand, the density curve for population 3 appears to have two modes, which is a feature exhibited in neither population 1 nor any other populations. Note that population 3 is the only population which donates DNA to population 2. We might use the bi-modality of a density curve as a potential indicator of gene flow to the reference population. In total 32 BAPS populations are identified, where three populations (numbered as 8, 29 and 32) belong to the N. lactamica species and the remaining 29 populations are labeled as N. meningitidis species. For accessing the robustness of the identified population structure, the partition determined using the whole data set was compared with the partition using bootstrap data generated according to the simulation scenarios. Figure 6 shows the adjusted Rand Index as a result of the comparison. Our partition method is able to identify the population structure with good accuracy, even though the performance may decrease as the complexity level of the data increases and when the number of available strains per population is quite low. It should be noted that the number of strains in the bootstrap samples was typically much smaller than the number of strains assigned to a particular population in the analysis of the original data. This illustrates that the population identification becomes highly stable when the sample sizes are sufficiently large. The results of admixture analysis for the Neisseria data set are summarized in Figure 7. The graph was obtained by fixing the admixture significance threshold at 0.05 and then pruning the arrows with gene flow strength below 0.03. It can be seen from the grey box highlighted in Figure 7 that two admixture arrows that imply inter-species gene flow remain significant, where two of the N. meningitidis populations (11 and 19) are constantly influencing the genetic makeup of one population of N. lactamica (population 29). The admixture arrows are uniformly directed from N. meningitidis towards N. lactamica, implying that N. meningitidis might donate genetic materials into N. lactamica, while the gene flow in a reversed direction is not supported by the analysis. The admixture estimates for the 32 clusters obtained under the bootstrap analysis over the genes are summarized in Table 1. To see how much the exclusion of a particular gene changes the estimates, we may look at the overall consistency of the inferred average outgoing and incoming gene flow using the complete data case as reference. It is observed that the exclusion of gene gdh seems to affect the admixture consistency most, as in this case the changes of outgoing (incoming) gene flow reach the maximum value in 20 (13) of the 32 clusters. In contrast none of the clusters is experiencing the largest gene flow changes when either the gene adk or aroE is excluded, suggesting that recombination signals on these two genes are more marginal. We plotted in Figure 8 the estimated densities of for the three N. lactamica populations (8, 29, 32), relative to N. meningitidis populations 11 and 19 separately. The densities for population 29 have a tendency towards zero, suggesting a close genetic affinity with populations 11 and 19. In contrast, the densities of populations 8 and 32 are much further away from zero, implying a distinctive difference in their genetic makeup compared to N. meningitidis populations 11 and 19. This result is consistent with the admixture pattern presented in Figure 7. The eBURST analysis of the Neisseria database resulted in 253 groups and 1165 singleton strains. The biggest group consists of 795 strains and there are additionally four groups containing more than 200 strains. Table 2 shows the degree of concordance between the eBURST groups and BAPS populations. Due to the very large number of eBURST groups, only groups containing at least 20 strains were included in this comparison. Table 2 shows that the largest eBURST groups harbour many strains from multiple BAPS populations, whereas the vast majority of strains in smaller groups are typically found only in a single BAPS population (in some cases in two populations). As these cases represent single-locus variants of one another from eBURST analysis being clustered into different populations by BAPS, it means that there must exist a very large amount of anomalous variation at the nucleotide level within the other locus to allow the model to identify such subgroups. It should also be kept in mind that the BAPS model used for the identification of these populations is not a phylogenetic method in contrast to eBURST, which is an important distinction particularly in the presence of highly recombinogenic data. Out of the three BAPS populations of N. lactamica strains (populations 8, 29 and 32), only one forms a group in the eBURST analysis (group 14, Table 2). Strains in the other populations are primarily assigned to singleton groups. This difference is further explored below using a phylogenetic analysis at the nucleotide level. To facilitate comparison of the phylogenetic analysis with the partition yielded by BAPS, we labelled strains with colors indicating population memberships. However, given the large number of strains included in the analysis and the large number of populations inferred by BAPS, it would be very challenging to visually extract information from a single NJ tree harbouring all the populations simultaneously. Therefore, four separate NJ trees are displayed in Figures 9–12, each of which shows a subset of the BAPS populations indicated with distinct colors. The strains remaining outside this particular subset are indicated by white circles. Since it is difficult to specify more than approximately 20 colors which remain clearly distinguishable from each other, independent coloring schemes were used for each tree to show the phylogenetic composition of the populations. Thus, it is not possible to compare the color codes directly with those in the gene flow network in Figure 7. The color coding scheme for the populations is shown in Figure 13 to enable comparison of the phylogenetic analysis and the gene flow network. The assignment of the populations to the NJ trees reveals that while a considerable number of them form relatively tight groups of lineages, there are also many populations in which the strains are spread over several separate lineages in the tree. This illustrates the dilution of phylogenetic signals in the presence of considerable levels of recombination between populations of strains. The population (population 14) which according to the inferred gene flow network is the most prominent recipient of genetic material from a multitude of sources, is seen (Figure 10) to include some dense and relatively large groups of strains that are found in separate parts of the tree. In addition, this population harbors a number of tiny groups of strains scattered over the three. In the BAPS analysis, strains identified as N. lactamica fell into three populations: 8, 29 and 32. Figure 7 indicates that we found no evidence for significant admixture involving populations 8 and 32. Population 29 however was found to be associated with variation characteristic of populations 11 and 19, which were composed of meningococcal strains. The positions of the STs composing these five BAPS populations and one other (8, 29, 32, 19, 11 and 20) are shown in Figure 7. The isolated status of population 8 is apparent as a well resolved group, whereas the recombinant status of 29 is clear from the way these STs are scattered around the tree with long branch lengths originating apparently separately from the main N. lactamica population. The role of meningococcal strains in populations 11 and 19 in this is evident, in that the recombinant N. lactamica strains (population 29, shown in red) apparently originate close to these populations in the main meningococcal radiation. Population 32 is intermediate on the tree between the majority of N. lactamica strains and the main meningococcal radiation. Hence these STs may be considered as examples of the so-called fuzzy fringes which have been proposed for recombinogenic species [25]. As noted however, they were not associated with significant admixture in the estimated gene flow network (Figure 7). Close examination of population 32 shows that 4 of the 22 STs in the population exhibited significant admixture with population 20 (shown in blue), receiving on average 12.3% from this population (which is composed of strains identified as meningococcus). It is interesting to note that populations 32 and 20 adjoin each other in the tree. In the present work we have introduced statistical tools implemented in the BAPS software that enable analyses of bacterial population structures on a previously unprecented scale, as the computational complexity of the earlier standard Bayesian methods prevents their application to large databases associated with complex patterns of admixture. This is particularly important when at least a moderate number of recombination events have plausibly taken place in a population, as a statistically valid characterization of the population structure then requires fairly extensive sampling of strains. It was also noted that a standard MCMC-based approach is not expected to yield a viable strategy for such analyses in practice, due to both the time constraints as well as the statistical accuracy of the resulting estimates. The BAPS analysis (inference about the populations and the levels of admixture) of the Neisseria database was completed within roughly 95 CPU hours on a standard PC with a 2.8 GHz Pentium 4 processor. As a comparison, our initial experiments with the STRUCTURE software [8] suggest conservatively that a comparable analysis had taken at least several thousands of CPU hours on the same machine. Moreover, the convergence problems associated with the Gibbs sampler algorithm, when applied to mixture models (e.g. [10]), suggest that statistically reliable estimates of the population structure are likely not accessible for data sets of such a high degree of complexity. The presented methods can be effectively utilized in a variety of contexts, where the genetic population structure is relevant, e.g. for the investigation of epidemiological questions and experimentally derived features of bacterial strains. For instance, outlying groups with specific characteristics with respect to virulence or antibiotic resistance may be detected from large population samples. The concept of a genetic shape, which was introduced here to represent the molecular variability of an identified population and its affinity towards other populations as a whole, is an intriguing characteristic associated with considerable potential for further theoretical research. Namely, the average change of log predictive likelihoods between populations can also be interpreted as the change in ‘free energy’ associated with a gene flow event. The larger this quantity, the more likely the ‘reaction’ of gene flow could occur spontaneously. However, it is not trivial to determine the minimal energy level that triggers such events. From the analysis of simulated data (results not shown) we are expecting that such an energy threshold depends on the identified population sizes. In particular, the analogy with physics-based characterizations of molecular interaction systems could yield mathematical ways to predict horizontal transfer events. Although this integrated approach advocated here provides a feasible means of handling data from thousands of strains and a multitude of genes, several issues remain. Firstly, if a very large number of genes are considered, it is likely that not all of them will be present or functional in overall in a heterogeneous population at a genus level. Under such circumstances it would be necessary to develop further the Bayesian model for a population structure and admixture to take into account that not all molecular information is shared by the sampled strains. Secondly, the scalability of the stochastic learning algorithms should be improved to ensure that models could still be fitted to data without access to supercomputing facilities. Given the present rate of improvement in sequencing facilities, it is likely that the need for such large-scale analyses will be a reality within a relatively short time-span. In order to meet these needs in the future, we are currently investigating several theoretical approaches to develop further the statistical population genetic tools available in BAPS. The findings from the combined phylogenetic and population genetic analyses suggest possible events of convergence between N. lactamica and N. meningitidis that have arisen on multiple occasions and have occurred for clearly separate lineages of the two species. As the former is a non-pathogen and N. meningitidis represents a pathogen of considerable importance in human health, exchanges of genetic material between them might have consequences for our understanding of their evolution. Moreover, the diversity and the extent of recombination indicated among the N. meningitidis populations highlight that it is necessary to consider these pathogens as a heterogeneous population, and that multiple pathways of evolution may arise among them as a response to treatment strategies, including antibiotics and vaccines, as also recently discussed in [31]. For details concerning the currently available Meningococcal vaccines, see, e.g. [32] and [33]. Contrary to some of the previous studies of recombination and population structure in Meningococci, e.g. [34]–[36]), where only very limited sample sizes were considered, we have here focused on the detailed exploration of a more extensive database using multiple model-based statistical tools. In summary, our combined results illustrate crisply the possibility of using large-scale MLST sequence data to draw attention to currents in the gene pool, i.e. specific populations that seem more likely to undergo recombination, including recombination with different species. More detailed exploration of such groups of strains could then shed new light on the mechanisms that shape the joint evolution of pathogens and non-pathogens sharing ecological niches.

10.1371/journal.pcbi.1000148

Coordinate Regulation of G Protein Signaling via Dynamic Interactions of Receptor and GAP

Signal output from receptor–G-protein–effector modules is a dynamic function of the nucleotide exchange activity of the receptor, the GTPase-accelerating activity of GTPase-activating proteins (GAPs), and their interactions. GAPs may inhibit steady-state signaling but may also accelerate deactivation upon removal of stimulus without significantly inhibiting output when the receptor is active. Further, some effectors (e.g., phospholipase C-β) are themselves GAPs, and it is unclear how such effectors can be stimulated by G proteins at the same time as they accelerate G protein deactivation. The multiple combinations of protein–protein associations and interacting regulatory effects that allow such complex behaviors in this system do not permit the usual simplifying assumptions of traditional enzyme kinetics and are uniquely subject to systems-level analysis. We developed a kinetic model for G protein signaling that permits analysis of both interactive and independent G protein binding and regulation by receptor and GAP. We evaluated parameters of the model (all forward and reverse rate constants) by global least-squares fitting to a diverse set of steady-state GTPase measurements in an m1 muscarinic receptor–Gq–phospholipase C-β1 module in which GTPase activities were varied by ∼104-fold. We provide multiple tests to validate the fitted parameter set, which is consistent with results from the few previous pre-steady-state kinetic measurements. Results indicate that (1) GAP potentiates the GDP/GTP exchange activity of the receptor, an activity never before reported; (2) exchange activity of the receptor is biased toward replacement of GDP by GTP; (3) receptor and GAP bind G protein with negative cooperativity when G protein is bound to either GTP or GDP, promoting rapid GAP binding and dissociation; (4) GAP indirectly stabilizes the continuous binding of receptor to G protein during steady-state GTPase hydrolysis, thus further enhancing receptor activity; and (5) receptor accelerates GDP/GTP exchange primarily by opening an otherwise closed nucleotide binding site on the G protein but has minimal effect on affinity (Kassoc = kassoc/kdissoc) of G protein for nucleotide. Model-based simulation explains how GAP activity can accelerate deactivation >10-fold upon removal of agonist but still allow high signal output while the receptor is active. Analysis of GTPase flux through distinct reaction pathways and consequent accumulation of specific GTPase cycle intermediates indicate that, in the presence of a GAP, the receptor remains bound to G protein throughout the GTPase cycle and that GAP binds primarily during the GTP-bound phase. The analysis explains these behaviors and relates them to the specific regulatory phenomena described above. The work also demonstrates the applicability of appropriately data-constrained system-level analysis to signaling networks of this scale.

Throughout the eukaryotes, G proteins convey information from receptors for diverse stimuli—neurotransmitters, hormones, light, odors, and pheromones—to intracellular regulatory proteins collectively known as effectors. G proteins function by transiting a dynamic cycle of activation and deactivation. Receptors accelerate activation, which allows G proteins to regulate effectors, and receptors thus increase signal output. GTPase-activating proteins, GAPs, accelerate deactivation. GAPs can thus attenuate signaling, but GAPs can also accelerate signal termination when stimulus is removed without inhibiting signal output while stimulus is present. Surprisingly, some effectors are also GAPs for the G proteins that activate them, essentially turning off their activator. We developed a mathematical model that describes control of G protein signaling by receptor and GAP and used experimental data to determine its important parameters. We show that GAPs actually potentiate G protein activation by receptor, a previously unsuspected effect. Further, GAPs indirectly stabilize receptor–G protein binding during stimulation, which we had previously predicted based on inconsistencies among other experimental results. The present results elucidate how GAPs can independently control signaling kinetics and amplitude and thus clarify how effectors can both respond to G proteins and act as G protein GAPs.

G protein-mediated signaling modules display a variety of dynamic input-output behaviors despite their use of a single, relatively simple biochemical mechanism. Signal amplification, the ratio of effector proteins activated to agonist-bound receptors, can vary from unity to hundreds. Activating ligands may bind receptors with affinities ranging from picomolar through millimolar. GAPs, which can accelerate hydrolysis of bound GTP over 2000-fold, can accelerate both activation and deactivation in cells with variable inhibitory effect [1]. Activation and deactivation rates upon addition and removal of agonist can thus range from ∼10 ms to minutes. Heterotrimeric G proteins convey signals by traversing a cycle of GTP binding and hydrolysis: the GTP bound state of the Gα subunit is active and deactivation is caused by hydrolysis of bound GTP to GDP [2]. The rates of activation and deactivation, and consequent effects on signal output at steady state, are regulated by interactions of the Gα subunit with receptors [3], Gβγ subunits [4], GTPase-activating proteins (GAPs) [1] and multiple other proteins [5]. The net effect of these inputs depends on the identities of the individual proteins, their concentrations and their own regulatory controls. Regulatory inputs to G protein modules are interactive, and it has been difficult to establish quantitative understanding of how they cooperate to control signal output. While some signals, particularly G protein-gated channels, can be monitored accurately in cells in real time, it has been harder to quantitate the intermediary reactions of the GTPase cycle and protein–protein binding or dissociation. Recently developed optical sensors are promising [6]–[10] but still do not provide complete or simultaneous coverage of multiple events and often do not provide absolute (i.e., molar) data. Conversely, quantitative biochemical assays using in vitro reconstituted systems have provided absolute biochemical data [11],[12] but have not adequately described the simultaneous regulatory interactions that are so important. Consequently, quantitative understanding of the dynamic behavior of an intact G protein module remains elusive. Computational modeling is used frequently to clarify mechanistic thinking about complex biochemical systems, including G protein signaling. Quantitative models can potentially combine information on individual reactions to simulate the behavior of a complex system, or use system-level data to test the validity of a proposed mechanism. The work of Linderman and colleagues, for example, has provided consistent examples of these approaches to G protein signaling [13]–[16]. The G protein-mediated yeast pheromone response has also been the focus of significant modeling efforts because of its presumed paucity of off-pathway inputs [17]–[19]. In at least one case, the failure of a simple model of this pathway motivated discovery of a novel mechanism for feedback regulation and subsequent refinement of the model [17]. However, modeling of G protein modules has often been descriptive, with parameters arbitrarily chosen for a few reactions such that model output mimics an experimental result. Alternatively, the inner workings of the G protein module itself have been condensed into an arbitrary function of agonist concentration and receptor regulation to allow analysis of a downstream event such as Ca2+ release or protein phosphorylation or, even more distal, transcription. A major problem in developing quantitative models of G protein modules has been accurate assignment of parameters to the many processes that are known to occur. These include both the GTPase cycle reactions and the multiple protein-protein interactions that govern these reactions. This problem is significant because local protein concentrations at the plasma membrane and the regulated association of these proteins are both unknown, either for resting cells or during dynamic signaling. In this study, we have used purified proteins, heterotrimeric Gq, m1 muscarinic acetylcholine receptors and phospholipase C-β1, reconstituted at uniform and controllable concentrations into unilamellar phospholipid vesicles, to overcome this first limitation. We estimated formation of multi-protein complexes according to their individual activities. The second major problem in modeling signaling through G protein modules is the difficulty in assigning correct, or even plausible, values of rate or equilibrium constants for the reactions included in the model. Despite their apparently small size, an informative model of a single G protein module will contain multiple parameters that are not readily accessible from individual measurements. These parameters may vary widely among different modules (receptors, G proteins, GAPs), which prohibits most literature-mining approaches. If all or most of the relevant parameters are not individually available for the module of interest, then an adequately large and diverse dataset must be produced to allow parameters to be fit to the data. Last, even with a presumably adequate dataset, the numerical fitting process that extracts values for the parameters and subsequent validation of the fit are both central problems in modeling signaling systems. We have adapted and extended several approaches to deal with the difficulty of fitting a model with a fairly large number of parameters using a modest amount of data. We present a modestly complex model of signal output in a G protein model that contains many of the salient regulatory interactions that characterize G protein signaling. We used steady-state GTPase data to support a Metropolis-Monte Carlo fitting strategy, and argue that most parameters are reasonably assigned, with statistical data to help qualify fits for individual parameters. The resultant parameter set shows that receptor accelerates both GDP dissociation and GTP binding, and that GAPs potentiate the receptor's nucleotide exchange catalyst activity. Further, the model argues strongly that GAP activity indirectly favors continued binding of receptor to G protein throughout the GTPase cycle, thus further potentiating the receptor's activity. Such indirect stabilization of receptor-G protein binding, referred to as kinetic scaffolding to distinguish it from direct interaction, was suggested as a mechanism for how a GAP can accelerate deactivation upon removal of agonist without substantially inhibiting signaling [1],[11],[16],[20]. Model-based simulation of signal output describes how GAPs combine these mechanisms to independently control signal amplitude and kinetics. The biochemical model of the GTPase catalytic cycle (Figure 1) includes GTP binding, hydrolysis of bound GTP and simultaneous release of inorganic phosphate (Pi), and the dissociation of GDP. At each stage of the reaction, G protein is allowed to bind agonist-liganded receptor, GAP or both. Receptor is assumed to be agonist bound and active at all times; agonist-stimulated GTPase data were obtained in the presence of saturating carbamylcholine (1 mM). Possible dissociation of Gβγ from Gα and protein oligomerization were not included (see Discussion). All reactions were considered to be reversible to allow imposition of path-independence constraints on closed reaction loops during the fitting process (see below). For the same reason, even presumably unlikely reaction paths were retained to create symmetry in the reaction map. For calculation of G protein activation (see below), all GTP-bound species were considered to be equally active, and fractional activation was calculated as the fraction of all species that contain bound GTP. The kinetic model for G protein signaling (Figure 1) includes 48 parameters, first- and second-order rate constants, only a few of which have been determined directly. We therefore fit all the parameters to a relatively large and diverse set of steady-state GTPase rates determined in a purified and reconstituted system in which protein concentrations were known and where data could be obtained over a wide dynamic range. Data for fitting came from 8 scans of GTPase activity as the concentration of one assay component, GTP, GDP or GAP, was varied from zero to saturation in the presence or absence of saturating agonist (Table S1; Figure 2). Data were fit simultaneously to minimize the cost function, defined as the sum of the squares of deviations between experimental data and data predicted by the model (Materials and Methods). Values for the 48 kinetic parameters were adjusted simultaneously by constrained simulated annealing to best match all available data while satisfying thermodynamic constraints (path independence, i.e. cyclicΔG = 0, for all potential cycles; and net ΔGhydrol for GTP [21]). The progress of cost minimization for a typical fitting run is shown in Figure S2. The cost function is initially quite high (off-scale in Figure S2) and decreases rapidly. The initial decrease is followed by relatively quick adjustments of the parameters interspersed with long metastable stages, reflecting occasional escape of the Monte Carlo search from local minima in the cost manifold. Improvement in the fit is negligible past a few thousand iterations. To further test the adequacy of the Monte Carlo search, it was repeated with thermodynamic constraints applied as a quantitative penalty for nonconformance in the cost function rather than as an absolute constraint (Materials and Methods) (Figure S3). In this case, initial convergence was slower, but subsequent enforcement of strict thermodynamic constraints decreased the value of the cost function to a level similar to that achieved if thermodynamic constraints are applied throughout the fitting process. Because this more ergodic search method did not lead to lower values of the cost function, it is likely that imposing path-independence constraints initially does not seriously limit the ergodicity of the fitting process. The initial test of such a modeling process is the ability of the model to simulate experimental data using the parameter set determined by fitting (Figure 2). Simulations based on the model and parameters derived from 41 fitting runs (Table S2) approximated the experimental data well over a 105-fold range of GTPase activities and a wide variety of experimental conditions. Values of Km for GTP, Ki for GDP and EC50 for the GAP activity of PLC-β1 were all matched closely in each experiment. Relative increases and decreases in activity were also simulated well, as were curve shape and steepness. The largest recurrent discrepancy between data and prediction was in the absolute value of the maximal activity. Disagreement was negligible in some experiments, but was significant in others. In part, this reflects real difficulty in fitting such a diverse dataset, but it also arises from variation in specific activity among the experiments. The data were obtained using several preparations of m1AChR-Gq vesicles that varied in maximum specific activities, with standard deviation of ∼40% among 13 batches of vesicles prepared during the study. Variation between fits and data in Figure 2 are within this margin. The values of the rate constants obtained by fitting the steady-state rates also compare well with those few that have previously been determined directly in pre-steady-state kinetic measurements [12] (Figure 3). For five reactions, nucleotide association and dissociation and GTP hydrolysis, agreement was within a factor of 4. The direct determinations were performed with different preparations of vesicles and by different investigators. Agreement is thus even more striking. Importantly, the pre-steady-state kinetic data were not used in the present fit. The rate constants obtained here also compare well with predictions from data obtained in non-identical preparations (detergent-solubilized proteins, free Gαq subunits, etc.) [11],[12],[22],[23]. Fitting is a stochastic process that, upon repeat, converges to different minima of comparable cost in a complex manifold. For these datasets, multiple fitting runs yielded a family of parameter sets with cost functions in the range 650–800 (not shown). The extent of variation among repeated fits reflects the size of the error on each parameter (Figure 3). For some of the parameters, reproducibility was excellent, but for others error was large. Error may reflect the absence of necessary data or experimental error, but an additional difficulty in fitting some parameters arises from the structure of the model. To allow imposition of path-independence constraints, the model contains all possible interactions of proteins and nucleotides, including species that are quantitatively negligible and reactions that do not contribute detectably to flux through the GTPase cycle. Thus, some individual rate constants cannot be fit well, and some pairs of forward and reverse rate constants that describe rapid equilibria are poorly fit because the data only constrain their ratios. To evaluate possible sources of errors associated with some of the parameters, we repeated the fitting process with synthetic data and asked whether the fitting process could accurately return the parameters used in the synthesis. Simulated data equivalent to the original experimental data were generated using the model and a chosen parameter set. To simulate experimental noise, Gaussian errors (standard deviation/mean = 10%) were convoluted with the predictions. The parameter set returned in this process simulated the synthetic data extremely well, and did not show the significant errors in maximal velocity observed when the real data were fitted (not shown). The parameter set obtained by fitting synthetic data was then compared with the set used in its generation (Figure 4A). The histogram shows that 32 of the 48 constants were fit to within 10-fold of the generating value, with 19 within 2-fold. Examination of the outliers indicates that they describe reactions that either are not appreciably populated or are much faster than the reaction that they precede, and therefore could not be constrained. The fitting process is thus adequate to determine most parameters well, and those that are not well fit do not contribute appreciably to overall flux through the GTPase cycle. To see whether rapid equilibria contribute to error in evaluating individual kinetic constants, we also compared the fitted equilibrium constants for each reaction (i.e., the ratios of forward and reverse rate constants) with the values used to generate the synthetic data (Figure 4B). Deviations from the generating values were fewer and smaller, indicating that equilibrium constants were constrained by the thermodynamic relationships used to construct the model. The quality of the fit was further assessed by thermal ensemble analysis [24] (Text S2). The analysis consists of generating statistically equivalent fits to the data and measures the extent to which parameters are coupled (Text S2). We found lack of generalized mixing suggesting (1) a reasonable match between the model and the underlying phenomena, (2) the absence of severe over- or under-parameterization, and (3) the availability of sufficient data for accurate determination of many of the parameters. The parameter set shown in Figure 3 and Table S2 provides the first reasonably complete set of experimentally determined rate constants for a G protein signaling module, and thus provides insights into regulatory interactions that were not previously accessible. While the parameters are interpretable only to within the errors of the fit, several novel observations stand out at this level. First, examination of the rates of nucleotide binding and release indicate that the salient function of receptor is to open an otherwise inaccessible (“closed”) nucleotide binding site on Gq to permit GDP/GTP exchange. In addition to accelerating GDP dissociation, receptor also markedly accelerates both GDP and GTP association (Table 1). Receptor thus promotes GDP/GTP exchange by two distinct mechanisms. It accelerates GDP dissociation over 104-fold and GTP association more than 103-fold. Receptors have been thought to act by binding G protein negatively cooperatively with respect to nucleotides; i.e., that receptor decreases affinity for GDP by increasing the dissociation rate (Kassoc = kassoc/kdiss). In the case of the M1 muscarinic receptor and Gq, the decrease in affinity for GDP (∼3-fold) is dwarfed by acceleration of GDP dissociation (∼20,000-fold; because GDP binding to the open site is also fast). Opening and closing of the nucleotide binding site is also reflected in the remarkably slow nucleotide association rates observed in the absence of receptor. The slow basal association rate constant for GTP, ∼500 M−1·s−1, is particularly striking, but all GDP and GTP association rate constants are less than 104 M−1·s−1 without receptor stimulation. Receptor increases the association rates about 104-fold to 106–107 M–1·s−1, values that are more commonly observed for binding of small ligands to proteins. Taken together with the slow rates of spontaneous nucleotide dissociation, the slow association rates indicate that the nucleotide binding site on Gq is essentially closed in the absence of receptor and that receptor stabilizes the open conformation regardless of whether GTP, GDP or no nucleotide is bound (see Discussion). Second, comparison of the rate constants for nucleotide exchange shows that GAP potentiates the ability of the receptor to accelerate the dissociation of bound GDP by about 20-fold (Table 2). Thus, even though GAP has negligible effect on GDP binding by itself, its facilitation of GDP/GTP exchange helps minimize potential inhibition of signaling during stimulation by receptor. GAPs were not previously known to modulate GDP binding [1],[25], but this effect was probably overlooked because GAPs do not bind tightly to GDP-bound G protein; the RGAD complex will only be formed during steady-state GTPase turnover. GAP displays little effect on the rate of GTP dissociation because the binding of GAP and GTP to G protein is positively cooperative [1]. The parameter set also indicates that receptor and GAP bind G protein negatively cooperatively, and that cooperativity depend on the binding of GDP or GTP (Table 3). Receptor and GAP reciprocally decrease the affinity of Gq for each other by 25-fold when GTP is bound and by ∼120-fold when GDP is bound, but there is essentially no cooperativity displayed for binding to nucleotide-free Gq. The most striking result of this interaction is the rapid dissociation of GAP from the receptor-Gq–GDP complex after GTP is hydrolyzed. The t1/2 for GAP dissociation is about 300 ms, about 90-fold faster than in the absence of receptor (Table S2). In contrast, GAP dissociation from GTP-bound Gq is slow, about 170-fold slower than hydrolysis, such that essentially every GAP binding event results in GTP hydrolysis. In summary, GAP dissociates virtually immediately after GTP hydrolysis during receptor-mediated signaling, and is thus potentially available to accelerate hydrolysis on other G proteins. The nucleotide-dependent, negatively cooperative binding of receptor and GAP to G protein also helps determine the reaction pathway through the GTPase cycle: what intermediate species are populated and for how long (Figures 5 and 6; see below). For example, GTP accelerates the dissociation of receptor from G protein by ∼70-fold whereas GDP has a much smaller effect. This difference further biases receptor-promoted GDP/GTP exchange toward the forward (activating) direction. Qualitatively, destabilization of receptor binding by nucleotides confirms the observation that nucleotides drive dissociation of receptor from G protein [26]. To examine the overall regulatory behavior of the G protein module, we used the complete reaction model and average fitted parameter set to simulate signal output as the fraction Z of all G protein complexes to which GTP is bound. Figure 5A shows a contour plot of fractional activation at steady-state as a function of varying concentrations of receptor and GAP, using typical in vitro assay conditions to allow us to compare prediction with experiment (300 nM GTP, 10 pM GDP, no Pi). At low concentrations of active receptor, signal output is predictably low regardless of GAP concentration. In the absence of GAP (bottom of figure), increasing the concentration of receptor raises Z to about 93% activation. At saturating concentrations of GAP (top of figure), Z increases with increasing concentrations of active receptor to about 4% of maximal activation. This limiting value reflects the ratio of the rates of GTP hydrolysis to GDP/GTP exchange when GAP and receptor are both bound to G protein throughout the catalytic cycle. At high receptor concentration (right side), increasing concentrations of GAP causes Z to fall from 85% to 12%. These transitions are relatively smooth, although slopes are asymmetric and steeper than predicted by a Hill coefficient of 1. The values of Z at the corners agree with analytical calculations, which are only possible at these limits. While the precise output obviously depends on the values of the rate constants, the overall topography of this plot had sufficient similarity among fitted parameter sets to indicate that errors in the fit do not modify the essential behavior of the model. The most striking feature of the Z contour plot lies in the region where the concentrations of G protein, receptor and GAP are similar. Here, Z contour lines are contorted and create an abrupt transition, a “ridge” at which activity peaks and then declines with increasing concentration of receptor. In a few locations, increasing the concentration of receptor causes Z to decrease, and in a tiny region, increasing the concentration of GAP actually increases Z. This unintuitive topography is not idiosyncratic to the average parameter set, but appears in various shapes for all the parameter sets obtained with repeats of the fitting procedure. To clarify the origin of this behavior, we calculated the fluxes and steady-state concentrations of intermediates at locations on either side of the ridge to determine what reactions and molecular species contribute to Z near the ridge (Figure 5C; see Figure S5 and Figure S6 for examples). To the left of the ridge, the major reaction path is RG→RGT→GT→GD→RGD→RG. GT is the major activated species. The receptor dissociates upon GTP binding and reassociates after hydrolysis, the mechanism referred to as collisional coupling [27]. GAP is not significantly involved in the reaction scheme and Z is low. Figure 5B indicates that the major active species is GT in this region. Across the ridge, the reaction pathway becomes a comparable mixture of RG→RGT→RGD→RG and RG→RGT→RGAT→RGAD→RGD→RG. Species RGT is the major active species (Figure 5A). Receptor remains bound throughout the GTPase cycle, and significantly, GAP is recruited to the receptor–G protein complex during the GTP-bound phase (Table 1). Z has a higher value despite involvement of the GAP in net GTPase turnover. The ridge thus reflects the coincidence of the peak in the concentration of GT in a region where the concentration of RGT is increasing significantly (Figure 5B). The change in pathway is governed by choice of the reaction that follows the branch-point species RGT (Figure 5A and 5C). With increasing concentration of receptor, net flux switches from RGT→GT to RGT→RGAT and RGT→RGD as the concentration of receptor crosses the ridge. The peak in activity reflects the transient accumulation of GT as the concentration of free R increases and drives GDP/GTP exchange but before it reaches the level at which GAP is recruited. Above the Z ridge, flux through the GTPase cycle is maintained entirely by complexes that include receptor; i.e., where receptor remains bound throughout the catalytic cycle. The occurrence of a ridge in Z with increasing receptor concentration, rather than a monotonic increase, is caused by the negatively cooperative binding of receptor and GAP to G protein (described above). The importance of this mechanism is indicated by the location of the ridge in the R-A plane. It lies just to the left of the line [A]tot = [G]tot−[R]tot, where the sum of the concentrations of total receptor and total GAP equals the concentration of total G protein. This straight line appears as a curve on log–log plots (Figure 5A). Negatively cooperative binding of receptor and GAP to G protein make accumulation of RG and GA species far more likely than accumulation of RGA species and thus causes the abrupt shift of pathway and consequent peak in G protein activation. The crest of Z is displaced from the line because the GTPase cycle is not at equilibrium under steady-state reaction condition. We also used the model and parameter set to simulate G protein activation under typical cytoplasmic conditions—0.2 mM GTP, 0.02 mM GDP, 1 mM Pi [28] (Figure 6). Activation of Gq responds to receptor and GAP in a pattern generally similar to that seen under laboratory assay conditions, but the higher cytoplasmic concentration of GTP allows substantial activation by receptor at high GAP concentrations. Signal output is thus significant, Z∼0.25, even in the presence of saturating GAP. Output remains high in the presence of GAP because GTPase flux is almost entirely from the RGA–>RGAT–>RGAD–>RGA pathway over a large part of the R-A plane (Figure S6, Figure S7, and Text S3). Given this pathway, high values of Z result in part from the GAP's potentiation of receptor-promoted GDP release (Table S2). GAP exerts this effect under cytoplasmic conditions because, at 0.2 mM GTP, nucleotide-free G protein binds GTP quickly (t1/2<2 ms) and because GAP does not dissociate appreciably. Equally important, receptor remains bound because GTP is hydrolyzed rapidly, before appreciable receptor can dissociate, and therefore catalyzes GDP/GTP exchange promptly after hydrolysis. The principal potentiating effect of cytoplasmic GTP concentration is thus to support continued association of receptor, GAP and G protein during the GTPase cycle. A novel and unintuitive result of this simulation is the decline and subsequent increase in Z at high receptor concentrations as the concentration of GAP is increased. As shown in Figure 6, Z is minimal at about 10−4 M GAP and increases at higher GAP concentrations. This effect is not predicted for lower concentrations of GTP and is relatively small for the conditions and parameters used here. The occurrence and extent of this behavior depends sensitively on multiple rate constants, as do the relative plateau values of Z at high and low GAP concentration. In general, the ability of GAP to increase fractional G protein activation at high concentrations depends on its potentiation of the receptor's exchange catalyst activity and its indirect stabilization of receptor binding to G protein, as discussed above. Its mechanism is discussed in the Text S3. In cells, GAP activity often accelerates signal termination when agonist is removed but does not inhibit signaling significantly while agonist is present [1]. To determine whether this behavior is accurately predicted by the present model and to study its mechanism, we simulated signal termination upon removal of a rapidly dissociating agonist by first allowing the system to reach steady state and then instantaneously setting the concentration of activated receptor to zero (Materials and Methods). We first scanned the receptor and GAP concentrations shown in Figure 6 for regions where increasing the GAP concentration causes minimal inhibition but significantly accelerates signal termination. Quantitative search criteria were chosen to mimic published experiments (reviewed in [1]; see legend to Figure 7), but their exact values are not crucial (results not shown). As shown in the inset to Figure 7, addition of GAP can accelerate deactivation with minimal steady-state inhibition at all concentrations of active receptor. A wide range of initial and final GAP concentrations also meet the initial criteria. This behavior is thus robust to initial conditions. Within this region, addition of GAP can accelerate signal termination up to 180-fold, which actually exceeds the acceleration that has been observed in cells. Figure 7 shows the deactivation time course for a representative simulation that compares signal termination at high and low concentrations of GAP, shown as red dots in the inset. The higher GAP concentration accelerated Gq deactivation more than 15-fold, measured as time to 50% of initial activity, but inhibited receptor-stimulated G protein activation by only 5%. Qualitatively similar behaviors are observed over much of the area of Figure 6, indicating that fast termination combined with minimal inhibition is a common outcome of G protein GAP activity. Neither termination time course in Figure 7 is monoexponential, and complete deactivation is markedly delayed at the lower GAP concentration (right inset). Some GAP activity thus appears to be required for reasonably fast decay of signal output to basal levels. Simulations with intermediate GAP concentrations (not shown) indicate that GAPs can also facilitate return to basal activation without accelerating signal termination to the extent shown in Figure 7, and a variety of termination behaviors can be observed at different points on this activation surface. While multiphasic decay of G protein signals has also been observed experimentally, we do not know whether the separate phases in Figure 7 correspond to specific cellular turn-off events. Flux analysis of the deactivation events indicates that there is a single mechanism for accelerated signal termination by GAPs. At low GAP concentrations, the species RGT and RGAT both contribute significantly to activity in the presence of activated receptor. Upon removal of receptor, GT and GAT are rapidly created. GAT is then rapidly deactivated at a rate of 8.6 s−1 (p21 in Table S2), the initial phase of deactivation. The second, very slow phase is deactivation of GT. In contrast, at higher GAP concentrations almost all G protein activity is due to RGAT. When activated receptor is removed, the GAT that is formed hydrolyzes rapidly to cause fast deactivation. While deactivation is not really monophasic even at fairly high GAP concentrations, slow hydrolysis of GT is not significant because there is not much of it and because the GAP that dissociates from the GAD hydrolysis product binds remaining GT to accelerate its deactivation. In this way, GAP provides a pathway for fast signal turn-off without inhibiting signaling. A mechanistic model of signal transduction should provide quantitative understanding of how time-dependent outputs arise from the underlying binding, conformational and chemical reactions. This study attempts to address three unresolved mechanistic questions in G protein signaling. First, what are the underlying dynamics of the GTPase catalytic cycle that integrate the regulatory activities of receptors and GAPs, their reversible binding to the G protein, and their control of G protein activation? Which effects are important and what functions do they serve? Next, how can a GAP accelerate signal turn-off when agonist is removed, yet not inhibit activation while agonist is present? Both these questions are vital to understanding how G protein-regulated effectors such as phospholipase C-β and p115RhoGEF can act as GAPs for their G protein activators without blocking their own activation. Last, can we use a data-constrained model to quantitate the interactions and activities of multiple interacting proteins during steady-state signaling where one-by-one measurements are not feasible? Quantitative modeling and simulation can provide this kind of understanding, but only if the underlying physical model is adequate and if the parameters of the mathematical model are objectively derived from experimental data. Even a relatively small G protein signaling module is a complex, non-linear system in which reaction pathways and modes of regulation may be both unintuitive and resistant to the simplifying assumptions of classical enzyme kinetics. We used a thermodynamically complete model, in which all reactions are reversible and all states are connected (Figure 1). Such a model assures that relevant reactions are not omitted, assures compliance with the laws of thermodynamics and uses detailed balance to help constrain parameters during the fitting process. The present version of the model does omit two relevant reactions. First, the concentration of agonist-bound active receptor is used as a surrogate for the agonist-induced activation of a fixed number of receptors. This simplification precludes some pharmacological inferences, but no currently available mechanism quantitatively and accurately relates agonist binding, receptor activation and G protein regulation [29],[30]. Second, we omitted activation-induced dissociation of Gαq and Gβγ. G protein subunits can dissociate in detergent solution [4], but physical dissociation in membranes is not universal [6],[31]. The binding of Gαq to Gβγ in detergent solution suggests that dissociation is slower than the reactions studied here [11],[32], and preliminary data on fluorescence resonance energy transfer between Gαq and Gβγ in phospholipid vesicles indicate that binding is relatively tight even for GTPγS-activated Gq (C. Hoang and E.M. Ross, unpublished). Thus, while certain behaviors determined here for Gq may reflect actions of both Gαq and Gβγ subunits, kinetically significant dissociation is probably not an important factor. We also did not consider any direct effects of Gβγ on the actions of receptor or GAP because they are subsumed in the rate constants for the reactions of these multi-protein species. For example, it is plausible that Gβγ contributes to the stable association of receptor with GTP-bound Gα during rapid GTPase turnover, but we have no independent evidence for this effect. Values for the rate constants for the model were derived from fits to steady-state GTPase data obtained with known concentrations of proteins, over widely varied concentrations of GAP, GTP and GDP, and in the presence or absence of agonist. Activities and ligand concentrations spanned several orders of magnitude. Such a dataset is appropriate for parameterizing a model of this complexity because steady-state activities encompass all the simultaneous reactions that modulate flux through the catalytic cycle, including those that cannot be measured individually. Indeed, most of the parameters could not have been determined by individual rate measurements regardless of desired accuracy or precision. We did not include pre-steady-state kinetic data in the fitting process, but individual rate constants that were previously directly determined in quenched flow experiments [12] agree well with those obtained here (Figure 3). The Metropolis-Monte Carlo fitting procedure yields a family of parameter sets that, with repetition, provides mean parameter values with quantitative statistical measures of accuracy. Most of the parameters also passed two other tests for validity: they were reproduced well in multiple fits to data (Figure 2) and, in fits to synthetic data, the fitted values reproduced the target values well (Figure 4). Further, thermal ensemble analysis indicated that the model was not significantly over- or under-parameterized (Figure S4 and Text S2). Thus, the data were sufficient in quality, quantity and diversity to produce reliable values for most of the rate constants. While the error windows on several of the parameters are larger than what would be expected from typical pre-steady-state measurement of a single enzymatic reaction rate, many are excellent even by traditional standards. The analysis also points out what parameters were not fit well, which prevents overinterpretation. For many of the poorly fit parameters, the chemical reactions do not take place to a significant extent, and their rates therefore do not contribute appreciably to steady-state GTPase activity or to G protein activation. Thus, they do not impact on interpretation of reaction rates or allosteric interactions, nor do they invalidate model-based simulations. Comparison of this parameter set with that of Bornheimer et al. shows several disagreements in values of reasonably well fit parameters for GTP and GDP binding in addition to expected disagreement with poorly fit values. Several are important for interpretation of allosteric interactions. Those authors chose their parameter set based on previously published pre-steady-state data from this laboratory, but did not fit them to a suitably diverse dataset. A significant value of the present fitting strategy is that it provides statistical descriptions of the reliability of individual rate constants, such that conclusions can be quantitatively evaluated. Having the complete set of rate constants allows simulation of signaling behavior with verifiable limits of accuracy. This systems level kinetic analysis of Gq signaling provides three distinct but interrelated sets of mechanistic information. First, the fitting process provided values for previously inaccessible kinetic parameters and thus revealed novel cooperative interactions among receptor, G protein, GAP and nucleotides. Second, model-based simulation demonstrated how paths through the GTPase cycle vary with the concentrations and activities of the individual proteins. Third, these analyses combine to allow description of regimes where GAPs can facilitate rapid signal termination upon removal of agonist without substantially inhibiting signaling. Because many of the important rate constants that describe the G protein signaling module were reasonably well determined by the fits to experimental data, this study identified several new regulatory interactions that control the rate and extent of G protein activation. A major finding was that GAP potentiates the GDP/GTP exchange catalyst activity of the receptor (Table 2). GAP both accelerates GDP dissociation from the receptor-G protein complex and inhibits GDP rebinding, decreasing equilibrium affinity for GDP more than 200-fold. This effect of GAP contributes significantly to its ability to accelerate GTP hydrolysis without proportionately decreasing steady-state G protein activation by receptor. This effect could not be determined directly by standard pre-steady-state kinetics methods because it impacts only transient, low-affinity intermediates in the GTPase cycle. GAP had no significant effect on GDP binding in the absence of receptor, consistent with previous data [1], and had no significant effect on GTP binding to the receptor-G protein complex, although it increased the affinity for GTP of free G protein about 25-fold. This increase is consistent with the ability of GTP analogs to increase the affinity of G protein for GAPs [1]. Note that Gβγ contributes to the kinetics of nucleotide binding to Gα subunits and is intimately involved in its regulation by receptors [4] and GAPs [1],[23]. Our data do not distinguish the contributions of the individual subunits to the regulation of Gq, but the net effects should represent the normal responses of intact G proteins in a biological membrane. A second novel finding is that receptor significantly accelerates nucleotide binding to G protein in addition to promoting dissociation (Table 1). Fast GTP binding at cytosolic concentrations is crucial for maintaining high steady-state G protein activation (Figure 6). Acceleration of nucleotide binding also clarifies the mechanism of receptor-mediated nucleotide exchange. The receptor-promoted increase in the equilibrium Kd is much smaller than the increases in kassoc and kdissoc for both GTP and GDP (Table 1). The receptor acts thus primarily to open the nucleotide binding site, presumably by moving the switch regions away from the entrance, but does not drastically distort the binding site itself. Such movement is demanded by the structure of the Gα subunit because bound nucleotide is essentially covered by a protein lid in the closed conformation [33]. The intrinsic high affinity of G protein for GDP that derives from the covered site is crucial to maintain low basal activation in the absence of agonist-bound receptor. The site-opening mechanism described here allows the receptor to act as a highly efficient GDP/GTP exchange catalyst while maintaining adequate affinity of receptor for the nucleotide-bound forms of the G protein. The idea that receptor opens the GTP binding site on G proteins actually dates to early studies of the GTPase cycle [34], but few studies have indicated that receptor actually increases kassoc [35]–[37]. In contrast, the prototypical GTP-binding protein Ef-Tu is regulated primarily by negatively cooperative binding of the exchange factor Ef-Ts [38], and this is true for several other monomeric GTP-binding proteins and their exchange factors (GEFs) [39]–[41]. For these proteins, GDP dissociation is the primary regulated step and the increase in kdissoc is roughly proportional to the increase in the equilibrium Kd; effects on kassoc are minimal. Negative cooperativity, defined as the reciprocal decrease in the equilibrium affinity of G protein for nucleotide and receptor when the other is present, is less significant for heterotrimeric G proteins than the ability of receptor to open the nucleotide binding site. Given the need for a low basal exchange rate, a purely negatively cooperative interaction with receptor would require a huge increase in Kd for GDP to allow receptor to promote physiologically fast exchange. The reciprocal effect on the Kd for receptor at physiological nucleotide concentrations would also compromise the stability of receptor binding. Heterotrimeric G proteins have thus evolved to use the lid of the binding site to allow low basal exchange without putting an energetically impractical demand on cooperative interaction with receptor. The negative cooperative binding of receptor and GAP to Gq was also unexpected. This interaction could not readily be detected by conventional binding measurements because of the low affinity of GAPs for the GDP-bound form of G proteins (where negative cooperativity is greatest; see Table 3). It should now be possible to test this interaction directly using the parameter values found here to guide experimental design. Note that the reaction model (Figure 1) does not demand any direct or indirect interaction between receptor and GAP, and their negatively cooperative binding was shown by fitting to experimental data. The importance of this interaction is not intuitive, but it underlies the shape of the activation surfaces shown in Figures 5 and 6. Such a surface was also predicted by Bornheimer et al. [42], who based their model on data from this laboratory. Kinzer-Ursem and Linderman [43] also described a biphasic effect of receptor based on sensitivity analysis of a model that focused on receptor function without consideration of GAP. Our analysis indicates that the ridge of maximal activation approximates the line at which the total concentrations of receptor plus GAP equal that of G protein, and this prediction can now be used to analyze other systems where these concentrations vary. Interaction between receptor and GAP also largely dictates the pathways of intermediary reactions through the GTPase cycle as functions of the concentrations of receptor and GAP, and thus contribute to the transient kinetics displayed when agonist is either added or removed. Simulations based on the parameterized model suggest mechanisms for how GAP activity promotes fast deactivation when agonist is removed without attenuating the signal while agonist is present. Receptor-generated signal output at steady-state can be significant over a wide range of GAP concentrations sufficient to accelerate signal turn-off (Figure 7). Such apparently paradoxical behavior is often observed for G protein-gated ion channels, whose cellular activation and deactivation kinetics can be studied directly [44],[45], reviewed in [1]. A major reason that a GAP can exert these two functions is its potentiation of the exchange-catalyst activity of the receptor, which is apparent by examining the rate constants that govern the GTPase cycle (Table 2). A second mechanism, which is evident only upon examining GTPase fluxes under the appropriate conditions, is that the GAP's multiple activities shift the path through the GTPase cycle such that receptor largely remains bound to G protein throughout the catalytic cycle and thus obviates the relatively slow step of reassociation with GDP-bound G protein after hydrolysis (Figure S5, Figure S6, and Figure S7). Thus receptor can initiate GDP/GTP exchange immediately after hydrolysis. Several properties of the GTPase reaction contribute to this effect, but it primarily results from the simple fact that GAP-stimulated GTP hydrolysis is faster than the rate of dissociation of receptor from the GTP-activated G protein. Because GDP dissociates faster than receptor, GDP dissociation occurs first and is followed by rapid GTP binding because the receptor maintains the nucleotide binding site in the open configuration. We refer to this mechanism as “kinetic scaffolding”, the ability of the GAP to promote long-term receptor binding by accelerating alternative reactions. We proposed this phenomenon previously [1],[11],[20], although we assumed that GAP also remains bound. The present analysis suggests that GAP binding to receptor-G-GDP is in rapid equilibrium, with dissociation likely to occur during each pass through the GTPase cycle. Because the affinity of G protein for GAP is poorly determined by these data (Figure 3), real quantitation of GAP binding is imprecise at best. Receptor binding is also not defined precisely in the fits to the present dataset, but examination of activation contours of the sort shown in Figure 6 show similar, although hardly identical, patterns when based on each of the 41 fitted parameter sets. The overall pattern of transit through the GTPase cycle is thus robust to variation in binding affinities over a reasonable range. Kinetic scaffolding was also proposed by Zhong et al. [16] based on nucleotide exchange kinetics. Kinetic scaffolding does not suggest any direct interaction, physical or allosteric, between receptor and GAP, but describes functional and temporal stabilization of receptor binding because alternative paths for receptor-G protein complex occur faster than dissociation. Kinetic scaffolding does not minimize the role of physical scaffolds, which can stabilize signaling complexes prior to activation by agonist (reviewed in [46]), and protein scaffolds may in some cases obviate the need for kinetic scaffolding. Kinetic scaffolding becomes efficient during signal transduction, however, by maintaining signaling proteins in their active complex. Further, kinetic scaffolding maintains receptor and G protein in contact and correctly oriented, whereas physical scaffolds may provide loose tethers which may be less effective. Examination of the activation contour shows that deactivation upon removal of receptor (or, in cells, of agonist) is accelerated by GAP over a large and biologically important region of receptor-GAP space (Figure 7, inset). Deactivation is to some extent multiphasic at all points because activated species to which GAP is bound deactivate most rapidly, and further relatively fast deactivation depends on binding of GAP to other GTP-bound, activated species (Figure 7). Precise pathways vary depending on the concentration of GAP and fractional activation at the time receptor is removed. It is likely that such multiphasic deactivation occurs in cells upon removal of agonist, but determining the precise shape of such deactivation time courses is experimentally taxing, and determining the molecular events underlying each phase is not yet experimentally approachable. However, we can now use simulations of the sort shown in Figure 7 as guides to designing experimental studies of deactivation pathways. Using computational modeling to analyze a specific dataset is valuable in that conclusions are based on real data and are statistically verifiable. However, the conclusions are to some extent unique to the particular proteins used in the experiments, and the experimental system used here is clearly simplified in comparison to the natural plasma membrane. However, a biochemically defined experimental system of intermediate size, such as this one, allows studies of complex regulatory interactions and their mechanisms that would be impossible in a plasma membrane where local protein concentrations are unknown and where effects of other components are difficult to rule out. It will be important to analyze other G proteins, effectors and GAPs in this way, both to determine important differences among G protein modules at the mechanistic level and to verify that this approach is generally valid. The details of agonist interactions with receptors in the context of a functioning signaling module is also of enormous interest, but there is insufficient understanding of these phenomena to incorporate them into a thermodynamically complete, data-driven model. Our approach is in this sense complementary to the rigorous but mechanistically speculative work of the sort pioneered by Linderman and coworkers [43],[47],[48]; see also [49]. We also need to engage questions of how GAPs function as effectors, and the present work will both guide these experiments and motivate direct measurements of the key interactions discovered so far. Steady-state GTPase activity was measured in large, unilamellar phospholipid vesicles that contain purified m1 muscarinic cholinergic receptor, Gαqβ1γ2 and phospholipase C-β1 [11]. Vesicles were prepared as described and phospholipase was added subsequently. The average diameter of the vesicles is 71 nm diameter (SD = 5 nm) according to negatively stained electron microscopic images. Concentrations of each protein and the amount of recovered lipid were measured as described previously [11],[12]. For modeling, protein concentrations are calculated according to the volume of the vesicle bilayer (see below), which is itself calculated according to the concentration of total phospholipids in the vesicle suspension [11] and their averaged partial specific volume. Because the phospholipid bilayer is homogeneous, the concentration of each protein in each vesicle is assumed to be uniform. Vesicles contain an average of 0.8 to 5 receptors and 2 to 12 Gq molecules depending on their concentrations, which probably approximates their molar ratios in natural membranes [11]. The specific activity of agonist/GAP-stimulated GTPase activity in these vesicles varied by 37% (SD) among six preparations prepared over several months. GTPase activity was assayed as described [11],[50]. The assay times and the amounts of vesicles used were adjusted to maintain steady-state activity high enough for reliable determinations. Specific activities were calculated according to receptor-accessible Gq in cases where agonist stimulation was measured [11]. Activity with no input from receptor was determined either in the presence of atropine, an inverse agonist, or in vesicles that did not contain receptors. Receptor-free vesicles probably displayed slightly lower activity than receptor-replete vesicles assayed with atropine, although the difference was uncertain because of difficulty in quantitating total Gαq [22]. The GTPase datasets used in parameterization of the model are listed in (Table S1). In each, the concentration of one component (GTP, GDP, GAP) was varied while others were held constant. When the concentration of GTP is listed as equal to its Km, the value of Km was determined under that set of assay conditions. The concentration of receptor varied among vesicle preparations, but was not itself varied systematically. The biochemical model is implemented as a system of 14 ordinary differential equations that describe the concentrations of each of the protein species shown in Figure 1, plus free receptor and GAP (Figure S1). Concentrations of free GTP, GDP, and Pi are constants (i.e., steady-state conditions) for the modeling and simulation reported here. There are 48 kinetic constants, labeled as shown in Figure 1. Concentrations of receptor, G protein and GAP are calculated according to the volume of the lipid bilayer of the vesicles, and the total volume available for all proteins in the system is therefore the sum of the bilayer volumes of all the vesicles in the suspension [51]. This convention yields both second-order association rate constants and equilibrium association constants for protein-protein binding that are about 13,000-fold higher than would be calculated if concentration were expressed as the total aqueous assay buffer volume. First-order dissociation rate constants are not altered by this convention (Text S1). Proteins are assumed to be homogeneously distributed among all vesicles, and any local variation in concentration are assumed to be negligible. Specifically, the number of vesicles with one or more of the proteins absent is assumed to be negligible. Concentration of GTP, GDP, and Pi are calculated according to the aqueous assay volume. To assign values to the kinetic constants appropriate to the m1 muscarinic receptor-Gq-phospholipase C-β1 system, we simultaneously fitted all parameters listed in Figure S1 to all the data of the experiments in Table S1. Fitting minimizes the cost function, the total mean square deviation between the predictions of the model (vmod) and the data (vdata), adjusted for the standard deviation (σ) of triplicate determinations.(1)To search parameter space, we used simulated annealing, an iterative stochastic search of multi-parameter space guided by the Metropolis algorithm [52],[53] (and references therein). At each iteration, the model is numerically integrated to yield steady-state GTPase activities and the cost function is calculated. Parameters are then changed randomly and the model is re-evaluated. Changes that decrease the cost function are accepted. Changes that increase the cost function may also be accepted, but only probabilistically according the Boltzmann probability function that depends on the cost difference of the proposed change scaled by an order parameter analogous to temperature in statistical physics. Simulated annealing applies the Metropolis algorithm while decreasing the temperature control parameter. The process allows escape from local minima of the cost manifold and discovery of the global minimum [53]. A thermodynamically complete model, with all possible interactions of species included and all reactions considered to be reversible, allows the use of thermodynamic constraints during the fitting process in addition to adjusting parameters to minimize the cost function. These constraints include both the path-independence of ΔG for reactions connecting two species (ΔG = 0 for any closed loop) and the net ΔG of hydrolysis of GTP to GDP and Pi that is enzyme-independent. In most fits, the parameter set was adjusted to meet thermodynamic constraints at each cycle of the search. Alternatively, thermodynamic constraints may be used quantitatively as part of the cost function. Deferring imposition of strict thermodynamic constraints may potentially allow broader, more ergodic search of parameter space during fitting, and this strategy was also evaluated. In searches strictly constrained by path independence, each newly generated candidate parameter set was adjusted before recalculation of the cost function. Parameters to be recalculated to comply with path independence were chosen at random. A symbolic manipulator (Mathematica, Wolfram Research, Inc., Champaign, IL) was used to derive explicit expressions for all the possible combinations of recomputed parameter sets in terms of randomly generated ones. The subset of parameters to be recalculated was then chosen. This approach is valid because the constraint equations effectively reduce the number of independent kinetic parameters (degrees of freedom) in the system. When strict constraints were deferred, each new, randomly generated parameter set was used whether or not it satisfied thermodynamics constraints to increase the potential ergodicity of the algorithm. In order to remediate this violation, a penalty term based on stoichiometric network theory (SNT) [54] was added to the cost function for the fit shown in Equation 1. SNT provides a method to compute sums of the chemical potential drops over each of the elemental loops I of the reaction network [54]. These target sums, shown as si for loop i in Equation 2, may be zero or non-zero depending on whether a particular loop includes a non-zero chemical motive force (hydrolysis of GTP). The penalty term expresses the weighted effect of deviation from the target values for all the elemental loops of the network. Its addition to the cost function thus causes the simulated annealing process to drive the fit toward simultaneously satisfying the thermodynamics constraints and minimizing the least-squares fit to the data. Overall, fits using SNT penalties were found to be comparable to fits using strict thermodynamic constraints, although SNT-constrained searches converged less rapidly. Ending SNT-constrained searches with a strict thermodynamically constrained search was an efficient way to combine both methods.(2) The system of coupled differential equations (Figure S1) was solved using the ode15s solver, which is designed for stiff systems of ordinary differential equations (Matlab, The MathWorks, Natick, MA). For efficiency, Matlab source code was automatically translated to C and compiled as a UNIX executable. The process was maximally parallelized because each data point can be calculated independently. A typical run employed 80 to 100 processors. Most runs were performed on the UNIX clusters of the Texas Advanced Computing Center, Austin, TX. Model-based simulations were also generated using ode15s, values of the kinetic constants shown in Table S2 and concentrations of proteins, nucleotides and Pi given in the text. Simulations were run to steady-state unless shorter times are specified. The integrity of numerical computations was verified throughout by checking for conservation of molecular types and by agreement with analytical solutions in limiting regimes where possible. Each independent fitting search settles on a different parameter set which equivalently fits the data. Variability among fit results is due to the intrinsic coupling between parameters and the stochastic nature of the fit. We have verified that distributions of the logarithms of the association and dissociation constants from multiple search repeats are all peaked, unimodal and thus well approximated by single Gaussians. We derived a best estimate for each model parameter from the means of their logarithms. Similarly, we derived a measure of error on each fit parameter from the standard deviations of these distributions. This procedure is justified because the logarithm of a rate constant is proportional to activation energy; the average of logarithms preserves the validity of the thermodynamic relationships among them. To simulate the response of G protein signaling to addition and removal of agonist, we first brought the system to an initial steady-state without receptor. We then instantaneously introduced a finite amount of activated receptor and allowed the model to reach a new steady-state. After 200 s, activated receptor was instantaneously removed and the system was allowed to return to the original steady-state. To reveal the mechanisms underlying the observed dynamics, the fractional activity Z, the fluxes and the concentrations of all species were computed as a function of time. Figure 7 shows a typical simulated output pulse shape (Z as function of time) and the reaction pathways responsible for it. We also surveyed the response to a pulse over a grid of receptor and GAP concentrations (2,500 grid points). At each point on the grid, we computed the time required for fractional activity to drop to ZMax/e where ZMax is the plateau level of signaling output. To study mechanisms of GAP-accelerated signal termination under conditions where GAP minimally inhibits receptor-stimulated signaling, we searched the grid for locations where increasing the GAP concentration approximately 500-fold inhibited output by ≤5%. Locations where the higher GAP concentration accelerated signal at least two-fold are shown on inset of Figure 7. The mechanisms underlying the dynamic response were studied at selected points (Results).

10.1371/journal.pgen.1003088

The Genomes of the Fungal Plant Pathogens Cladosporium fulvum and Dothistroma septosporum Reveal Adaptation to Different Hosts and Lifestyles But Also Signatures of Common Ancestry

We sequenced and compared the genomes of the Dothideomycete fungal plant pathogens Cladosporium fulvum (Cfu) (syn. Passalora fulva) and Dothistroma septosporum (Dse) that are closely related phylogenetically, but have different lifestyles and hosts. Although both fungi grow extracellularly in close contact with host mesophyll cells, Cfu is a biotroph infecting tomato, while Dse is a hemibiotroph infecting pine. The genomes of these fungi have a similar set of genes (70% of gene content in both genomes are homologs), but differ significantly in size (Cfu >61.1-Mb; Dse 31.2-Mb), which is mainly due to the difference in repeat content (47.2% in Cfu versus 3.2% in Dse). Recent adaptation to different lifestyles and hosts is suggested by diverged sets of genes. Cfu contains an α-tomatinase gene that we predict might be required for detoxification of tomatine, while this gene is absent in Dse. Many genes encoding secreted proteins are unique to each species and the repeat-rich areas in Cfu are enriched for these species-specific genes. In contrast, conserved genes suggest common host ancestry. Homologs of Cfu effector genes, including Ecp2 and Avr4, are present in Dse and induce a Cf-Ecp2- and Cf-4-mediated hypersensitive response, respectively. Strikingly, genes involved in production of the toxin dothistromin, a likely virulence factor for Dse, are conserved in Cfu, but their expression differs markedly with essentially no expression by Cfu in planta. Likewise, Cfu has a carbohydrate-degrading enzyme catalog that is more similar to that of necrotrophs or hemibiotrophs and a larger pectinolytic gene arsenal than Dse, but many of these genes are not expressed in planta or are pseudogenized. Overall, comparison of their genomes suggests that these closely related plant pathogens had a common ancestral host but since adapted to different hosts and lifestyles by a combination of differentiated gene content, pseudogenization, and gene regulation.

We compared the genomes of two closely related pathogens with very different lifestyles and hosts: C. fulvum (Cfu), a biotroph of tomato, and D. septosporum (Dse), a hemibiotroph of pine. Some differences in gene content were identified that can be directly related to their different hosts, such as the presence of a gene involved in degradation of a tomato saponin only in Cfu. However, in general the two species share a surprisingly large proportion of genes. Dse has functional homologs of Cfu effector genes, while Cfu has genes for biosynthesis of dothistromin, a toxin probably associated with virulence in Dse. Cfu also has an unexpectedly large content of genes for biosynthesis of other secondary metabolites and degradation of plant cell walls compared to Dse, contrasting with its host preference and lifestyle. However, many of these genes were not expressed in planta or were pseudogenized. These results suggest that evolving species may retain genetic signatures of the host and lifestyle preferences of their ancestor and that evolution of new genes, gene regulation, and pseudogenization are important factors in adaptation.

Cladosporium fulvum and Dothistroma septosporum are two related fungal species belonging to the class Dothideomycetes. C. fulvum is a biotrophic pathogen of tomato that has served as a model system for plant-microbe interactions since its first effector gene, Avr9, was cloned in 1991 [1]. It is not related to species in the genus Cladosporium sensu strictu, and has recently been renamed Passalora fulva [2]. However, to be consistent with past literature it will be referred to here as C. fulvum. Phylogenetic analyses based on sequences of the internal transcribed spacer (ITS) region of the ribosomal DNA revealed that C. fulvum is closely related to D. septosporum and other Dothideomycete fungi such as species of Mycosphaerella isolated from eucalyptus [3]. D. septosporum is an economically important hemibiotrophic pathogen of pine species that is well known for its production of an aflatoxin-like toxin, dothistromin [4]. A taxonomic revision also occurred for this species: prior to 2004 the name Dothistroma pini (syn. D. septosporum syn. D. septospora) was widely used. The revision involved a split into two species: the best-studied and most widespread species was named D. septosporum, and a less common species retained the name of D. pini [5]. The disease caused by C. fulvum, leaf mold of tomato, likely originates from South America, the center of origin of tomato [6]. The first outbreak of the disease was reported in South Carolina, USA, in the late 1800s [7]. Since then, disease outbreaks have occurred worldwide in moderate temperature zones with high relative humidity. The disease was of high economic importance during the first half of the 20th century, but its importance waned after introgression of Cf (for C. fulvum) resistance genes by breeders into tomato cultivars began providing effective control [8]. However, recent outbreaks have been reported in countries where tomato cultivars lacking Cf resistance genes are grown, and in areas where intensive year-round cultivation of resistant tomato plants led to fungal strains overcoming Cf genes [9], [10]. In contrast, the foliar forest pathogen D. septosporum (Dorog.) Morelet has a relatively recent history and has been less intensively studied than C. fulvum. D. septosporum infects over 70 species of pine, as well as several minor hosts including some Picea species [11]. During the 1960s–1980s, Dothistroma needle blight (DNB) was largely a problem of Southern hemisphere pine plantations, where primary control was achieved by fungicide applications or planting of resistant species (reviewed in [12]). Since the early 1990s DNB incidence has increased greatly in the Northern hemisphere, with some epidemics causing unprecedented levels of mortality [13], [14]. In northwest British Columbia, disease outbreaks are correlated with summer rainfall levels, suggesting that climate change could have unpredictable and severe effects on DNB outbreaks in forests [15]. Infection in both the C. fulvum-tomato and D. septosporum-pine pathosystems starts with conidia that germinate on the leaf surface and produce runner hyphae that enter the host through open stomata. Subsequently, the fungi colonize the apoplastic space between mesophyll cells. In the case of C. fulvum, conidiophores emerge from stomata 10–14 days later producing massive amounts of conidia that can re-infect tomato [16], [17], [18] (Figure 1A–1D). D. septosporum produces conidia, several weeks after infection, on conidiomata that erupt through the needle epidermis where they can be spread to other pines by rain splash [19], [20] (Figure 1E–1H). Whilst C. fulvum is considered a biotroph, D. septosporum is assumed to be a hemibiotroph based on similarities of its lifecycle to other Dothideomycete fungi. There is no evidence that C. fulvum has an active sexual cycle, although both mating type idiomorphs occur in its global population [21]. Although D. septosporum also has a predominantly asexual lifestyle, it is known to be sexually active in some parts of the world. The sexual stage Mycosphaerella pini Rostr. (syn Scirrhia pini Funk & Parker) has been reported in some forests in Europe and North America but has not yet been found in other regions, such as South Africa or the United Kingdom, even though both mating types are known to be present [22]. The rare sightings of the sexual stage are due partly to difficulties in identification, but also reflect findings from population studies that show mixed modes of reproduction with a significant clonal component [23], [24]. So far, attempts to induce a sexual cycle between opposite mating types of D. septosporum in culture in our laboratory or others (Brown A, unpublished data) have failed. Further research is required to determine environmental conditions conducive to sexual reproduction. The D. septosporum isolate whose genome was sequenced is derived from a clonal population with a single mating type that was introduced into New Zealand in the 1960s [22], [25]. The C. fulvum-tomato interaction complies with the gene-for-gene model [2], [26]. During infection C. fulvum secretes effector proteins into the apoplast of tomato leaves which function not only as virulence factors, but also as avirulence (Avr) factors when recognized by corresponding tomato Cf resistance proteins. This recognition leads to Cf-mediated resistance that often involves a hypersensitive response (HR) preventing further ingress of the fungus into its host plant tomato [8]. To date many cysteine-rich effectors have been cloned from C. fulvum, including Avr2, Avr4, Avr4E and Avr9, that can trigger Cf-2-, Cf-4-, Cf-4E-, and Cf-9-mediated resistance, respectively, and Ecps (extracellular proteins) like Ecp1, Ecp2, Ecp4, Ecp5 and Ecp6 that trigger Cf-Ecp-mediated resistance [27], [28], [29]. Specific functions for some C. fulvum effectors have been determined: Avr4 is a chitin-binding protein that protects fungi against the deleterious effects of plant chitinases [30], [31], Ecp2 is a virulence factor that occurs in many fungi [32], [33] and Ecp6 sequesters chitin fragments released from fungal cell walls by chitinases during infection thereby dampening their potential to induce pathogen-associated molecular pattern (PAMP)-triggered immunity [28]. Initially, the Avr and Ecp effectors seemed unique to C. fulvum, but in recent years homologs of Avr4, Ecp2 and Ecp6 with functions in virulence have been found in other fungal genomes, including members of the Dothideomycetes [27], [28], [33]. Whilst most studies of C. fulvum have focused on effectors and their interactions with components in both resistant and susceptible plants, studies of D. septosporum have instead focused on dothistromin, a toxin produced by the fungus that accumulates in infected pine needles. Dothistromin is a broad-spectrum toxin with structural resemblance to a precursor of the highly toxic and carcinogenic fungal metabolite, aflatoxin [34]. Although dothistromin is not essential for pathogenicity [35], recent observations suggest it to be a virulence factor, affecting lesion size and spore production (Kabir MS and Bradshaw RE, unpublished data). Some dothistromin biosynthetic genes were identified in D. septosporum but unexpectedly they were in several mini-clusters rather than in one co-regulated cluster of genes as reported for aflatoxin-producing species of Aspergillus [36], [37], [38]. The similarity of dothistromin to aflatoxin enabled predictions to be made about other D. septosporum genes involved in dothistromin production [39]; the complete set of dothistromin genes will help us understand the evolution of dothistromin and aflatoxin gene clusters. Here we report the sequence and comparison of the genomes of C. fulvum and D. septosporum, which have very similar gene contents but differ significantly in genome size as a result of different repeat contents. We found unexpectedly high levels of similarity in genes previously studied in one or other of these fungi, including those encoding Avr and Ecp effectors of C. fulvum, and dothistromin toxin genes of D. septosporum. Surprisingly, compared to D. septosporum, C. fulvum has higher numbers of genes normally associated with a necrotrophic or hemibiotrophic lifestyle such as genes for carbohydrate-degrading enzymes and secondary metabolite biosynthesis. However, in C. fulvum some of these genes were lowly or not expressed in planta and others were pseudogenized. Other C. fulvum genes that are absent in D. septosporum are putatively involved in virulence on its host plant tomato, such as the α-tomatinase gene. We suggest that regulation of gene expression and pseudogenization, in addition to evolution of new genes, are important traits associated with adaptation to different hosts and lifestyles of the two fungi that, however, also retained some signatures of their common ancestral host. The 30.2-Mb genome of D. septosporum (http://genome.jgi.doe.gov/Dotse1/Dotse1.home.html; GenBank AIEN00000000) was sequenced at 34-fold coverage (Table S1) and then assembled into 20 scaffolds (>2-kb), 14 of which were 407-kb or larger, have telomere sequences at one or both ends (Table S2) and mostly match chromosome sizes estimated from pulsed-field gel electrophoresis [36]. The six smallest scaffolds ranged from 2.3- to 5.2-kb in size so are not significant parts of the genome. The excellent assembly of the D. septosporum genome was facilitated by its very low repeat content of only 3.2% (Table 1; Table 2; Protocol S1). In contrast, the repeat-rich genome of C. fulvum (http://genome.jgi-psf.org/Clafu1/Clafu1.home.html; GenBank number AMRR00000000) was very difficult to assemble. Fourteen 2-kb paired-end or shotgun 454 sequencing runs for C. fulvum resulted in a 21-fold coverage of the 61.1-Mb assembly in 2664 scaffolds >2-kb (Table 1) with a total repeat content of 47.2% (Table 2). The sequencing strategy was initially based on the assumption of a genome size of around 40-Mb, but soon it appeared that the C. fulvum genome was much larger due to the high repeat content. Problems with the assembly are not caused by the sequencing coverage of C. fulvum because it is estimated to be sufficiently high for good coverage of the gene-encoding areas. Instead, they are a consequence of its high repeat content. An estimated additional 26-Mb of C. fulvum DNA reads could not be assembled as they were predominantly repeat sequences (Figure 2). In the remainder of the manuscript we refer to chromosomes (1 to 14) for D. septosporum and scaffolds for C. fulvum. Summary statistics for the two genomes are shown in Table 1 and at the Joint Genome Institute (JGI) Genome portal (jgi.doe.gov/fungi) [40]. The C. fulvum and D. septosporum genomes are predicted to encode approximately 14 and 12.5 thousand gene models, respectively. Nevertheless, the C. fulvum and D. septosporum genomes share more than 6,000 homologous gene models with at least 80% similarity at the predicted amino acid level, whereas this number drops to 3,000 gene models this similar when comparing C. fulvum or D. septosporum with other closely related Dothideomycete species such as Mycosphaerella graminicola and M. fijiensis (Figure 2, Figure S1). Similarly, most introner-like element clusters found in C. fulvum and D. septosporum are closely related, more than to elements in other Dothideomycetes [41]. Phylogenetic analysis of C. fulvum and D. septosporum genomes in the context of nine other Dothideomycetes [42] confirms that these two species are the most closely related of the sampled species (Figure 2), as was inferred earlier from ITS [3] and mating type sequences [21]. This gives us two very closely related genomes with drastically different genome sizes mostly due to the greatly increased repeat content of C. fulvum. The massive increase in repetitive elements in C. fulvum might result from expansion of one or more repeat families that are also present in D. septosporum. Therefore, we classified the different repeat families in D. septosporum and compared them with those in C. fulvum. This revealed that some of the repetitive element families present in D. septosporum have expanded in C. fulvum (Table 2). This is most remarkable for the Class I retrotransposons which comprise over 90% of the repetitive fraction in C. fulvum and together account for over 26-Mb of the assembled genome. Retrotransposons are also highly abundant in the large repeat-rich genome of the hemibiotrophic sexual pathogen Mycosphaerella fijiensis (Dhillon B, Goodwin SB and Kema GHJ, unpublished data). Both Copia and Gypsy LTR retroelements are expanded in C. fulvum compared to the D. septosporum genome, whereas LINEs are detected only in C. fulvum (Table 2). Some other fungal species that are closely related to each other, but have a different lifestyle, also differ in repeat content, such as the Leotiomycetes of which Botrytis cinerea (<1% repeats) and Sclerotinia sclerotiorum (7% repeats) are necrotrophs, while Blumeria graminis f. sp. hordei (64% repeats) is an obligate biotroph [43], [44]. The latter species is particularly enriched in Class I elements and one of several biotrophs that show expansion of genome size associated with high repeat content [44], [45]. In contrast to the retroelements, Class II DNA transposons comprise only a small percentage (4.7%) of the overall repetitive elements in C. fulvum, but 46.2% of the repeats in D. septosporum, although they make up only a small portion of the genome overall. Interestingly, helitron-like DNA transposons comprise 40% of all repeats in D. septosporum and are 25.8-fold higher in terms of sequence coverage than in C. fulvum, whereas the DDE-1, hAT, and MuDR_A_B DNA transposons present in C. fulvum are not present in D. septosporum. Helitrons are transposons that replicate by a rolling-circle mechanism and are found in a wide range of eukaryotes, including the white rot fungus Phanerochaete chrysosporium [46], and are thought to have a role in genome evolution [47]. Helitron-like repeats are particularly abundant on D. septosporum chromosomes 3, 6 and 11 (Figure 3) and usually occur in clusters, sometimes along with other types of repetitive elements. The organization of repeats in D. septosporum is striking in that for the majority of the chromosomes, most repeats are localized into just one or two large regions containing a mixture of repeat element types (Figure 3), although other small repeat clusters also occur. In many eukaryotes, centromeres are characterized by repetitive DNA [48], and therefore we propose that some of the larger complex repeat regions are centromeres, in line with similar suggestions made for other fungal genomes [49], although experimental confirmation is required. The absence of any repeat cluster from chromosome 14, along with the observation that it harbors only one telomere, suggests that it is a chromosome fragment. Repeats in fungi are affected by repeat-induced point mutation, also referred to as RIP, a defense mechanism employed by fungi to suppress transposable element activity that was first described in Neurospora crassa [50]. RIP is a process by which DNA accumulates G:C to A:T transition mutations. It occurs during the sexual stage in haploid nuclei after fertilization but prior to meiotic DNA replication. Clear evidence of RIP was found in both the C. fulvum and D. septosporum genomes (Table 3) and is mainly confined to repeat-rich regions. In total 25.9-Mb were RIP'd in C. fulvum and 1.1-Mb in D. septosporum, which represent 42.4% and 3.7% of their genomes, respectively. RIP occurred mainly on large repeated sequences (≥500 nucleotides) that represent 97.2% of all repeats in C. fulvum and 98.0% in D. septosporum (Table 3). The high rate of RIP in repeat regions is in the same range as that seen in other Dothideomycetes such as S. nodorum (97.2%; Table S3) [42], [51]. Although RIP is present at high levels in C. fulvum, we propose that it has not been able to prevent transposon expansion possibly due to very rare sexual activity. Of the RIP'd loci, C. fulvum has almost none (0.5%) and D. septosporum little (16.9%) outside the main classified repeat regions. This is different from N. crassa (Table S3), where 35.2% of all RIP'd loci are predicted to be non-repeat-associated. For N. crassa it has been shown that even single gene duplication events are prey to the RIP machinery, thereby exemplifying its efficiency and sensitivity [50]. Clearly such sensitivity is not applicable to C. fulvum and D. septosporum, nor for three other studied Dothideomycetes (Table S3). In the Dothideomycete phytopathogenic fungus Leptosphaeria maculans, RIP slippage is found in regions adjacent to repetitive elements. In that species RIP has occurred in genes encoding small secreted proteins, such as the effector genes AvrLm6 [52] and AvrLm1 [53] that are located in repeat-rich regions of the genome [54]; mutations in these genes caused by the RIP process enabled the fungus to overcome Lm6 and Lm1-mediated resistance, respectively. However, we found no evidence of RIP slippage into the known effector genes of C. fulvum and related effector genes in D. septosporum. One way to assist the assembly of a fragmented genome is to use synteny with a well-assembled genome of a closely related species to order the scaffolds [55]. We attempted to use the D. septosporum genome to improve the C. fulvum assembly in this way. However, although it was possible to map C. fulvum scaffolds onto the assembled D. septosporum genome (Figure S2A), individual C. fulvum scaffolds are not collinear along their length, but have only short blocks of synteny to different parts of the D. septosporum chromosomes. The syntenic regions of the C. fulvum and D. septosporum genomes are associated with just 461 of the C. fulvum scaffolds (Table 4). In contrast, the remaining >4,000 C. fulvum scaffolds are non-syntenic. A more detailed analysis with the ten largest C. fulvum scaffolds (two are shown in Figure S2B) revealed that they each match primarily to only one D. septosporum chromosome, suggesting predominantly intrachromosomal rearrangements (mesosynteny), as described for other Dothideomycete fungi [42], [56] (Condon B et al., unpublished data). As found in other fungi [43], [57] non-syntenic regions are repeat-rich; for C. fulvum 79.7% of the repeat sequences are present in non-syntenic regions (Table 4). Secreted proteins are important for communication of plant-pathogenic fungi with their hosts. They comprise not only enzymes required for penetration and growth on plant cell walls, but also proteins needed to compromize the basal defence system of plants by either suppressing or attacking it, as has been reported for several fungal effector proteins [58]. The percentage of proteins predicted to be secreted is similar for both C. fulvum (8.5%) and D. septosporum (7.2%), and in the same range as that predicted for other Dothideomycete fungi such as M. graminicola (9.1%) and S. nodorum (10.8%) [40], [42], [51], [59]. Genes encoding secreted proteins including effectors are subject to evolutionary selection pressure imposed by environmental and host plant factors [58], and they often show a high level of diversification. Repeat-rich, gene-poor regions have been proposed to contain genes involved in adaptation to new host plants. For example, in some Phytophthora species and in L. maculans significantly higher proportions of in planta-induced species-specific effector genes encoding secreted proteins are found in repeat-rich compared to repeat-poor regions [50], [59] and in pathogenic strains of Pyrenophora tritici-repentis transposable elements are associated with effector diversification (Manning V, Ciuffetti L, unpublished data). We hypothesized that we would find more genes encoding secreted proteins in repeat-rich regions that are less syntenic between the C. fulvum and D. septosporum genomes than in repeat-poor syntenic regions. We therefore compared the number of genes and their similarity at the nucleotide and protein levels in syntenic and non-syntenic regions of these two genomes (Table 5) using C. fulvum as the reference sequence due to its higher overall content of repeat elements in non-syntenic regions. The regions syntenic between C. fulvum and D. septosporum, representing 22.3-Mb of the C. fulvum genome, contain 70% of all predicted genes whereas 30% of the genes are located in the non-syntenic repeat-rich regions representing 38.8-Mb of the C. fulvum genome (Table 5). The syntenic regions contain most of the homologous genes that encode proteins with the highest level of conservation between the two genomes, whereas the proteins encoded by genes located in the non-syntenic, repeat-rich regions are less conserved. In syntenic regions, 89.9% of gene models have a bi-directional best BLAST hit (BDBH) to a D. septosporum gene model, with a mean predicted amino acid similarity of 85.2%, compared to non-syntenic with only 51.7% of gene models with BDBH and 65.1% amino acid similarity (Table 5). As expected, we found the repeat-rich, non-syntenic regions to have higher proportions of gene models encoding secreted proteins (10.4%, with a mean predicted amino acid similarity of 60.7%) and small secreted cysteine-rich proteins (2.8%) than in syntenic regions (7.6%, with a mean amino acid similarity of 81.1%, and 1.5% respectively) (Table 5), as has been reported for L. maculans [54]. Some C. fulvum effector homologs have previously been reported to occur in other Dothideomycete species including M. fijiensis, M. graminicola, and several Cercospora species [33], but in the D. septosporum genome we found the highest number of C. fulvum effector homologs discovered to date, including Avr4, Ecp2-1, Ecp2-2, Ecp2-3, Ecp4, Ecp5 and Ecp6. Of those, Avr4, Ecp2-1 and Ecp6 are core effectors [33] and show the highest identity (51.7%, 59.8% and 68.6% amino acid identity, respectively) with those present in C. fulvum, whilst Ecp4 and Ecp5 are pseudogenized. We were interested to know whether the D. septosporum effectors would be functional in triggering a Cf-mediated hypersensitive response (HR). Therefore, we inoculated plants of tomato cultivar Moneymaker (MM) carrying the Cf-Ecp2 resistance trait with Agrobacterium tumefaciens expressing potato virus X (PVX) containing D. septosporum Ecp2-1 and used PVX-containing C. fulvum Ecp2-1 as a positive control. D. septosporum Ecp2-1 triggered a Cf-Ecp2-1-mediated HR (Figure 4A), whilst MM tomato plants lacking Cf-Ecp2 did not show any HR when inoculated with PVX containing Ds-Ecp2-1 (results not shown). We also showed that the D. septosporum homolog of C. fulvum Avr4 is functional in triggering a Cf-4-mediated HR in Nicotiana benthamiana, as determined with an Agrobacterium transient transformation assay (Figure 4B). This is remarkable because D. septosporum infects a gymnosperm which is only distantly related to tomato, but apparently produces effectors that can be recognized by tomato Cf resistance proteins. It would be interesting to examine whether gymnosperms carry functional homologs of the well-studied Cf tomato resistance gene homologs [33], [60] or other major R genes that could confer resistance to D. septosporum. Major R genes have been shown to be involved in resistance of some pine species to Cronartium spp. rust pathogens [61], [62] and are thought to function in a gene-for-gene manner [63]. In C. fulvum, adaptation to resistant tomato cultivars is sometimes associated with deletion of effector genes [64]. Presence of repeats or location near a telomere can cause repeat-associated gene deletion [65]. We analyzed the location of all cloned C. fulvum effector genes in its genome. Many scaffolds containing an effector gene are very small (Figure 5), suggesting that they are surrounded by large repeats hampering assembly into larger scaffolds. The location of the C. fulvum effectors is shown in Figure 5 and the types of flanking repeats are detailed in Table S4. The well-characterized effector gene Avr9 is located on a very small (20-kb) scaffold (Figure 5) and is likely flanked on both sides by repeats; on one side there are 11-kb of repeats on the scaffold and on the other side probably also repeats just outside the region shown that prevented further scaffold assembly. This suggests that the absolute correlation found between deletion of the Avr9 gene in C. fulvum and overcoming Cf-9-mediated resistance [64] is most likely due to the close proximity of Avr9 to large, unstable repeat regions. As well as causing deletions, transposons can contribute to genome plasticity by mutation due to transposition into coding sequences. During co-evolution, transposons have inserted into effector genes causing their inactivation and overcoming Cf-mediated resistance in C. fulvum, as has been reported for inactivation of both Avr2 [64] and Avr4E [66]. The C. fulvum homologous effector genes present in D. septosporum are also often in close proximity to repeat-rich areas that may represent centromeres (Figure 3), but the biological significance of this is not yet clear. Pseudogenization of two D. septosporum effector genes, Ecp4 and Ecp5, homologous to those reported for C. fulvum [67], could point to host adaptation in the DNB fungus at the pine genus, species or cultivar level. Future population analysis of both fungal strains and host genotypes will reveal the mechanism behind this phenomenon. Another class of well-studied C. fulvum small cysteine-rich secreted proteins is the hydrophobins. These amphipathic proteins are implicated in developmental processes in filamentous fungi and are localized on the outer surface of fungal cell walls [68]. They are divided into class I and class II hydrophobins based on sequence differences that also correlate with their different solubility [68]. Six hydrophobin genes (Hcf-1 to Hcf-6) had previously been identified from C. fulvum [66], [67]. We identified five additional hydrophobin genes in the C. fulvum genome [two class I (Cf187601 and Cf189770) and three class II (Cf197052, Cf188363 and Cf183780)] (Figure S3), which makes C. fulvum the Ascomycete species with the largest number of hydrophobin genes reported so far. In the D. septosporum genome only four hydrophobin genes were found, one of which (Ds75009) is predicted to encode a class II hydrophobin and was highly expressed both in culture and in planta. Based on EST data the 11 C. fulvum hydrophobin genes show a range of different expression patterns. Of the six class I C. fulvum hydrophobins, two were only expressed in culture [Cf184635 (Hcf-2) and Cf189850 (Hcf-4)], three were expressed both in culture and in planta [Cf189770, Cf187601 and Cf193176 (Hcf-1)], and one was not expressed in culture or in planta [Cf184193 (Hcf-3)]. Three of the class II C. fulvum hydrophobins were only expressed in culture [Cf197052, Cf188363 and Cf193013 (Hcf-5)], whilst Cf193331 (Hcf-6) and Cf183780 were expressed neither in culture nor in planta. None of the C. fulvum hydrophobin genes were expressed in planta only. It has been proposed that hydrophobins may act as ‘stealth’ factors, preventing the invading fungus from detection by its host plant [69] or protecting it against deleterious effects of plant chitinases and β-1,3 glucanases as reported for C. fulvum [70]. Early functional studies focused on the hydrophobin genes Hcf-1 (Cf193176) and Hcf-2 (Cf184635). Knocking down expression of Hcf-1, Hcf-2, or both genes by homology-dependent gene silencing did not compromise virulence [71],[72]; a similar result was reported for knock-down mutants of class I Hcf-3 and Hcf-4 and class II Hcf-6 genes [73]. A phylogenetic tree (Figure S3) shows that the four class I genes (Hcf-1 to Hcf-4) are paralogs, suggesting functional redundancy that might explain the lack of a phenotype; functional redundancy may also exist between different classes. It would be interesting to examine the role in virulence of the two most similar hydrophobin class I and class II genes of C. fulvum and D. septosporum (Cf 189770/Ds67650 and Cf197052/Ds75009, respectively) either by knock-out or knock-down strategies. Because C. fulvum and D. septosporum have very different plant hosts and pathogenic lifestyles, we expected that their capacity to degrade carbohydrates would also differ and that this might be reflected in their gene complements and expression profiles. We compared numbers of genes predicted to encode carbohydrate-active enzymes (CAZymes) [74] in these two fungi to those in other fungi representative of different lifestyles. As seen for grouped families of CAZyme genes in Table 6 (e.g., GH family of glycoside hydrolases), both C. fulvum and D. septosporum have gene numbers in the same range as hemibiotrophic and necrotrophic fungi, and many more than the obligate biotroph B. graminis f. sp. hordei. Despite this, both C. fulvum and D. septosporum have fewer predicted cellulolytic enzyme genes (e.g., GH6, GH7) as well as fewer genes classified in carbohydrate binding module gene families (e.g., CBM1) than most of the other fungi shown except for M. graminicola (Table 6, Table S5). The reduced number of predicted genes for cell wall-degrading enzymes in M. graminicola was hypothesized to represent an adaptation to avoid host defenses during stealth pathogenicity [59], which also may apply to C. fulvum and D. septosporum. However, it is known that even a small number of genes can enable high levels of enzymatic activity, as has been shown for the strongly cellulolytic fungus Trichoderma reesei [75]. Next we focused on CAZyme gene families that appear to differ in gene number between C. fulvum and D. septosporum. Because small differences in gene number could be due to mis-annotation, only families that differed by two or more genes were considered and examples of these are shown in Table 6 (full data in Table S5). Potentially interesting is the expansion of genes associated with pectin degradation in C. fulvum. For example, in the GH28 family that includes many pectinolytic enzymes, C. fulvum has 15 genes whilst D. septosporum has only four. A higher pectinolytic activity in C. fulvum is concordant with the higher pectin content of its host, tomato, compared to the pine host of D. septosporum [76], [77], but larger numbers of genes encoding pectin-degrading enzymes have generally been associated with a necrotrophic rather than a biotrophic lifestyle in fungi [78]. High pectinolytic activity is observed in fungi such as Botrytis cinerea [79], [80] that invades soft, pectin-rich plant tissues causing a water-soaked appearance of the infected tissues [43]. However, during colonization of tomato leaves by C. fulvum this type of symptom is never observed [79], [80]. Instead of contributing to the destruction of host cell walls, the C. fulvum pectinolytic enzymes may facilitate local modification of primary cell walls of mesophyll cells allowing the fungus to thrive in the apoplast of tomato leaves, as suggested for the ectomycorrhizal fungus Laccaria bicolor that thrives on plant roots [81]. Although C. fulvum has a large arsenal of pectinolytic genes compared to D. septosporum, not all of them appear to be functional. For example, two of the six GH78 and one of the two GH88 pectinolytic genes are pseudogenized in C. fulvum, whilst the corresponding D. septosporum families do not contain pseudogenes. Another constraint to function is that gene expression appears to be tightly regulated. As shown in Table 6, none of the 15 C. fulvum GH28 genes appear to be expressed in planta, whilst all four D. septosporum GH28 genes are expressed. Indeed in all gene families with predicted pectinolytic function shown in Table 6 (GH28, GH78, GH88, GH95, PL1, PL3), expression in planta was only detected for 2 of the 31 C. fulvum genes, whilst all 6 genes in these pectinolytic gene families were expressed in D. septosporum. It is possible that C. fulvum pectinases are only expressed very locally to modulate complex primary cell wall structures. The location and accessibility of pectin structures embedded in the cell wall is an important consideration for its enzymatic degradation. For instance, the Basidiomycete Schizophyllum commune grows predominantly on beech and birch wood which is poor in pectin [82]. However, the pectin in these cell walls is concentrated around the bored pits that are used by S. commune to enter the wood, explaining why this fungus contains a higher number of pectinase genes than would be expected based on the overall host pectin content. Differences in pectinolytic gene content and expression between C. fulvum and D. septosporum may therefore be related to their different strategies of host invasion and subsequent colonization. In addition to increased numbers of pectinolytic genes compared to D. septosporum, C. fulvum has more genes for enzymes that degrade hemicelluloses (e.g., families GH31, GH35 and GH39) [83] and hemicellulose-pectin complexes (GH43) (Table 6). It also contains 11 genes (compared to 4 in D. septosporum) encoding CE5 enzymes; these include cutinases that are required for early recognition and colonization of the host by fungal pathogens [84], [85]. The presence of so many genes encoding enzymes for plant cell wall and cuticle degradation in a biotrophic fungus like C. fulvum that enters its host via stomata is unexpected. However, the number of cutinase genes, and other secreted lipase genes is particularly low in the D. septosporum genome compared to other Dothideomycetes, a feature shared with the other tree pathogens Mycopshaerella populorum and M. populicola [42]. Overall our comparison shows a similar complement of CAZy genes between C. fulvum and D. septosporum, but an increased number of particular CAZyme families in C. fulvum including genes encoding pectin- and hemicellulose-degrading enzymes. However, a large proportion of genes in the C. fulvum CAZyme families lack expression in planta and some genes are pseudogenized. A second aspect of carbohydrate metabolism that we considered was a comparison of growth on defined and complex carbon substrates (Figure 6 and Figure S4; www.fung-growth.org). It was anticipated that growth profiles could illuminate differences between pathogens with dicot and gymnosperm hosts and show correlations with their respective gene complements. In a study of polysaccharide hydrolysis activities of many fungal pathogens, King et al. [86] showed preferential substrate utilization based on host specificity (dicot or monocot). In general D. septosporum grows more slowly on minimal control medium [87] than C. fulvum, but surprisingly overall the growth profiles of the two fungi are similar on most substrates (Figure 6 and Figure S4; Table S6) and both appear to utilize a broader range of substrates than M. graminicola (Figure 6). This is not only the case for the oligomeric and polymeric carbon substrates, requiring CAZymes for degradation, but also for monomeric carbon substrates, suggesting a diverse and efficient carbon catabolism in C. fulvum and D. septosporum. The good growth of D. septosporum on sucrose is particularly striking, suggesting that it can utilize sucrose available in apoplastic fluid during its early biotrophic colonization phase. In terms of complex carbon sources, D. septosporum shows a slightly better capacity than C. fulvum to utlise apple and citrus pectin (Figure 6 and Figure S4). This seems to contradict the higher pectinolytic gene numbers in C. fulvum compared to D. septosporum, but is supported by the expression of fewer C. fulvum pectinolytic genes during infection of tomato when compared to the D. septosporum pectinolytic genes during infection of pine needle (Table 6). Interestingly, good growth on pectin is also observed for M. graminicola, despite an even lower number of putative pectinases than D. septosporum. This suggests that regulation of expression is a more dominant factor in pectin degradation by these plant pathogens than the number of pectinase-encoding genes in their genomes. In contrast, pectinase gene numbers correlate well with growth profiles of Aspergillus nidulans, A. oryzae and A. niger [88]. Compared to growth on controls lacking a carbon source, D. septosporum also showed slightly better growth than C. fulvum on lignin. This would be consistent with the higher proportion of lignin in pine needles, estimated to be 25–30% of dry weight [89], compared to less than 10% in dicots [90]. However, due to the very slow growth of both fungi and the non-uniform growth habit of D. septosporum on these media, firm conclusions about their abilities to utilize lignin cannot be made. Tomato plants produce the antimicrobial saponin, tomatine. The tomato pathogen Fusarium oxysporum produces α-tomatinase, which functions as a virulence factor as it degrades tomatine into the non-toxic compounds tomatidine and lycotetraose [91]. A gene predicted to encode α-tomatinase, classified as a GH10 enzyme, was found in the C. fulvum genome (JGI ID 188986) but is absent from the D. septosporum genome. Another gene found only in C. fulvum shows predicted similarity to the GH5 family enzyme hesperidin 6-O-α-L-rhamnosyl-β-glucosidase that can degrade hesperidin [77]. Hesperidin occurs most abundantly in citrus fruits [92] and is a member of the flavonoid group of compounds that is well known for its antimicrobial activity. Flavonoid-degrading enzymes such as hesperidin 6-O-α-L-rhamnosyl-β-glucosidase might enable C. fulvum to detoxify hesperidin or related compounds present in tomato. Chemical defence molecules in pine needles include antimicrobial monoterpenes. Thus it is expected that D. septosporum is adapted to tolerate or degrade these compounds whilst C. fulvum is not. Recent work on the pine pathogen Grosmannia clavigera revealed several classes of genes that are upregulated in response to terpene treatment [93]. After 36 h, major classes of upregulated genes included those involved in β-oxidation as well as mono-oxygenases and alcohol/aldehyde dehydrogenases that may be involved in activating terpenes for β-oxidation. A drug transporter, GLEAN_8030, was functionally analyzed and found to be required for tolerance of the fungus against terpenes, enabling G. clavigera to grow on media containing these compounds. A search for three of these genes, including GLEAN_8030, showed that both C. fulvum and D. septosporum genomes contain putative homologs and share a similar gene complement to each other (Table S7). However, since these genes have not all been functionally characterized in G. clavigera and all are predicted to encode proteins involved in general metabolic processes, further work is required to determine the roles of the homologs found in both C. fulvum and D. septosporum. As well as chemical mechanisms, plants employ basal structural defence mechanims including lignification of cell walls [94], [95]. Due to the abundance of lignin in pine needles that block access to usable cellulose, fungal pathogens and saprophytes living on pines have a particularly challenging environment [96]. For D. septosporum to complete its lifecycle, degradation of pine needle tissue must occur so that conidiophores bearing conidia can erupt through the epidermis (Figure 1H), which contains lignin [97]. This is in contrast to C. fulvum whose conidiophores emerge from tomato leaves through stomatal pores (Figure 1D). Thus, we investigated genes that may be involved in lignin degradation. Some saprophytic fungi utilize oxidoreductases, particularly class-II peroxidases such as lignin peroxidases, manganese peroxidases and laccases, and a number of H2O2-producing enzymes, to achieve lignin breakdown [98], [99]. However, the number of genes encoding oxidoreductases in D. septosporum is no higher than those of other Dothideomycetes (C. fulvum, M. graminicola and S. nodorum) that infect plants with lower levels of lignin (Table S8). D. septosporum appears to have a similar complement of laccase genes as C. fulvum and only one distant relative of a class-II peroxidase, missing in C. fulvum, but also present in M. graminicola and S. nodorum. Interestingly, the classical Ascomycete laccases found in C. fulvum, D. septosporum and M. graminicola bear a carbohydrate-binding domain (CBM20, putative starch binding domain). This type of laccase is only found in Dothideomycetes but the significance of this novel modular structure is unclear. Brown-rot saprophytes such as Serpula lacrymans have a reduced complement of ligninolytic genes compared to lignin-degrading white-rot fungi and are proposed to initially weaken lignocellulose complexes by non-enzymatic use of hydroxyl radicals prior to enzymatic assimilation of accessible carbohydrates [81]. It is likely that D. septosporum uses a similar strategy to breach the lignin-rich components of pine needles, as complete degradation of this polymer is not required to complete its life cycle. Secondary metabolites (SMs) are important compounds for the colonization of specific ecological niches by fungi. In particular, plant-pathogenic fungi can produce non-specific and host-specific toxic SMs [100]. SMs also include mycotoxins that contaminate food and feed and are harmful to mammals [100]. The only currently known SMs produced by C. fulvum and D. septosporum are cladofulvin and dothistromin, respectively [101], [102]; both compounds are anthraquinone pigments. In fungi, SM biosynthetic pathways often involve enzymes encoded in gene clusters [103] and always require the activity of at least one of four key enzymes: polyketide synthase (PKS), non-ribosomal peptide synthetase (NRPS), terpene cyclase (TC) or dimethylallyl tryptophan synthase (DMATS) [104]. It has been suggested that loss of SM biosynthetic pathways is associated with biotrophy [44], thus we searched for SM gene pathways in both genomes. Surprisingly, the biotroph C. fulvum has twice the number of key SM genes (23 in total) compared to the hemibiotroph D. septosporum (11 in total) (Table 7), of which 14 and 9, respectively, are organized into gene clusters along with other SM-related genes. The numbers of key SM enzyme-encoding genes are comparable to those of M. graminicola, but are lower than those in most other sequenced Dothideomycetes [42]. Like all Ascomycetes [105], the majority of key SM enzymes in C. fulvum and D. septosporum are PKSs, NRPSs and hybrid PKS-NRPSs. Annotation of all key SM genes was manually checked and two truncated (Pks4 and Nps1) and five pseudogenized (Pks9, Hps2, Nps5, Nps7 and Nps10) genes were found in the C. fulvum genome, while all D. septosporum genes except Pks4 (truncated) are predicted to encode functional enzymes. Overall, the number of predicted functional pathways suggests that C. fulvum and D. septosporum can produce at least 14 and 10 different SMs, respectively. Surprisingly, only three of the key SM genes are predicted to belong to biosynthetic pathways shared between the two species (Table S9) suggesting a diverse SM repertoire. This is much lower than expected given the overall level of similarity in gene content between the two genomes, and suggests that this SM repertoire is under strong selection. The three common genes are predicted to be involved in production of a pigment related to melanin (Pks1), a siderophore (Nps2) and dothistromin (PksA) based on similarities to other characterized genes. In C. fulvum, the three other functional non-reducing PKS enzymes are candidates for production of cladofulvin. The genomic locations of the 11 biosynthetic SM genes in D. septosporum do not show any enrichment at sub-telomeric positions, as reported for Aspergillus spp. and Fusarium graminearum [106], [107], or near putative centromeres (Figure 3). However, 8 out of the 11 genes are located on chromosomes smaller than 2-Mb (chromosomes 8 to 12; Figure 3). The genomic regions immediately surrounding all 11 D. septosporum SM genes are conserved in the C. fulvum genome, although 8 of them lack the key SM gene itself and, sometimes, putative accessory genes also (Figure S5). Reciprocally, only 9 C. fulvum SM genomic regions out of 23 are conserved in D. septosporum with 6 of these lacking the key SM gene, suggesting either gain or loss of SM genes has occurred. For two of the regions where flanking genes are conserved but SM gene(s) are missing in C. fulvum (regions corresponding to those surrounding Pks3 and Nps3 in D. septosporum), the presence of repeats suggests that SM gene loss may have occurred in C. fulvum (Figure S5). The C. fulvum-specific SM loci Pks5, Pks6, Nps5/Dma1 and Nps9 include many transposable elements and genes that have similarity to genes scattered in the D. septosporum genome, often on the same chromosome, leading to the hypothesis that these SM loci were assembled by gene relocation as recently proposed for the fumonisin gene cluster in F. verticillioides [108]. Analysis of the D. septosporum 1.3-Mb chromosome 12 revealed that the three previously identified mini-clusters of dothistromin genes [38] are widely dispersed, confirming fragmentation of this gene cluster (Figure 3 and Figure 7). Candidates for additional dothistromin genes, previously predicted based on aflatoxin pathway genes [39], are also present. Although three of these genes (OrdB, AvnA, HexB) are located in the published VbsA mini-cluster, the others are dispersed over different regions of chromosome 12 as shown in Figure 7. The end of the Nor1 gene cluster (Nor1, AdhA, VerB) is less than 10-kb from one predicted telomere, whilst Ver1 (previously called dotA [35]) is only 81-kb from the other telomere. As expected, a gene similar to the aflatoxin AflR regulatory gene is present and, like in aflatoxin-producing species of Aspergillus, is divergently transcribed with an adjacent AflJ regulatory gene candidate. Functional analysis of these genes is in progress. Although C. fulvum is not known to produce dothistromin, the complete set of predicted dothistromin genes is present in its genome, encoding proteins with amino acid identities ranging from 49% (AflJ) to 98% (Ver1) when compared with those of D. septosporum (Table S10). The arrangement of predicted dothistromin genes in C. fulvum reveals a high level of synteny with some rearrangements. With the exception of the Ver1 gene cluster, the mini-clusters contain the same genes in the same orientations in the two species (Figure 7A). The three mini-clusters on C. fulvum scaffold 130775 are much closer together than in D. septosporum, but are still separated from each other by considerable distances (approximately 24-kb between Est1 and the VbsA gene cluster, and 40-kb between the VbsA and Nor1 gene clusters). A comparison of the relative locations of the mini-clusters in the two species suggests inversions (AflR/J and VbsA gene clusters) as well as rearrangements over relatively small (VbsA-Nor1) and large (Ver1-AflR/J) distances. This is consistent with the overall pattern of intrachromosomal rearrangements observed between these two genomes. Given the presence of the dothistromin biosynthetic pathway genes, we tested whether dothistromin is produced by C. fulvum. However, no dothistromin was detected by HPLC analysis of extracts from C. fulvum PDB cultures, which is a condition favorable to dothistromin production by D. septosporum. Despite the lack of dothistromin production under these conditions, a strong evolutionary constraint on dothistromin biosynthetic genes was seen by analyzing the ratio of non-synonymous to synonymous mutations (Ka/Ks) between C. fulvum and D. septosporum. The low Ka/Ks ratios seen for dothistromin genes (range 0.018–0.169) are indicative of purifying selection [109] and did not differ from the distribution observed for four housekeeping genes (Tub1, Eif3b, Pap1, Rps9; range 0.003–0.073) (P = 0.561). Evidence for purifying selection was also shown for aflatoxin pathway genes in Aspergillus flavus and A. nomius [110]. On the basis of this we propose that C. fulvum might produce dothistromin, or a metabolite related to dothistromin, under certain environmental conditions when it is required. Many fungal SM biosynthetic pathways are cryptic, meaning that they are not expressed in wild-type strains under laboratory conditions. However, manipulation of genetic regulatory pathways or environmental conditions has shown that some of these cryptic pathways are functional [111], [112]. As seen for other gene families such as CAZyme genes, C. fulvum appears to be more economical in its expression of SM genes than D. septosporum, particularly in planta. In C. fulvum, EST support was obtained from in vitro conditions for all key SM genes except Hps2, Nps7 and Nps10, which are pseudogenized. The two truncated genes (Pks4 and Nps1) and the pseudogenized Nps5 genes also have EST support but the resulting proteins are unlikely to be functional. However, no evidence for in planta expression could be obtained for any of the C. fulvum key SM genes from this EST library. In contrast, all D. septosporum key SM genes have EST support from both in vitro and in planta libraries, with the unique DsPKS2 being one of the most highly expressed genes during pine needle infection. Differences in dothistromin pathway regulation were confirmed by quantitative PCR. In D. septosporum, Ver1, PksA, AflR and VbsA show higher expression during pine infection than in controlled culture conditions used to induce dothistromin production (Figure 7B). In contrast, the same genes show a low expression level in C. fulvum during infection and in vitro (Figure 7C). Because no dothistromin could be detected in vitro, this low expression likely represents background transcription with no biological relevance. Such an expression pattern is significantly different from the upregulation of Avr4 and Avr9 genes during tomato infection (Figure S6). SM production is associated with development in fungi and involves common regulators [113]. We searched the genomes of C. fulvum and D. septosporum for conserved regulators of development and SM production and, based on predicted protein sequences, found clear homologs for most of these genes in both fungi (Table S11). The two species appear to lack a PpoB oxygenase, but PpoA and PpoC are sufficient to produce all psi factors (oxylipins) identified in Aspergillus species [114]. In addition, C. fulvum lacks clear homologs of the G-protein regulators FlbA and RgsA, while possible homologs are found in D. septosporum. In Aspergillus species both proteins are negative regulators of G-protein signaling pathways. Neither C. fulvum nor D. septosporum have a homolog of BrlA, an essential regulator of conidiation in Aspergillus species [115], suggesting that they use another regulator for this role. Future studies analyzing expression of the SM genes, and the roles of regulatory genes, will help to determine fundamental differences in how C. fulvum and D. septosporum differentially regulate their SM gene expression. We embarked upon a comparative genomics analysis of C. fulvum and D. septosporum to test for differences that might explain their host specificity and lifestyles. The comparison revealed surprising similarities, such as the presence of dothistromin toxin genes in C. fulvum and functional Avr4 and Ecp2 effector genes in D. septosporum. However, the genome sizes of the two fungi are remarkably different, mainly due to a vast expansion of transposable elements in C. fulvum, and show several key differences in gene content. Adaptation of C. fulvum to its host plant tomato is exemplified by the specific presence of a gene encoding α-tomatinase, likely involved in degradation of tomatine. In contrast, the dothistromin gene cluster is present in both fungi, but while it is strongly expressed in D. septosporum at late stages of pine needle infection, it is lowly or not expressed in C. fulvum during infection of tomato leaves. Both fungi contain additional key SM genes, but the majority of these are not in common, contrasting with the high degree of homology between the two genomes. We suggest that this lack of conservation of key SM genes in the C. fulvum and D. septosporum genomes is a consequence of different evolutionary pressures that result from their different lifestyles, either as a pathogen inside their host or possibly as a saprophyte outside their host. Another key difference between the two fungi during pathogenesis concerns their differential gene regulation. Gene expression in C. fulvum is strictly regulated in planta, with many SM, hydrophobin and CAZy genes not expressed, while expression in D. septosporum is more constitutive. This differential regulation of expression may be crucial in determining differentiation between these fungi despite very similar gene profiles. Furthermore, this expression pattern is consistent with a biotrophic lifestyle without gene loss. Finally we suggest that the higher repeat content of the C. fulvum genome, along with evidence for gene pseudogenization (van der Burgt A et al., unpublished data) has facilitated the evolution of different lifestyles between C. fulvum and its sister species D. septosporum. Overall, our comparison of the two genomes suggests that even closely related plant pathogens could adapt to very different hosts and lifestyles by differentiating gene content and regulation, whilst retaining genetic signatures of a common ancestral way of life. The fungal strains of C. fulvum (race 0WU; CBS131901) and D. septosporum (strain NZE10; CBS128990) were isolated from tomato growing in an allotment garden in Wageningen, The Netherlands, in 1997, and from a needle from an eight-year-old Pinus radiata tree on the West Coast of the South Island of New Zealand in 2005, respectively. Monospore cultures, whose identities were confirmed by ribosomal ITS sequencing, were used throughout. Unless specified otherwise, cultures of these fungi were maintained on potato dextrose agar (PDA) or potato dextrose broth (PDB) media (C. fulvum) or Dothistroma Medium (DM; 5% w/v malt extract, 2.8% w/v nutrient agar or nutrient broth) at 22°C prior to use. Growth conditions used for generation of EST libraries (Protocol S1) are shown in Tables S12 and S13. Cultures were maintained for long-term storage in closed vials at −80°C stocks in 20% glycerol. Conidia of C. fulvum were harvested from two-week-old PDA plates with distilled water. The conidial suspension was filtered through Calbiochem Miracloth (EMD Millipore Chemicals, Philadelphia, PA) and washed once with water prior to calibration to 5×105 conidia/mL. Five-week-old Heinz tomato plants were sprayed on the lower side of the leaves with the conidial suspension (10 mL per plant). The plants were kept at 100% relative humidity for 48 h. The plastic-covered cages were then opened to grow the plants under regular greenhouse conditions (70% relative humidity, 23–25°C during daytime and 19–21°C at night, light/dark regime of 16/8 h, and 100 W/m2 supplemental light when the sunlight influx intensity was less than 150 W/m2). The 4th composite leaves of infected tomato plants were harvested at 2, 4, 8, 12 and 16 dpi, and immediately frozen in liquid nitrogen. To highlight the phylogenetic relationships of C. fulvum and D. septosporum with Dothideomycetes and other fungi relevant to this study, conserved protein families were predicted by use of the MCL Markov clustering program [116] with pairwise blastp protein similarities and an inflation factor of 4. From this multi-gene family set, 51 orthologous groups of genes were identified. Predicted protein sequences were concatenated, aligned using MAFFT 6.717b [117] and a species tree calculated using RAxML 7.2.8 [118]. We also determined protein homology data based on bidirectional best hits when comparing the proteomes of eleven Dothideomycete species (Alternaria brassicicola, C. fulvum, Cochliobolus heterostrophus, D. septosporum, Hysterium pulicare, Mycosphaerella fijiensis, Mycosphaerella graminicola, Pyrenophora tritici-repentis, Rhytidhysteron rufulum, Septoria musiva and Stagonospora nodorum), together with four out-group species (Aspergillus nidulans, Fusarium graminearum, Neurospora crassa and Magnaporthe grisea). Repeat sequences in both genomes were identified using RECON [119]. To group repetitive elements together into different families the default RECON output was parsed to include families with 10 or more elements. The parsed RECON repeat library was used to determine the extent of the repetitive fraction in the D. septosporum and C. fulvum genomes using RepeatMasker [120] and to annotate repetitive families and identify structural features, such as Long Terminal Repeats (LTRs) and Terminal Inverted Repeats (TIRs), using BLAST. Sequences that had undergone Repeat-Induced Point mutation (RIP) were identified according to the composite RIP index (CRI) method [121]. The CRI was calculated for each 500-nt sequence window, which was shifted at each 25-nt step. Sequences were identified as having been subjected to RIP when the RIP product, RIP substrate and composite RIP indices were at least 1.2, at most 0.8 and at least 1.0 respectively. As a final constraint, a series of overlapping sequence windows had to exceed 750 nt in length and the CRI value of any of the windows peaked to 1.5 in order to be scored as a RIP'd locus. Syntenic regions shared between C. fulvum and D. septosporum were detected ab initio on their repeat-masked genome sequences using promer [122], blastp and a suite of custom made python scripts. A script called blastpmer obtained all translated ORFs above a threshold nucleotide length from both query and subject genomes, performed a blastp on these ORFs, and subsequently filtered on expected value and high-scoring segment pair (HSP) length. Protein matches (using C. fulvum as query and D. septosporum as subject) were obtained with promer (–maxmatch) and blastpmer (–ORF 500 nt –HSP 250 nt –expect 1e-9). Both genomes were masked for these protein matches before being subjected to a second round of searching for weaker and shorter protein similarities, again using promer (–b 50 –c 15 –l 5 –maxmatch) and blastpmer (–ORF 300 nt –HSP 110 nt –e 1e-7). These four searches yielded 57,270, 44,865, 1,864 and 2,367 matches, respectively, many of which were redundant and overlapping. This large set was reduced to 24,480 unique matches by removing all except the best alignment for each unique genomic locus. This step removed overlapping alignments with different phases or orientations, and excluded suboptimal alignments caused by paralogs and common protein domains. The product of amino acid similarity and match length was employed as a final alignment quality score. Matches were ordered by query scaffold position and joined into linked syntenic regions according to the following criteria: (i) adjacent matches were identical on the query and subject scaffolds; (ii) matches had the same strand orientation; and (iii) maximum and average nucleotide distance between adjacent matches on the query and subject scaffolds were <10-kb and <5-kb, respectively. This step resulted in a reduction to 1,875 collinear match regions, of which 1,277 were >5-kb. For comparison of protein-coding genes in syntenic versus non-syntenic areas, gene models were classified as syntenic if they overlapped with any of the 1,875 collinear syntenic areas. Thus, subsets of 9,890 syntenic and 4,237 non-syntenic genes were inferred for C. fulvum. To investigate mesosynteny on a whole-genome scale, a refined synteny dataset was created with correction for inversions and rearrangements, and removal of spurious, small alignments. Match regions were compared and merged further if (i) adjacent groups had opposite orientations; or (ii) groups with identical query and subject scaffolds were separated by at least one (group of) matches on a conflicting subject scaffold, but maximum and average nucleotide distances between match regions were at most 20-kb or on average <10-kb apart; and finally (iii) match regions <5-kb were rejected. The final refined dataset contained 1,103 syntenic regions between 5 and 226-kb (average 22,194-bp), representing 22,700 matches from the original 24,480 unique matches. To identify potential C. fulvum and D. septosporum-specific proteins, the total protein sets from both fungi were used in comparative blastMatrix [123] searches against sequences from the nine additional members of the Dothideomycetes listed in the phylogenetics section. Initially, subcellular localizations for all C. fulvum and D. septosporum proteins were predicted using WoLF PSORT (wolfpsort.org; [124]). Only proteins containing a signal peptide and a signal peptide cleavage site, but lacking transmembrane (TM) domains or proteins containing a single TM that overlaps with the secretion signal, were selected. Signal peptides and cleavage sites were predicted using SignalP version 3.0 [125], where a final D-Score cut-off of 0.5 was used to increase specificity while retaining sensitivity. Subsequently, all proteins with signal peptides (1,886 and 1,591 for C. fulvum and D. septosporum, respectively) were analyzed for the presence of TM domains using the web servers Phobius [126] and TMHMM (version 2.0; [127]). The servers identified different, partially overlapping, sets of proteins with putative TM domains. On average Phobius detected 22% more TM domain proteins than did TMHMM, and about 75% of the predictions were shared between the servers. For further analyses, all proteins with putative TM domains as predicted by either of the two servers were removed from the dataset. Then, the proteins that contain a putative mitochondrial targeting signal as predicted by TargetP version 1.1 [128] were removed. Finally, proteins containing a potential GPI-anchor signal as predicted by the PredGPI web service were discarded [129]. A C. fulvum Avr4 (Cf-Avr4) gene homolog was identified in the genome of D. septosporum (Ds-Avr4) by blastp, with an E-value of 1×10−4. To determine Cf-4-mediated HR-inducing ability of Ds-Avr4 of D. septosporum, the Agrobacterium tumefaciens-mediated transient gene expression assay (ATTA) was performed in N. benthamiana as described by Van der Hoorn et al. [130]. The Cf-Avr4 and Ds-Avr4 genes were each fused to a PR-1A signal peptide sequence [131] for secretion into the apoplast. Subsequently a Gateway cloning strategy was performed to clone them into a pK2GW7 binary expression vector [132] containing the CaMV 35S promoter. A. tumefaciens (strain GV3101) was finally transformed with pK2GW7 binary vectors containing Cf-Avr4 or Ds-Avr4 genes by electroporation. Agroinfiltration of Cf-4 transgenic N. benthamiana leaves with Cf-Avr4- and Ds-Avr4-containing A. tumefaciens clones was performed as described by van der Hoorn et al. [130]. Photographs were taken at six days post inoculation. Three D. septosporum homologs of C. fulvum Ecp2 genes (Ds-Ecp2-1, Ds-ecp2-2 and Ds-Ecp2-3) were identified as described for Avr4. A binary Potato Virus X (PVX)–based vector, pSfinx, was used for transient expression of the Cf-Ecp2-1 ortholog, Ds-Ecp2-1, in MM-Cf-Ecp2 tomato lines based on methodology described by Hammond-Kosack et al. [131]. The recombinant viruses were obtained by cloning Ds-Ecp2-1 (an intron-less gene), encoding the mature protein, downstream of the PR-1A signal sequence for secretion into the apoplast and under the control of the CaMV 35S promoter. Recombinant pSfinx::Ecp2-1, corresponding to the C. fulvum Ecp2 (Cf-Ecp2-1), and pSfinx::Empty viruses were as published [33]. A. tumefaciens (GV3101) was transformed with the pSfinx::Ds-Ecp2-1 construct by electroporation. A. tumefaciens strains containing the pSfinx constructs for the expression of Cf-Ecp2-1 and Ds-Ecp2-1 proteins were inoculated on MM-Cf-Ecp2 tomato lines containing the cognate R gene, and MM-Cf-0 tomato lines that contain no R genes, mediating recognition of the Ecp2-1 effector. Photographs were taken four weeks post inoculation. All six previously reported hydrophobin genes from C. fulvum [133] were found in the automated gene predictions performed on the genome sequence. Five of the hydrophobins (Hcf-1 to Hcf-5) are predicted to contain an interpro motif common in fungal hydrophobins (IPR001338), while Hcf-6 has an interpro motif, which is restricted to Ascomycetes only (IPR010636). To identify putative hydrophobin-encoding genes in other genomes, all secreted gene models of C. fulvum, D. septosporum and M. graminicola were computationally annotated using Interpro scan and Gene Ontology terms. Then, gene models with IPR001338 and IPR010636 Interpro scan terms were identified as putative hydrophobin candidates. Also, a HMM profile search (which was built based on the conserved cysteine motifs in class I hydrophobins) was performed to identify hydrophobins missed by standard similarity searches. In this way five additional hydrophobin genes were identified in the C. fulvum genome. Hydrophobin sequences were aligned with ClustalW and edited in GeneDoc software. Then a consensus phylogenetic tree of predicted hydrophobin amino acid sequences was constructed using MEGA5 software [134] performing the minimum-evolution algorithm with default parameters and 1000 bootstrap replications. The carbohydrate-active enzyme catalogs of C. fulvum and D. septosporum were compared with the corresponding catalogs from other Dothideomycete fungi [42]. The boundaries of the carbohydrate-active modules and associated carbohydrate-binding modules of the proteins encoded by each fungus in the comparison were determined using the BLAST and HMM-based routines of the Carbohydrate-Active-EnZymes database ([74]; www.cazy.org). For determining the growth profiles on different carbohydrate substrates Aspergillus minimal medium [87] adjusted to pH 6.0 and containing 1.5% agar (Invitrogen, 30391–049) was used. Carbon sources were added at concentrations as indicated in the text and using standard methods as described at www.fung-growth.org. Duplicate plates were inoculated with 2 µL of a suspension containing 500 conidia/µL. Cultures were grown at 22–25°C for two weeks for C. fulvum and four weeks for D. septosporum, and representative plates were photographed. Genes encoding polyketide synthases (PKSs), non-ribosomal peptide synthases (NRPSs), hybrids of PKS and NRPS, terpene cyclases (TCs) and dimethylallyl tryptophan synthases (DMATSs) were sought in the two genomes using tblastn/blastp and several Ascomycete protein sequences as queries (Ace1 for PKS and hybrids; MGG_00022.7 protein for NRPS; tri5, cps/ks, all TCs from B. cinerea for TCs; Dma1 from Claviceps purpurea for DMATSs). For each tblastn/blastp hit, search for conserved domains (CDS at NCBI, InterproScan) and blastp analysis at NCBI and InterproScan confirmed the functional annotation. The locus of each key gene was analyzed for genes that could potentially be involved in a biosynthetic pathway. Functional annotation of downstream and upstream genes was confirmed using blastp at NCBI. In addition, homologs to genes that were shown to be involved in the regulation of fungal development and secondary metabolism were sought using tblastn/blastp with the sequences of the characterized proteins as queries. Ka/Ks calculations were carried out to estimate evolutionary constraints on putative dothistromin genes (PksA, VbsA, Ver1, HexA, AvfA, CypA and MoxA) in comparison to four housekeeping genes (Tub1 JGI PIDs Cf-186859 Ds-68998, Eif3b Cf-190521 Ds-75033, Pap1 Cf-190301 Ds-180959 and Rps9 Cf-196996 Ds-92035). DNA sequences from D. septosporum and C. fulvum were aligned with the codon-aware multiple sequence alignment software, RevTrans [135]. Sequence alignments were trimmed in codon units to remove missing data across both species with the sequence editor, Jalview [136]. The non-synonymous/synonymous amino acid ratio (Ka/Ks or ω) was obtained using the Ka/Ks Calculator [137] with the algorithm of Nei and Gojobori [138]. Statistical differences between Ka/Ks values for dothistromin and housekeeping genes were determined using Student's two-sided t test [139]. For determination of dothistromin production, previously published extraction and hplc methods were followed [140]. For quantification of dothistromin gene expression in D. septosporum, RNA was extracted from sporulating lesions on Pinus radiata needles collected from a forest in New Zealand (in planta sample) or grown in PDB or B5 [141] broths for 6 days as described previously [140]. cDNA synthesis and relative quantitative RT-PCR were carried out using primers and methods described earlier [140], with three biological replicates and two technical replicates. For C. fulvum, similar protocols were followed except that tomato infections, RNA extraction and cDNA synthesis followed the protocols of van Esse et al. [142] and four biological replicates were used. Oligonucleotides were designed with Primer3Plus [143] and are shown in Table S14. Their efficiency and specificity were tested on a genomic DNA dilution series. For both species, quantitative PCR was performed with the Applied Biosystems 7300 Real-Time PCR system (Applied Biosystems, USA) using the default parameters. Raw data were analyzed using the 2−ΔCt method [144].

10.1371/journal.pcbi.1005807

A machine learning approach for predicting CRISPR-Cas9 cleavage efficiencies and patterns underlying its mechanism of action

The adaptation of the CRISPR-Cas9 system as a genome editing technique has generated much excitement in recent years owing to its ability to manipulate targeted genes and genomic regions that are complementary to a programmed single guide RNA (sgRNA). However, the efficacy of a specific sgRNA is not uniquely defined by exact sequence homology to the target site, thus unintended off-targets might additionally be cleaved. Current methods for sgRNA design are mainly concerned with predicting off-targets for a given sgRNA using basic sequence features and employ elementary rules for ranking possible sgRNAs. Here, we introduce CRISTA (CRISPR Target Assessment), a novel algorithm within the machine learning framework that determines the propensity of a genomic site to be cleaved by a given sgRNA. We show that the predictions made with CRISTA are more accurate than other available methodologies. We further demonstrate that the occurrence of bulges is not a rare phenomenon and should be accounted for in the prediction process. Beyond predicting cleavage efficiencies, the learning process provides inferences regarding patterns that underlie the mechanism of action of the CRISPR-Cas9 system. We discover that attributes that describe the spatial structure and rigidity of the entire genomic site as well as those surrounding the PAM region are a major component of the prediction capabilities.

The CRISPR-Cas9 system, a microbial adaptive immune system, was recently exploited for modulating DNA sequences within the endogenous genome in many organisms. This system has emerged as a technology of choice for genome editing with promising therapeutic and research advancements. However, these exciting developments were not paralleled by deep understanding of CRISPR-Cas9 cleavage efficiency. Indeed, while numerous studies have been conducted in order to define better guidelines to determine CRISPR-Cas9 specificity, much ambiguity remains surrounding its mechanism of action. Here, we present a machine-learning based algorithm that was trained on genome-wide experimental data. The algorithm considers a broad range of features that describe different attributes that potentially impact the cleavage efficacy of CRISPR-Cas9 including genomic attributes, RNA thermodynamics, and those concerning sequence similarity. We further found that incorporating the possibility for DNA or RNA bulges play an important role in prediction accuracy. Together, these result in a predictive model that can be used both to predict the cleavage propensity of a new genomic site according to the genomic context, as well as to learn on the importance of different features on CRISPR-Cas9 efficiency and selectivity.

The Clustered, Regularly InterSpaced, Palindromic Repeats (CRISPR), and its associated protein 9 (Cas9) constitute a microbial adaptive immune system that was exploited in recent years for modulating DNA sequences within the endogenous genome in cultured cells and whole organisms [1–6]. The Cas9 endonuclease is directed by a programmable single guide RNA (sgRNA) to induce double strand breaks at specific genomic sites [7,8]. Recognition and cleavage occur via complementarity of a 20-nt sequence within the sgRNA to a genomic site, upstream to a Protospacer Adjacent Motif (PAM) at its 3’-end. Early studies demonstrated that multiple mismatches as well as DNA or RNA bulges can be tolerated [9–15], resulting in cleavage of unintended genomic sites, termed off-targets. This gave rise to devising key considerations for the design of an optimal sgRNA, namely, an efficient guide with minimal off-target effect. Such rules asserted that the number of mismatches should not exceed a specified bound, that mismatches at PAM-proximal positions are more influential than those occurring at PAM-distal positions, that spatially-dispersed mismatches are better tolerated, and that cleavage would not occur at sites that follow PAM sequences other than the canonical NGG (and occasionally NAG) [9–11,13]. However, early studies were not performed on a genome-wide scale as they analyzed off-targets that were pre-selected according to sequence similarity. Thus, such analyses were not designed to detect features outside the scope of pairwise sequence similarity. Subsequently, several experimental methods for unbiased genome-wide profiling of off-targets were introduced, including those based on integration of oligonucleotides into double strand breaks detected by sequencing (GUIDE-Seq) [16–18], high-throughput genome-wide translocation sequencing (HTGTS) [19], direct in situ breaks labelling (BLESS) [20,21], integration-deficient lentiviral vectors (IDLV) [22], and in-vitro nuclease-digested whole-genome sequencing (Digenome-seq) [23,24]. These studies demonstrated that CRISPR off-targets can be located at unexpected sites, such as sites that harbor alternative PAM sequences, sites that contain a large number of mismatches, and off-targets that were cleaved at higher frequencies than the intended on-targets. Thus, it is becoming clear that an intricate set of attributes play a role in CRISPR-Cas9 function. To date, several computational methods for sgRNA design were developed based on different design rules [25–42]. For example, CCTop [25] considers the distance of the mismatch from the PAM site when evaluating the specificity of candidate sgRNAs, ‘Optimized CRISPR Design’ [26] incorporates a position-specific mismatch penalty and additionally considers the spatial distribution of mismatches, and the CFD score [28] penalizes each mismatch according to its specific substitution type and position. Importantly, while these and other widely-used methods have been developed based on empirical data, they mostly neglect the genomic context surrounding the target sequence and instead focus on predicting off-target effects for a given sgRNA using basic sequence features [25,29,34,35,43]. Here, we introduce CRISTA, a novel methodology based on the machine learning paradigm for predicting the cleavage propensity of a genomic site by a given sgRNA. The method accounts for the possibility of bulges and incorporates a wide range of features encompassing those that are specific to the genomic content, features that define the thermodynamics of the sgRNA, and features concerning the pairwise similarity between the sgRNA and the genomic target. We show that CRISTA achieves a higher predictive accuracy than widely-used alternatives. We further examine our approach using a leave-study-out cross-validation procedure, thereby demonstrating that the predictive model represents general patterns of the cleavage machinery across different detection techniques. In addition to its predictive value, our method suggests additional information on the underlying mechanism of action of the CRISPR-Cas9 system, including attributes that were previously overlooked. The training dataset was assembled from published data obtained using several genome-wide unbiased methods for CRISPR-Cas9 cleavage sites profiling: GUIDE-Seq, HTGTS, and BLESS [16,17,19–21]. These datasets are termed hereafter Tsai [16], Kleinstiver [17], Frock [19], Ran [20], and Slaymaker [21]. The data in these studies are composed of collections of experimentally verified genomic targets throughout the genome, such that each target is denoted with the frequency of cleavage by a given sgRNA. We note that additional systems for cleavage sites detection are available, but these are not compatible with our objective to reveal genomic effects on CRISPR efficacy. For example, Digenome-Seq [23,24] does not provide cleavage frequencies in-vivo; the integrase-defective lentiviral vectors (IDLV) method can be used to detect off-targets in-vivo, but does not provide their cleavage frequencies [22]. Furthermore, a number of studies employed targeted sequencing approaches [15,22] to examine the cleavage frequencies of several genomic sites that were pre-selected based on prior deductions, and thus are lacking the information at the genomic scale. In total, data from five genome-wide studies were assembled, spanning 33 collections of sgRNAs and their respective targets obtained from 25 unique sgRNAs (S1 and S2 Tables). Combined, these sgRNAs cleaved 872 and 491 genomic targets across the genome before and after data filtration, respectively (see “Training dataset assembly” below). We refer to these data as the set of cleaved sites. Notably, the collection of targets was obtained from multiple methodologies and under different experimental conditions, hence, their reported cleavage efficiencies are not comparable and were thus transformed to a common scale. To this end, for each platform we extracted the set of sgRNAs that are in common with those from Tsai et al. [16], which is the most inclusive dataset. We then fitted the cleavage frequencies of the mutual targets of each study and Tsai data using linear regression. The inferred regression parameters were then used to transform the rest of the data obtained from the respective study (for more details see S1 Text, S1 and S2 Figs). In an initial exploratory phase, we observed that the pairings of the sgRNAs and the corresponding genomic sites, as originally reported, occasionally contained an exceedingly large number of mismatches. For example, 243 out of 872 sites retained five to ten mismatches, 22 of these had cleavage frequencies that were ranked among the highest 25% (S2 Table). This is in contrast to previous reports that showed that observing more than five mismatches is highly unlikely [9–11,13]. While these studies mainly concentrated on the number of mismatches, more recent evidence suggested that DNA/RNA bulges are also possible [12], and these can be represented as indel events in the context of sequence alignment. To account for this possibility and for additional specific characteristics of the CRISPR-Cas9 system, we modified the Needleman-Wunch pairwise alignment algorithm [44] to include two additional components: (i) Up to three single gaps are allowed over the whole alignment–a bound that was rarely met (and was never exceeded) in the set of cleaved sites but was necessary in order to detect potential off-targets in a computationally efficient manner. (ii) Since three gaps are allowed, each 20nt long genomic target is extended or shortened by up to three nucleotides, and the best pairwise alignment score over seven independent alignments between the DNA site (of length 17-23nt) to the corresponding sgRNA is selected. The pairwise alignment is determined by the match, mismatch, and gap parameters, such that a bulge (i.e., a gap), would be preferred over a mismatch only if the penalty paid for its insertion is compensated by the matches it induces. To determine the ideal parameters for pairwise alignment, we repeated the alignment procedure by ranging over different combinations of parameter values. The parameters that resulted in the maximal averaged squared Pearson correlation coefficient (r2) between the cleavage intensities and the pairwise-alignment scores were then selected. In this optimization procedure, targets of exact match were removed since these always result in the highest possible score and could shift the obtained r2 values. This procedure was performed either across the whole dataset, as well as for the partial data used in cross-validation (see below). A total of 119 targets, as obtained from the original studies, follow a non-NGG PAM (54 in Tsai data, 31 in Kleinstiver data, 34 in Frock data). Originally, the coordinates of the cleaved sites were detected by matching sequences to the reference genome while considering mismatches only. Thus, for example, if bulges are disregarded, a possible DNA-bulge upstream to a canonical PAM would be interpreted as a target with a non-canonical PAM. The introduction of gaps in the alignment allowed us to correct such instances. Hence, we re-evaluated the position of all non-NGG targets by shifting the PAM genomic coordinates 2-nt downstream or upstream in search for an NGG PAM or, if one did not exist, an NAG PAM at closest proximity. If none were found, the original PAM was preserved. We developed CRISTA, a tool for predicting the cleavage propensity of potential genomic targets given a specified sgRNA. CRISTA is based on learning a regression model using the Random Forest algorithm, and further allows the examination of the importance of features that determine the variation of cleavage efficiency. The development of a machine learning algorithm relies on (i) the assembly of a training dataset that encompasses a range of data inputs, and (ii) the incorporation of a set of features that can be used to predict cleavage efficiencies. The utility of the learning framework to distinguish between cleaved and uncleaved sites was also examined within a classification learning scheme (as opposed to a regression model). As the results were generally similar, those obtained with the regression model are presented throughout (see Discussion). We evaluated the prediction performance of CRISTA using two cross-validation procedures (Fig 1). We devised a leave-one-sgRNA-out procedure, such that in each iteration the samples of a single sgRNA were excluded and used as a test set. The algorithm, trained on the rest of the data, was then used to predict the cleavage probabilities for the test set. Each iteration of the cross-validation consisted of a preliminary step: the pairwise alignment parameters were first optimized as previously described using the training set only, and then were used to re-compute the pairwise alignment features for the training and the test sets. Similarly, we used a leave-study-out cross-validation strategy such that in each iteration all samples from a single study were excluded from the training data and used as a test set (note that Tsai data were divided to two datasets, S1 Text). Whereas the training dataset of CRISTA—which was used in the leave-one-sgRNA-out procedure and for all reported comparisons—did not include redundant sgRNAs to avoid overfitting of the model to the data, here we calculated the performance scores separately for sgRNAs that were uniquely inspected in one study (termed ‘unique guides’), and sgRNAs that were analyzed in more than one study (termed ‘common guides’; S1 Table). Several metrics (squared Pearson correlation coefficient and the area under the Receiver Operator Characteristics and Precision-Recall curves), were used to evaluate the performance of CRISTA and to compare it to three widely used alternatives; CCTop [25], the function for scoring single off-targets used in the online tool ‘Optimized CRISPR Design’ [26] (hereafter termed OptCD), and the CFD score [52]. The performance evaluation reported throughout was computed over the original set of cleaved sites for each sgRNA (without bootstrapping as was performed in the training set), and an equally-sized sample of uncleaved sites (see Results for the effect of this sample size on the performance evaluation). The Random Forest algorithm computes the relative contribution of the examined features to the regression model, termed feature importance. When the entire set of features is examined (S3 Table), some features may receive seemingly low importance values due to the presence of a correlated feature (e.g., the pairwise alignment score and the number of mismatches). To learn on the independent importance of the various features, we reduced the number of features by applying a forward selection procedure. Features were added iteratively by examining the performance of the leave-one-sgRNA-out cross-validation procedure for incremental sets of features. First, we tested which feature provides the highest Pearson r2 when examined independently. Then, in each iteration, the feature that increased the r2 the most was adjoined to the set. This procedure was repeated for 15 iterations. Random Forest was then applied to the resulting set of features and the relative importance of each feature was extracted. The introduction of gaps to the pairwise sequence alignment affected 18% of the targets in the training dataset, such that 87 of 491 sites contain 1.1 bulges on average (or an average of 1.23 in 175 out of 872 sites if considering the full dataset; S2 Table). This resulted in r2 = 0.34 (squared Pearson correlation coefficient between the pairwise alignment score and the observed cleavage frequencies) averaged over the sgRNAs datasets compared to r2 = 0.27 when gaps are not allowed. The optimized parameter values were 1 for a match, 0 for a mismatch and -1.25 for a gap (S3 Fig). We note that although mismatches are not explicitly penalized, matches are still awarded and so longer complementarity is generally preferred. Following this procedure, the number of mismatches was reduced from an average of 3.64 to 3.36 per target, such that six mismatches became very rare (S4 Fig). Reconsidering the PAM locations, such that NGG or NAG PAMs were found, resulted in a shift of 33, 17, and 22 instances (out of 54, 31, and 34 targets with rare PAMs) of Tsai, Kleinsteiver, and Frock data, respectively (S2 and S4 Tables). Notably, the pairwise similarity score explains merely 34% of the observed variation among the cleaved sites, which motivated us to integrate additional features in the prediction process. We devised CRISTA, a machine learning methodology that is based on the Random Forest regression model [47,48]. CRISTA was trained on several genome-wide experimental studies and combines a large set of explanatory features, to compute the cleavage propensity of a DNA target by an sgRNA. The resulting regression function of CRISTA is composed of a complex interaction between the incorporated features as represented by a set of decision trees. We evaluated the prediction performance of CRISTA in a leave-one-sgRNA-out cross-validation procedure, and compared it to the alternative tools. First, we calculated the squared Pearson correlation coefficient (r2) between the experimentally observed cleavage frequencies and the predictions. The scores that were predicted in the cross-validation conformed to the observed values with an r2 of 0.65. In comparison, OptCD produced an r2 of 0.13, the scores obtained using CCTop resulted in an r2 of 0.23, while the CFD score correlated best out of the three commonly-used alternatives with an r2 of 0.52 (Fig 2A–2D, S5 Table). A similar trend regarding the relative performance of the four scoring functions was obtained when Spearman rank correlation was computed (Spearman rho coefficients for CRISTA, OptCD, CCTop, and the CFD score were 0.81, 0.66, 0.64, and 0.74 respectively). Second, the receiver operating characteristic (ROC) curve was used in order to compare the abilities of the tools to discriminate between experimentally cleaved and uncleaved sites (thus, for this performance evaluation we treat these as the positive and negative sets, respectively), as measured by the area under the curve (AUC, values closer to 1.0 represent better predictions). To this end, we used the predicted scores as thresholds to delineate positives and negatives for the ROC calculation. Using this measure a similar trend was observed regarding the relative accuracy of the prediction methods (Fig 2E). CRISTA had the highest AUC score of 0.96 followed by the CFD score (AUC = 0.91), OptCD (AUC = 0.85) and CCTop (AUC = 0.85). Noticeably, all methods received high AUC scores, but this could be due to the large number of uncleaved sites that were included in the dataset. Hence, we further compared the ability to detect and to rank among the positive samples, as measured using the area under the Precision-Recall curve (PRC-AUC). Similar to the ROC curve, PRC-AUC values closer to 1.0 indicate highly successful predictions. Again, the ability of CRISTA to rank among the cleaved samples was favorable to the other three methods, with a PRC-AUC of 0.96, compared to 0.93, 0.88, and 0.87 that were obtained using the CFD score, OptCD, and CCTop, respectively (Fig 2F). The accuracy measures described above were computed while combining the predicted values across the whole dataset. Additionally, we tested whether the alternative prediction tools are consistent, that is, whether or not similar accuracies are obtained across different sgRNAs. The accuracy of CRISTA was found to be the most persistent across distinct sgRNA datasets, with an averaged r2 of 0.8 and a standard deviation of sd = 0.13. In comparison, the CFD score, OptCD, and CCTop obtained averaged r2 values of 0.65 (sd = 0.2), 0.32 (sd = 0.28) and 0.46 (sd = 0.25), respectively (Fig 2G; similar results were obtained when considering the ROC-AUC and PRC-AUC measures, Fig 2H–2I; averaged Spearman correlation coefficients were 0.88, 0.77, 0.76, and 0.72, respectively). Notably, while the uncleaved sites are an integral part of the learning process, as well as for assessing the accuracy of the different tools, the reported metrics could be biased to those sites with a “0” cleavage frequency. To examine to what extent the set of uncleaved sites affects the results, the averaged r2 was re-computed while altering the sample size of this set from 100% to 0% (relative to the size of the set of cleaved sites). Our results show that reducing the sample size has little impact on the relative success of the different tools. While the obtained r2 values decrease with lower proportion of uncleaved sites, the ones achieved by CRISTA are still better than the other alternatives (evidently, the decline for CRISTA is shallower than that obtained by the CFD score, which is the second-ranked tool; S6 Table). The learning dataset of CRISTA combines data from three experimental methodologies for genome-wide profiling of CRISPR cleavage sites with some of these applied in multiple experimental settings. Thus, we used a leave-study-out cross-validation procedure to examine whether the accuracy of CRISTA is dependent on a single platform that dominates the learning dataset. This allowed us to examine both the compliance of the different methods, and the performance of the predictive model on data that is similar to the training set (the set of common guides, see Methods, S1 Table) and on new data (unique guides). Our results demonstrated that, with the exception of the data by Frock et al., the different experimental procedures comply with one another (Fig 3, S7 Table). That is, when each study was used as a test set, without being included in the training set, the prediction made by CRISTA resulted in r2 higher than 0.8, and ROC-AUC and PRC-AUC values close to 1. In addition, the prediction accuracies of the common guides did not overwhelmingly exceed those of the unique guides, indicating that the prediction of cleavage efficiencies was accurate not only when the predictor was trained on similar sgRNAs as in the test data, but also when it was applied to unfamiliar data. Our analysis further demonstrated that the datasets obtained with HTGTS for unique sgRNAs are not comparable with those obtained with the other platforms. Therefore, Frock data was eliminated from the training dataset of CRISTA. A central component of the learning procedure implemented in CRISTA is the ample amount of data contained within the set of uncleaved sites as it conceals significant information regarding the features that hinder CRISPR-Cas9 action. Yet, such wealth of information was generally ignored by previous studies that aimed at devising rules regarding CRISPR-Cas9 specificity. To examine whether the enhanced accuracy achieved by CRISTA, as compared to other tools, stems from the inclusion of a large set of uncleaved sites, we repeated the leave-one-sgRNA-out procedure while retaining only the set of cleaved sites in the training set. The accuracy achieved by this model, referred to as CRISTA+, was substantially lower compared to CRISTA when trained on the whole dataset (S5 Fig), and is more similar to the one obtained using the CFD score. Beyond prediction capabilities, the learning process provided the opportunity to systematically learn the attributes that are most important for Cas9 function. When examining the entire set of features (S3 Table), three clusters emerged among the top first 25 (Fig 4): (i) features concerning the pairwise similarity between the sgRNA and the DNA site. Besides the pairwise alignment score, this cluster included the number of mismatches, the number of RNA/DNA bulges, and the mismatches types (i.e., whether they are transition, transversion, or wobble); (ii) features concerning the nucleotides content of the 20-nt site and its adjacent genomic region. These included the GC content, DNA enthalpy (a proxy for the DNA duplex stability [53]), and several measures that describe the spatial structure of the DNA including the minor groove width and the bending stiffness [54]; (iii) features concerning the PAM site and the surrounding nucleotides. These included the PAM type (i.e., NGG or NAG) and DNA geometry scores calculated in and around this region (i.e., NNGGNN if considering the canonical PAM). To learn about the features that are most important for prediction, and to reduce the redundancy introduced by correlated features, we obtained a succinct group of 15 elementary features using a forward selection process for which the relative importance was extracted (Fig 5, for the accuracy measurements achieved for the first 30 selected features see S8 Table). As expected, the pairwise alignment score was selected first and ranked as the most important. Additional attributes of the pairwise similarity, including the number of mismatches and their position, and the number of DNA/RNA bulges were also highly ranked. Additionally, a number of attributes describing the mismatch type (wobble, transversion, purine-purine, and pyrimidine-pyrimidine transitions) were found as important discriminative features. Particularly, we found that the relative frequency of wobble mismatches significantly increases with the total number of mismatches (p<0.05; S6 Fig) supporting the notion that wobble mismatches are better tolerated by Cas9 [16]. Extending beyond the pairwise similarity, our results revealed that the types of nucleotides in several positions also affect the sensitivity of CRISPR-Cas9. The selected features indicated the importance of the nucleotide at the second position upstream to the PAM, as was previously observed [28]. Additional nucleotides that were indicated to contribute to the prediction accuracy are the couple of nucleotides at positions 4–5, the site where cleavage occurs, and those in the first five positions downstream to the PAM (S7 Table). In addition, the results pointed at the significance of the nucleotide at the 20th position from the PAM site. Previous studies observed that there is a strong preference for guanine at the 5’-end of the genomic target [56,57]. However, given that all the sgRNAs in our data contain guanine in the 5’-end, the importance of the type of nucleotide at this position could well be an artifact of the assembled dataset. Among the genomic features that were examined, the presence of the target within DNase I Hypersensitive sites as well as within an exon (either on the coding strand or on the opposite one) were selected. These results support previous observations that reported higher tendency of targets near or around DNase I hypersensitive sites and in coding regions [58–60]. While both attributes signify an exposed DNA structure, the latter is also biased by the selection of on-targets. Interestingly, in addition to a simple categorization of the PAM type (i.e., NGG or NAG), the continuous measure that describes the width of the minor groove surrounding the PAM site was selected. Indeed, some DNA-binding proteins tend to interact with either the minor or major groove of the helix, and it was previously shown that changes in the groove width may affect their fit and therefore their function [61]. Cas9 crystallography highlighted that the PAM-interacting domain of Cas9 makes contacts with the major groove of the PAM duplex [62], and our results suggest that this interaction may be consequently influenced by the groove width. An additional feature that corroborates the importance of DNA geometry to Cas9 function is DNA enthalpy, which describes the binding affinity of the double helix in and around the genomic site. Our results revealed a symmetric pattern, whereby genomic sites with medium stability are more susceptible to Cas9 cleavage while sites at the extreme ends of the scale are significantly less so (p < 0.05 using a permutation test; S7 Fig). This feature, which correlates with other features concerning the local chromatin shape (Fig 4), is indicated to play an important role in predicting Cas9 efficacy. Such geometric features have been previously reported to affect binding of transcription factors and other DNA-binding proteins due to their contribution to the local shape of the double-helix [63,64]. To date, however, the contribution of these aspects to Cas9 affinity has not been explored. We postulate that highly rigid double stranded DNA (dsDNA) with high enthalpy prevents the Cas9 protein from melting the dsDNA and allowing the RNA/DNA duplex to form, while genomic sites with very low enthalpy tend to coil and block access of the enzyme. The learning dataset of CRISTA is based on genome-wide profiling of cleavage intensities of nuclear sites. Thus, targeted evaluation of nuclear sites that were pre-selected according to their similarity to a specified sgRNA could not be integrated within the learning dataset since they would bias the results towards certain features. Yet, those targets could be used as external validation to examine the performance of CRISTA on data that were not used for its training. To this end, datasets of targeted sequencing generated from two studies were examined. Cho et al. [15] analyzed the indel formation of 116 sites by 10 sgRNAs in the human genome using deep sequencing. Similarly, Wang et al. [22] examined 54 sites for two sgRNAs. Combined, these data provided 170 samples of on-targets, off-targets, and uncleaved sites (S2 Text). These datasets differ from the data that were used for the leave-one-sgRNA-out cross-validation procedure in two ways. First, cleavage sites were not detected in an unbiased manner, thus, cleavages of additional potential sites from the reference genome have not been validated and such ones could not be included as a set of uncleaved sites. Second, in contrast to the experimental systems used for our training dataset, the experimental systems used in the studies of Cho et al. and Wang et al. were not sensitive enough to differentiate among nuclear sites that were cleaved at low efficiencies [15,22]. Such sites, which were considered as ‘undetermined’ in the two studies, were marked with zero cleavage intensities for our validation procedure. Over the sets of 12 sgRNAs and their corresponding targets, CRISTA achieved an averaged Pearson r2 of 0.68, ROC-AUC of 0.7, and PRC-AUC of 0.72 (S9 Table; accuracy measurements of the four alternative tools for each dataset are denoted in S8 and S9 Figs). CRISTA, as well as the other three alternative tools, achieved lower accuracy measurements over the validation data in comparison to the leave-one-sgRNA-out cross-validation procedure. While CRISTA performed better than CCTop and the CFD score according to all three metrics, the averaged Pearson r2 obtained by OptCD (r2 = 0.92) was much higher than those of the other three scoring functions. This could be explained by the dichotomous nature of the OptCD score (see Fig 2, S8 and S9 Figs), which assigns a score of 1.0 to all on-targets and to some sites with a mismatch in unpenalized position, while assigning scores close to 0.0 to nearly all other targets. In contrast, the predictions made by CCTop, the CFD score, and CRISTA produce a more continuous scale. Consequently, assigning the ‘undetermined’ sites with zero cleavage intensities better matches scoring systems that highly penalize off-targets, like OptCD. CRISTA was developed for the assessment of the cleavage efficacy of a certain genomic target by a specific sgRNA. This assessment integrates two aspects that have been treated separately by currently available tools: those that are designed to predict off-target effects, and those that are aimed at ranking different sgRNAs according to their on-target efficiency. In contrast to the many computational tools that have been developed for these tasks, CRISTA accounts for wider genomic-related attributes in addition to sequence considerations. Additionally, CRISTA considers possible bulges within the DNA site or sgRNA, a concern that was mostly overlooked to date (but see [31,34]). Our results suggest that bulges are an integral part of the CRISPR system, as they are predicted to occur in approximately 20% of the targets in the evaluated dataset. While a large number of these are targets with low cleavage frequencies, several of them are cleaved at medium-to-high frequencies. These findings are in contrast to the conclusions of Haeussler et al. [27], who argued that bulges are rare and occur in targets that are cleaved at negligible frequencies. This discrepancy could partially be explained if certain combinations of mismatch-gap penalties are assumed when computing the pairwise alignment. While the relative importance of mismatches and bulges to Cas9 activity are underexplored, the experimental results presented by Lin et al. [12], Wang et al. [22], and Ran et al. [20] support our findings that bulges constitute an important component of the off-target spectrum. We showed that unbiased genome-wide methods for profiling CRISPR target sites generally comply with one another. The discrepancy in the results obtained with Frock data can be explained by the specificities of the HTGTS method [19]. In that study, two alternative approaches were presented: one using sgRNA-generated double strand breaks at on-targets to capture off-targets, and a second approach (termed “universal donor bait HTGTS”) uses known breaks of one sgRNA to capture targets of another. The latter technique was executed on two sgRNAs that were also examined in other studies, and hence belong to the ‘common guides’ set. For these two sgRNAs the predictions made by CRISTA using the leave-study-out procedure were similar to the results obtained for the other studies (Fig 3). In contrast, the sgRNAs that were examined using the first approach were all unique in our dataset. Our analysis demonstrated that the predictions of CRISTA on datasets obtained with this approach were not compatible with the other techniques. Possible explanations to this observation were previously described as bias for sites that are closer in proximity to the on-target [65,66], and we thus chose to eliminate Frock data from the training dataset of CRISTA. Besides the impact of some known attributes that are important to Cas9 action, namely, attributes that describe pairwise similarity and the nucleotide composition, our results highlight the importance of features that are associated with the DNA geometry, such as the DNA rigidity, double-helix groove width and DNA enthalpy. These attributes are usually used for predicting genomic elements, such as nucleosome organization and transcription factor binding sites, or for determining the optimal setting of empirical procedures (e.g., PCR). Here we found that these features are more influential for predicting CRISPR’s efficacy than measures that are based only on the DNA occupancy. Our findings suggest that integrating local DNA geometry and other genomic features could enhance the prediction and ranking of on-targets. To date, studies that analyzed large datasets of on-targets accounted for position-specific nucleotide identities to evaluate the cleavage efficacy of CRISPR-Cas9 efficacy, and used these to form predictive models [28,56,67–71]. We speculate that incorporating genomic features in the analysis of such data will enhance the ability to rank among on-targets. In addition, we did not find the features concerning the RNA thermodynamics to contribute much to the predictive model. However, the variance of these features in our dataset is low since they are clearly uniform for all samples of the same sgRNA. Possibly their importance will be highlighted when the efficacy of a large number of on-targets is examined. The CRISTA model described in this study was trained as a regression model, which was fitted to the (transformed) cleavage efficiencies reported in the experimental studies. One difficulty with this approach is the need to combine results from different experimental platforms into a single scale (as described in S1 Text)–a procedure which may bias the results. As an alternative, it is possible to analyze the data within a classification framework. Under such a setting, the data provided by genome-wide profiling of CRISPR-Cas9 could be interpreted as a binary outcome (i.e., all cleaved sites regarded as the set of positives while the uncleaved sites as the negatives). To assess the performance of the learning scheme under these two alternatives (i.e., regression and classification), we implemented a classification model using the Random Forest classification algorithm (S3 Text). Notably, the results obtained using the classification model were very similar—although slightly inferior—to those obtained using the regression model (S10 Fig). This might be expected since the regression model inherently accounts for the differential cleavage propensities among the cleaved sites, whereas the classification approach largely overlooks the complexity present in the experimental data. While it is possible to set a strict threshold on the cleavage propensities above which sites are considered as positives (in contrast to sites that were cleaved at low frequencies and might as well be considered as noise), this setting imposes the difficulty regarding the exact value of the threshold that should be chosen, and raises the question whether such a discretization process extracts the maximum amount of information from the experimental data. CRISTA was implemented using currently available data, which included published genome-wide profiling of off-targets by CRISPR-Cas9 (the learning dataset) and available predictive tools for feature extraction. The future development of CRISTA would benefit both from the further accumulation of genome-wide profiling of CRISPR-Cas9, as well as from additional features. In turn, an important benefit of CRISTA’s prediction framework is the ability to examine the contribution of various attributes. This use of CRISTA as a platform for hypothesis testing only entails that genome-wide assessment of the examined feature could be provided. A feature that is important for CRISPR-Cas9 mechanism of action would either be highly ranked, or ultimately increase the prediction accuracy. Genome engineering techniques have evolved rapidly since CRISPR-Cas9 first emerged, introducing alternative endonucleases for manipulating the genome. For example, manipulation of the active domains of the Cas9 enzymes to generate a single-strand break (Cas9-nickase; Cas9n [72,73]) requires targeting of two sites at opposing strands at once, thus yielding a complex with enhanced specificity. Structural biology has been employed to generate Cas9 variants by altering residuals that were identified to mediate the ability of Cas9 to cleave off-target sites, generating eSpCas9 (enhanced SpCas9 [21]) and SpCas9-HF1 (high fidelity SpCas9 [17]). In addition, SpCas9 homologs or other CRISPR endonucleases that differ in their PAM requirements, packaging size, and manner of action, including the Staphylococcus aureus Cas9 (SaCas9 [20]) and the class 2 CRISPR endonuclease, Cpf1 [74], were recently detected, and shown to reduce off-target effect. Nevertheless, Cas9 is still in wide use and protocols that rely on the use of the wild-type SpCas9 for genome engineering, therapeutics, and reverse-genetics have yet to be developed for its alternatives [75–77]. Notably, the learning scheme presented here is not reliant on any specific experimental system, granted this system is not biased towards specific regions of the genome. Thus, future genome-wide experiments can be easily integrated into the learning dataset, including those obtained with Cas9 variants and its orthologs, consequently revealing enzyme-unique characteristics. Taken together, while CRISTA was developed as an inferential tool, such a framework can be further employed to deepen our understanding and to shed light on future research of the CRISPR system.

10.1371/journal.ppat.1003609

Human Cytomegalovirus Gene UL76 Induces IL-8 Expression through Activation of the DNA Damage Response

Human cytomegalovirus (HCMV), a β-herpesvirus, has evolved many strategies to subvert both innate and adaptive host immunity in order to ensure its survival and propagation within the host. Induction of IL-8 is particularly important during HCMV infection as neutrophils, primarily attracted by IL-8, play a key role in virus dissemination. Moreover, IL-8 has a positive effect in the replication of HCMV. This work has identified an HCMV gene (UL76), with the relevant property of inducing IL-8 expression at both transcriptional and protein levels. Up-regulation of IL-8 by UL76 results from activation of the NF-kB pathway as inhibition of both IKK-β activity or degradation of Ikβα abolishes the IL-8 induction and, concomitantly, expression of UL76 is associated with the translocation of p65 to the nucleus where it binds to the IL-8 promoter. Furthermore, the UL76-mediated induction of IL-8 requires ATM and is correlated with the phosphorylation of NEMO on serine 85, indicating that UL76 activates NF-kB pathway by the DNA Damage response, similar to the impact of genotoxic drugs. More importantly, a UL76 deletion mutant virus was significantly less efficient in stimulating IL-8 production than the wild type virus. In addition, there was a significant reduction of IL-8 secretion when ATM -/- cells were infected with wild type HCMV, thus, indicating that ATM is also involved in the induction of IL-8 by HCMV. In conclusion, we demonstrate that expression of UL76 gene induces IL-8 expression as a result of the DNA damage response and that both UL76 and ATM have a role in the mechanism of IL-8 induction during HCMV infection. Hence, this work characterizes a new role of the activation of DNA Damage response in the context of host-pathogen interactions.

The importance of herpesviruses is evident by their prevalence in the human population and the diverse range of diseases that they provoke. Their ability to establish latency provides a compelling example of how herpesviruses successfully evade the immune system and manipulate cellular biology. One promising approach for the development of novel anti-viral strategies is to study viral proteins involved in these host-pathogen interactions. One example is the induction of the pro-inflammatory chemokine IL-8 by HCMV which enhances viral replication and dissemination of the virus by neutrophils. Here, we have identified HCMV UL76 gene, conserved in all herpesviruses, as an inducer of IL-8, and thus with an important impact on HCMV pathogenesis. The induction of IL-8 by UL76 results from activation of the DNA Damage response, which may also explain why UL76 also induces cell cycle arrest. These findings are a clear example of how viruses manipulate intracellular signaling pathways with different outcomes that will be beneficial for viral infection. Finally, the fact that UL76 is a non-homologous gene substantiates the premise that many such pathogen genes without homology may indeed have evolved for host manipulation, and are a repository of potential useful tools for experimental manipulation in health and disease.

Human cytomegalovirus (HCMV) is a β-herpesvirus that infects healthy individuals, usually asymptomatically, but can cause severe or fatal disease in immunocompromised individuals. Primary HCMV infection, as with other herpesviruses, is followed by establishment of lifelong latency with periodic reactivation and, in order to successfully establish itself in the host, the virus has evolved a broad range of host evasion strategies, modulating not only innate and adaptive immunity, but also host cell biology, for example, the cell cycle and apoptosis [1]. The induction of the interleukin-8 (IL-8) during HCMV infection is particularly important for viral replication and possibly contributes to the efficient dissemination of the virus by neutrophils [2], [3]. Interleukin-8 is a pro-inflammatory chemokine that attracts primarily neutrophils, and also monocytes and cytotoxic T cells, by interacting with the CXC chemokine receptors CXCR1 and CXCR2 [4]. Although expression of IL-8 is low or absent under normal conditions, it is highly inducible by a wide range of extracellular stimuli, such as the pro-inflammatory cytokine IL-1, the tumor necrosis factor alpha (TNFα) [5], bacteria and viruses [6], [7]. Besides its relevant role in inflammation, IL-8 is a key component in several viral infections, modulating viral dissemination and virus replication, in part due to inhibition of the impact of interferon-α [8]. On the other hand, excessive amounts of locally produced IL-8 can have deleterious effects, and so IL-8 gene expression is tightly controlled at both transcriptional and post-transcriptional levels. Activation of IL-8 expression in the majority of cell types is critically controlled by the NF-kB transcription factor. The AP-1 and NF-IL-6 transcription factors may also contribute to optimal IL-8 activation, depending on the stimulus or the cell type [4]. The NF-kB canonical pathway involves the activation of the IKK complex, consisting of two catalytic kinase subunits, IKKα and IKKβ, and a regulatory subunit, IKKγ/NEMO. In most unstimulated cells, NF-kB dimers (mostly p65/p50 dimers) are localized in the cytoplasm as a complex with the IkB proteins. Upon stimulation, IkB is phosphorylated by the IKK complex, ubiquitinated and targeted for degradation, thus releasing the NF-kB subunits that translocate to the nucleus and induce transcription of target genes [9]. Although most of the physiological inducers of NF-kB involve the canonical pathway, alternative mechanisms leading to NF-kB nuclear localization and DNA binding have been identified. One of these pathways is induced by activation of the DNA damage response and, in contrast to inflammatory stimuli such as TNFα or IL-1β, the signal originates in the nucleus [10]. Activation of NF-kB by genotoxic stress requires induction of two independent parallel pathways. The first one triggered upon DNA damage results in the phosphorylation and activation of ATM, a nuclear protein kinase which regulates cell cycle checkpoints in response to DNA double-strand breaks [11]. The second pathway leads to SUMOylation of NEMO through a mechanism dependent on PARP1, PIASy and Ubc9. Activation of both these pathways leads to phosphorylation and ubiquitination of sumoylated NEMO in an ATM-dependent way. Ubiquitinated NEMO associated with ATM is exported back to the cytoplasm, activating the IKK complex and subsequent NF-kB activation in a similar manner to the canonical pathway [10]. The HCMV UL76 protein is virion-associated and expressed with late kinetics [12]. The corresponding gene belongs to the UL24 gene family, conserved in all herpesviruses and the only core gene without an assigned function [13]. Bioinformatic analysis identified UL24 gene family as a putative novel PD-(D/E)XK endonuclease [14]. This superfamily of restriction endonuclease-like fold proteins includes several restriction endonucleases (e.g. EcoRI, EcoRII, BamHI, BglI, Cfr10I, NaeI), DNA repair enzymes (MutH and Vsr), Holliday junction resolvases (Hjc and Hje) and other nucleotide-cleaving enzymes [15]. However, no endonuclease activity has been demonstrated experimentally for any of the UL24 homologues. Global mutational analysis of the HCMV genome classified UL76 as an augmenting gene for viral replication [16], recently demonstrated to be involved in the regulation of the UL77 gene expression. Since UL77 is essential for viral replication, its regulation by UL76 may be important for efficient HCMV replication [17]. Expression of UL76 also induces cell cycle arrest at G2/M phase by inhibition of the mitotic Cdc2-cyclin B complex. Interestingly, this effect on the cell cycle is conserved in the UL24 human homologues representatives of the alpha, beta and gamma-subfamilies and the murine homologue from MHV-68 (ORF20) [18], [19]. The precise mechanism of cell cycle arrest induced by UL24 homologues remain to be clarified, but a recent report showed that HCMV UL76 induces chromosomal aberrations and DNA damage [20]. Here we identify a new function of UL76, the induction of IL-8 expression, mediated by the ATM kinase and activation of the NF-kB pathway. Thus, activation of NF-kB by UL76 results from induction of the DNA damage response, similar to genotoxic drugs. Importantly, induction of IL-8 by HCMV is significantly reduced in the absence of ATM or in normal fibroblasts infected with a HCMV UL76 deletion mutant. Thus, viral infection induces IL-8 in a similar manner to the UL76 gene alone and UL76 is essential for maximal activation of IL-8 by HCMV. The effect of UL76 on the activation of IL-8 transcription was demonstrated using a luciferase reporter construct containing the IL-8 promoter sequence. Transfection of a UL76 expression plasmid significantly activated transcription of IL-8 promoter in a dose-dependent manner (Fig. 1A). Furthermore, cells expressing UL76 were demonstrated to secrete significantly higher levels of IL-8 as compared to the control vector (p<0.01) (Fig. 1B). In conclusion, UL76 induces IL-8 expression at both the level of transcriptional activation and protein secretion. Expression of IL-8 is tightly regulated at the transcriptional level. The sequence of nucleotides -1 to -131 in the proximal promoter region of IL-8 gene is essential for its transcription regulation and contains binding sites for NF-kB, AP-1 and NF-IL-6 transcription factors (Fig. 2A) [4]. To determine the mechanism of induction of IL-8 by UL76, the luciferase activity of wild type IL-8 luciferase reporter was compared with its mutant derivatives containing a mutation in each of the three transcription factor binding sites. There was no significant difference in luciferase activity in response to co-transfection with UL76 when AP-1 or NF-IL-6 binding sites were mutated in the luciferase reporter construct, whereas IL-8 transcriptional activation was drastically reduced in the absence of the NF-kB binding site, indicating a critical role for the NF-kB pathway in the UL76-mediated induction of IL-8 (Fig. 2B). Consistent with the previous results, expression of UL76 significantly activated an NF-kB responsive promoter (Fig. 2C). To further characterize the activation of the NF-kB pathway by UL76, we used a catalytically inactive mutant IKKβ and a mutant IkBα (IkBαS32/36A) in which the two critical serine residues were mutated to alanine, thus no longer permitting its phosphorylation and degradation. Both constructs function as dominant negatives, inhibiting the activity of cellular wild type IKKβ and IkBα, respectively. Co-transfection of each dominant negative with IL-8 luciferase reporter and UL76 expression plasmid or control vector resulted in a reduced induction of IL-8 in cells expressing UL76 (Fig. 3A). After IkBα degradation by the proteosome, dimers of NF-kB subunits translocate to the nucleus where they bind to the target gene promoter region and activate transcription [9]. Thus, we evaluated the effects of UL76 expression on the subcellular localization of the NF-kB p65 subunit by immunofluorescence. As shown in Figure 3B, p65 was localized in the nucleus of the UL76-transfected HFF cells, in contrast to its cytoplasmic localization in control cells. This effect was not due to an increase in the expression of p65 (Fig. 3C). Furthermore, similarly immunoblotting of 293T nuclear extracts with anti-p65 antibody revealed an accumulation of p65 in the nucleus of cells expressing UL76 (Fig. 3D). This accumulation was specific, as can be seen from the constant levels of the nucleolin expression in the loading control. Consistent with these results, chromatin immunoprecipitation (ChIP) analysis demonstrated that expression of UL76 leads to NF-kB p65 binding to the IL-8 promoter (Fig. 3E). In summary, IL-8 induction by UL76 requires a functional IKKβ and the degradation of IkBα to promote translocation of p65 subunit to the nucleus where it activates IL-8 transcription. Expression of UL76 results in an increased number of double stranded DNA breaks and phosphorylation of γH2AX, indicating activation of the DNA Damage response [20]. Consistent with these results, expression of UL76 results in activation of ATM and consequent phosphorylation of p53 and H2AX proteins (Fig. 4A). Recently, several studies have characterized an alternative pathway to NF-kB activation that results from DNA damage. Based on this, we hypothesized that the ability of UL76 to induce DNA damage would lead to activation of NF-kB pathway and result in the induction of IL-8 expression. A characteristic feature of NF-kB pathway activation by genotoxic stress is the accumulation of IKKγ/NEMO in the nucleus where a series of post-translational modifications occurs [21]. The nuclear post-translational modifications of NEMO that are critical for NF-κB activation following genotoxic stress include ATM-independent sumoylation and ATM-dependent phosphorylation at serine 85 followed by monoubiquitination. Consistent with this mechanism, immunostaining using an anti-NEMO antibody revealed increasing amounts of nuclear NEMO in cells expressing the UL76-HA tagged protein (Fig. 4B). Furthermore, immunoblotting of similarly transfected cells with a specific antibody to NEMO(S85) phosphorylation demonstrated that expression of UL76 induces phosphorylation of NEMO as previously observed after genotoxic stress (Fig. 4C). To evaluate the impact of the ATM kinase in IL-8 induction by UL76, we used two different approaches: a specific ATM inhibitor, KU55933, and a human fibroblast cell line deficient in ATM. The amount of IL-8 secreted by 293T cells expressing UL76, or the control plasmid, in the presence or absence of KU55933 was determined by ELISA. Inhibition of ATM by KU55933 blocked the UL76-induced IL-8 secretion (Fig. 4D). Similarly, IL-8 concentration was determined in supernatants of ATM -/- cells expressing UL76 or the control plasmid. As shown in Figure 4E, UL76 or etoposide, a genotoxic drug, are unable to induce IL-8 in the absence of ATM. Although UL76 expression leads to higher levels of IL-8 than etoposide stimulation, there is no increase in IL-8 secretion when compared to control vector, so this basal induction is possibly due to transfection. Moreover, this result is not due to the incapacity of the cell line to produce IL-8 since stimulation with TNFα, a membrane receptor-triggering NF-kB canonical pathway independent of ATM, is still capable of inducing IL-8 secretion (Fig. 4E). Expression of UL76 was not affected in the ATM -/- cell line or in 293T cells cultured in the presence of the ATM inhibitor as confirmed by western blot (Fig. 4D,E). In summary, these results indicate that activation of NF-kB pathway and consequent IL-8 induction by UL76 are ATM-dependent and result from activation of DNA damage. There is clear evidence that UL76 activates the DNA damage response, however, the mechanism employed by UL76 for this activation is still unknown. The prediction that the UL24 gene family encodes a novel PD-(D/E)XK endonuclease is a possible explanation [14]. Comparison of UL24 gene family sequences identified the three conserved PD-(D/E)XK signature amino acids of the endonuclease motif which are conserved in all homologues (Fig. 5A) [14]. A UL76 gene with these three critical amino acids mutated was constructed and used to determine the impact of the putative endonuclease activity on IL-8 induction. Levels of IL-8 secreted by cells expressing the mutant UL76 gene were reduced when compared to wild type UL76 gene; however, they were still significantly higher than control vector-expressing cells (Fig. 5B). These results indicate that the putative endonuclease activity is not essential for the induction of IL-8. In order to evaluate the impact of UL76 on the up-regulation of IL-8 in the context of HCMV infection, we used a previously described UL76 transposon mutant HCMV [16]. Supernatants from human fibroblasts infected with wild type HCMV AD169 BAC or UL76 mutant virus (TNUL76) were collected at the indicated time points and secreted IL-8 was determined by ELISA. Consistent with previous studies [3], [22], HCMV infection resulted in high levels of IL-8 secretion during the course of the experiment (Fig. 6A). Induction of IL-8 in cells infected with the UL76 mutant virus, however, was significantly reduced. At each time point, the amount of IL-8 secreted by cells infected with UL76 mutant virus (TNUL76) was reduced by 42–52% compared with wild type HCMV. Equal infection by both viruses was confirmed by levels of the HCMV immediate-early 1 (IE1) protein. Thus, UL76 is essential for optimal induction of IL-8 by HCMV. There was, however, no difference in the phosphorylation of ATM, NEMO and IkB proteins in cells infected with HCMV wild type compared to UL76-deficient virus (data not shown). As the mutation of UL76 significantly increases the level of UL77 protein expression [17], the effect of UL77 in the expression of IL-8 was evaluated by ELISA. In contrast to UL76-transfected cells, there is no induction of IL-8 in cells expressing UL77. Moreover, UL77 has no inhibitory effect in the induction of IL-8 by different stimuli (Fig. S1). Overall, these results demonstrate that UL77 is not able to modulate IL-8 expression and thus, the reduction of IL-8 levels in cells infected with the UL76 mutant virus (TNUL76) is not due to the regulation of UL77 expression by UL76. A previous deletion mutant analysis of the IL-8 promoter in monocytic cells has shown that AP-1 and NF-kB transcription factors were required for optimal induction of IL-8 by HCMV [22]. The precise mechanism used by HCMV to induce IL-8 expression, however, is still not clear. As UL76 is required for maximal IL-8 induction by HCMV (Fig. 6) and HCMV infection activates ATM [23], [24], we hypothesized that ATM would also have a role in IL-8 up-regulation during viral infection. To test this hypothesis, a primary fibroblast ATM -/- cell line was infected with HCMV AD169 BAC virus. Supernatants were collected at the indicated time points and IL-8 concentration was determined by ELISA. When compared to normal human fibroblasts (Fig. 6), the amount of secreted IL-8 was significantly reduced in the HCMV infected ATM -/- cells (16,79 vs 3,63 fold induction) (Fig. 7). Similar results were obtained with a transformed ATM -/- fibroblast cell line (data not shown). Infection with HCMV was confirmed by the presence of the viral protein UL44 (Fig. 7, below). Similar to wild type HCMV, ATM -/- cells infected with the UL76 deficient virus (TNUL76) secreted lower levels of IL-8 compared to HFF infected cells. The levels of IL-8 of cells infected with TNUL76 were, however, even lower than the observed in ATM -/- cells infected with wild type virus (Fig. S2). Collectively, these results indicate that during HCMV infection, UL76, and possibly other gene(s), induces IL-8 expression, at least in part, through activation of ATM. The UL24 gene family is one of the approximately 40 core genes that are conserved in all three herpesviruses subfamilies and the only one which still has no assigned function [13]. Previously, functional assays demonstrated that all homologues of UL24 gene family induce cell cycle arrest [18], [19]. This work identifies another, and at first sight, apparently unrelated function of the UL24 homologue from HCMV (UL76), the induction of the expression of IL-8. Further exploration of the mechanism, however, suggests that both activities may result from viral activation of the DNA Damage response. Deletion mutant analysis of the IL-8 promoter demonstrated that UL76 up-regulates IL-8 expression through activation of NF-kB pathway requiring a functional IKKβ and degradation of the IkB protein. Moreover, expression of UL76 resulted in the translocation of the NF-kB p65 subunit to the nucleus and its binding to the IL-8 promoter. These events, characteristic of the canonical NF-kB pathway, typically occur in the cytoplasm and are usually activated by membrane-receptor stimulation [9]. Thus, the exact mechanism of how UL76 activates the NF-kB pathway is an interesting paradox as UL76 is a nuclear protein. In recent years an alternative mechanism of activation of NF-kB pathway triggered by genotoxic stress has been described. In contrast to inflammatory stimuli such as TNFα or IL-1β, the signal originates in the nucleus [10]. Since it has been shown that UL76 is able to induce double strand breaks, and consequently activate DNA damage [20], we hypothesized that UL76 might induce IL-8 expression as result of the DNA damage response. Indeed, and similar to the effect of genotoxic drugs such as etoposide, expression of UL76 resulted in an accumulation of nuclear NEMO and its activation (phosphorylation at serine 85). These findings suggest that the ATM kinase might play a role in the IL-8 induction by UL76 and thus, the predicted role of ATM was demonstrated by two strategies, one using a specific ATM inhibitor and the other employing an ATM knockout cell line. Abrogation of UL76-mediated induction of IL-8 occurred with both approaches. These observations indicate that induction of IL-8 by UL76 originates in the nucleus as a result of the DNA Damage response. The exact mechanism of activation of the DNA Damage response induced by UL76 is still not clear. A promising clue that we pursued was the identification of conserved putative PD-(D/E)XK endonuclease motifs in the UL24 gene family [14]. Its conservation suggests a critical role in the function of this gene family, thus, the DNA damage activation by UL76 could be a direct effect of its endonuclease activity. Mutation of the three predicted endonuclease signature amino acids was, however, inconclusive as it resulted in a reduction rather than an abolition of IL-8 induction. It is possible that the observed reduction may be related with an unknown function of these conserved domains rather than loss of a putative endonuclease activity. In fact, the bioinformatically predicted endonuclease motif might not be related to a functional endonuclease activity, as no endonuclease activity has been demonstrated experimentally for any of the UL24 homologues. Importantly, UL76 has a critical role in the up-regulation of IL-8 during HCMV infection as demonstrated by the significant reduction of secreted IL-8 in cells infected with an UL76 deletion mutant virus. This result is particularly important as IL-8 enhances HCMV replication and contributes to the efficient viral dissemination by neutrophils [2], [3]. The only HCMV gene that has been described as an activator of IL-8 is the Immediate Early 1 gene (IE1) [22]. Thus, the incomplete inhibition of IL-8 secretion by HCMV observed in the absence of UL76 may be due to the effect of IE1 gene. The existence of other gene(s) that may also contribute for HCMV-induced IL-8 expression, in addition to IE1 and UL76, is not excluded. Experiments to observe a similar impact of UL76 in the signaling pathway at the level of virus infected cells were negative, possibly due to alternative virus strategies activating these proteins. We emphasize, however, that the key observation is the diminished expression of IL-8 induced by the UL76 deficient virus, which clearly demonstrates a role for UL76 in the up-regulation of IL-8 in HCMV infected cells. The relevance of IL-8 in HCMV life cycle is emphasized by the fact that the HCMV UL146 gene encodes a homologue to CXC chemokines such as IL-8 (vCXCL1), which functions as a selective agonist for CXCR2 and, with lower affinity, for CXCR1 [25], [26]. The IL-8 production observed in cells infected with wild type HCMV or UL76 mutant virus, however, is independent of the presence of the viral CXCL1, as this gene is deleted from the HCMV AD169 strain used in this work [16]. Induction of IL-8 expression by HCMV requires activation of the NF-kB pathway [22]. Thus, one objective of this work was to elucidate the mechanism of NF-kB activation by HCMV that leads to IL-8 production. Here we demonstrate that ATM also has a critical role in the induction of IL-8 by HCMV as infection of ATM -/- cells with wild type HCMV resulted in considerably lower levels of secreted IL-8 compared to the similar infection of normal human fibroblasts. On the other hand, the incomplete inhibition of IL-8 expression in ATM-/- cells infected with HCMV suggests that other NF-kB pathways are involved. It is possible that these are not redundant effects, but activation of different NF-kB pathways, possibly through different viral proteins, may be necessary for the induction of optimal levels of IL-8 by infected cells that will be beneficial for HCMV replication as previously reported [2]. It may be significant that a major reduction in viral replication is observed when normal cells are infected with a UL76 deficient virus [16] or when ATM deficient cells are infected with HCMV wild type virus [27]. It is possible that this defect in viral replication is associated with the reduced IL-8 levels observed in cells infected with UL76 deficient HCMV or in ATM deficient cell infected with wild type HCMV. Supporting this hypothesis is the fact that IL-8 enhances HCMV replication [2]. In summary, the non-homologous UL76 gene of HCMV has not only evolved for manipulation of the host cell cycle, but also activates expression of the pro-inflammatory chemokine IL-8. Both of these activities appear to depend on activation of pathways triggered as a result of the DNA Damage response and may favor propagation of the virus. The fact that, in recent years, several viruses have been demonstrated to activate the DNA Damage response raised new important questions. It is not known if this activation results from recognition of DNA damage or if it is due to the recruitment of DNA repair proteins observed during viral infections such as HCMV. Furthermore, it is not completely understood how the activation of DNA Damage pathway is beneficial for viral replication. Our present work establishes a new role of the induction of DNA Damage response in the context of viral infection that may help to elucidate some of these questions, as it demonstrates how viruses exploit the complex crosstalk that occur between different cell signaling pathways. Human embryonic kidney 293T cells were cultured in 5% CO2 in Dulbecco's Modified Eagle's Medium (Gibco) supplemented with 10% fetal calf serum (Gibco) at 37°C. Human foreskin fibroblasts (HFF) (obtained from European Collection of Cell Cultures), a transformed (GM09607) and a primary (GM01588) A-T human fibroblast cell lines (obtained from the Coriell Institute for Medical Research) were cultured in Minimum Essential Medium with Earle's salts supplemented with 10% fetal calf serum (Gibco). The UL76 and UL77 gene from HCMV AD169 were cloned into pcDNA3 plasmid fused in frame with an amino-terminal influenza haemaglutinin peptide (HA) tag. The three putative endonuclease amino acids in the UL76 gene were mutated to glycine (pcDNA3HA-E/K mut plasmid) according to the Directed Mutagenesis kit protocol (Stratagene). The luciferase reporter constructs containing human IL-8 promoter (-131) or a mutation in the NF-kB, AP-1 or NF-IL-6 binding site were a gift from Dr Naofumi Mukaida and have been described before [22]. The reporter plasmid for NF-κB [p(PRD2)5tkΔ(-39)lucter] was a gift from Dr Steve Goodbourn. Dominant negative mutants of IKKβ and IkBα (S32/36A) plasmids containing an HA tag, were obtained from Dr Michael Karin [28] and Dr Dean Ballard [29], respectively. The pCMVβ plasmid contains a β-galactosidase gene under the control of human cytomegalovirus immediate early promoter. The HCMV laboratory strain AD169 bacterial artificial chromosome (BAC) DNA was obtained from Dr Ulrich Koszinowski. The UL76 mutant virus (TNUL76), a gift from Dr Thomas Shenk, was generated by site-directed transposon mutagenesis of HCMV AD169 BAC and has been previously described [16]. Wild type or UL76 mutant virus BAC DNA were transfected in HFF cells by electroporation. Supernatants of transfected cells were collected and used for virus stock production. To prepare virus stocks of wild type AD169 BAC virus and TNUL76 mutant virus, HFF cells were infected at a multiplicity of infection (MOI) of 0.01. After virus adsorption for one hour, infected cells were cultured at 37°C and medium was collected every three days. Pre-cleared supernatants were centrifuged two hours at 12000 rpm at room temperature. Virus aliquots were stored at −80°C. Virus stock titers were determined by plaque assay. Briefly, HFF cells were cultured with 10-fold dilutions of virus suspension and allowed to absorb for 1 h. Cells were then cultured with complete medium containing 10% carboxymethylcellulose (CMC) for 10–15 days. Cellular monolayers were fixed in 4% paraformaldehyde and stained with 0.1% toluidine blue. Quantification of the viral plaques was performed using a dissecting microscope. 293T cells were co-transfected in triplicate with 100 ng of IL-8 luciferase reporter plasmid or luciferase reporter constructs containing mutations in the IL-8 promoter (ΔNF-kB, ΔAP-1 and ΔNF-IL-6), 25 ng of β-galactosidase internal control plasmid (pCMVβ) and 300 ng of pcDNA3 or pcDNA3HA-UL76, according to the Lipofectamine 2000 (Invitrogen) protocol. A similar transfection protocol was performed using the NF-kB luciferase reporter plasmid. Cells were lysed 28 h–30 h post-transfection and the luciferase activity was measured using the luciferase assay system (Promega) according to the manufacturer's protocol. β-galactosidase activity was measured using the Galacton-Plus kit from Tropix (Bedford, MA). The luciferase activity was normalized relative to the β-galactosidase activity of each sample as control of transfection efficiency. Supernatants of 293T cells or ATM -/- (GM09607) transfected with pcDNA3 (negative control), pcDNA3HA-UL76 or pcDNA3HA-UL77 plasmids were collected at 48 h post-transfection. As control, cells were stimulated with etoposide (10 µM) (Sigma), TNFα (10 ng/ml) (Peprotech) or IL-1β (1 ng/ml) (Cell Signalling) for 5 h. The concentration of IL-8 secreted was determined using an IL-8 ELISA kit (BD Biosciences) following the manufacturer's instructions. Similarly, supernatants of HFF or ATM -/- (GM01588) cells infected with wild type HCMV or UL76 mutant HCMV (TNUL76) at a MOI of 3, or mock-infected, were harvested at the indicated time points and clarified by centrifugation before quantification of IL-8 by ELISA. For ATM inhibition experiments, ATM inhibitor KU55933 (10 µM) (Calbiochem) was added to cells 1 h before transfection or infection with HCMV and was maintained in the medium during the experiment. Plates were analyzed at 450 nm using a BioRad ELISA Reader (BioRad) and levels of IL-8 were determined by comparison to a standard curve. The 293T or HFF cells were cultured on sterile glass coverslips and transfected with pcDNA3HA-UL76 or control pcDNA3 plasmid according to the Lipofectamine 2000 (Invitrogen) protocol. As positive control cells were stimulated with recombinant human TNFα (20 ng/ml) (Peprotech) for 30 minutes. At the indicated times post-transfection, cells were washed with PBS and fixed with 4% paraformaldehyde for 20 minutes. Fixed cells were permeabilised with PBS-0.1% Triton X-100 for 20 minutes. After washing, the cells were blocked with PBS-0.05% Tween 20 containing 5% normal goat serum for one hour. The samples were incubated with a mouse monoclonal anti-p65 (F-6) or anti-IKKγ/NEMO (B-3) antibody (Santa Cruz Biotechnology) followed by incubation with goat anti-mouse Texas Red secondary antibody (Molecular Probes) and rat monoclonal anti-HA-FITC conjugated antibody (Roche) to visualize UL76 HA-tagged protein. After incubation with DAPI, the coverslips were mounted in “Slow Fade” (Invitrogen) and images were acquired with a DeltaVision microscope (Applied Precision/Olympus). 293T cells were transfected with pcDNA3HA-UL76 or control pcDNA3 plasmid and nuclear extraction was performed using a Nuclear Extraction Kit according to the manufacturer's indications (Active Motif). Briefly, at the indicated time points post-transfection, cells were collected in ice-cold PBS in the presence of phosphatase inhibitors. Cytoplasmic extracts were obtained by resuspending the cells in hypotonic buffer followed by addition of detergent. After centrifugation the pelleted nuclei were lysed and nuclear proteins were solubilized in the lysis buffer supplemented with a protease inhibitor cocktail. Protein concentrations were determined by Bradford assay (Bio-Rad Laboratories). Total lysates from cells transfected with pcDNA3HA-UL76 or control pcDNA3 plasmid were prepared using lysis buffer supplemented with a mixture of protease and phosphatase inhibitors (Calbiochem), for 30 minutes on ice. Protein concentrations were determined by Bradford assay (Bio-Rad Laboratories). Proteins from total or nuclear lysates were separated by sodium dodecyl sulfate polyacrylamide gel electrophoresis (SDS-PAGE) and transferred to polyvinylidene difluoride (PVDF) membrane (GE Healthcare). Membranes were blocked with 5% nonfat milk for one hour at room temperature. Primary antibodies used were: mouse monoclonal anti-p65 (F-6), mouse monoclonal anti-IKKγ/NEMO (B-3), rabbit anti-nucleolin/C23 (H-250), mouse monoclonal anti-p53, rabbit anti-ATM, mouse anti-IE1 HCMV, mouse anti-pp52 (UL44) (Santa Cruz Biotechnology), rabbit anti-IKKγ/NEMO(S85) (Assay Biotech), rabbit anti-phospho-Histone H2AX(Ser139), mouse monoclonal anti-phospho-p53(Ser15), mouse monoclonal anti-phospho-ATM(Ser1981) (Cell Signaling), mouse monoclonal anti-β-actin, anti-HA and anti-tubulin (Sigma). Mouse monoclonal anti-β-actin and rat monoclonal anti-HA horseradish peroxidase-conjugated antibodies were purchased from Sigma. IRDye 800CW anti- mouse and anti-rabbit antibodies were purchased from Li-Cor Biosciences. Immunoblots were developed by enhanced chemiluminescence detection according to the manufacturer's instructions (ECL, Thermo Scientific Pierce) or using the Odyssey Infrared Imaging System (Li-Cor; Lincoln, NE). Densiometry analysis was performed using ImageJ software or Image Studio Lite Analysis Software (Li-Cor). 293T cells were transfected with pcDNA3HA-UL76 or control plasmid according to the Lipofectamine 2000 (Invitrogen) protocol. Thirty hours post-transfection, cells were cross-linked with 1% formaldehyde (Calbiochem) for 10 minutes at room temperature. After washing with PBS, cells were resuspended in SDS lysis buffer with protease inhibitor cocktail (Sigma) and chromatin was sheared by sonication. Immunoprecipitation was performed overnight at 4°C, using 2 µg of rabbit polyclonal anti-NF-kB p65 (A) or control IgG antibody (Santa Cruz Biotechnology). After incubation with protein G magnetic beads (Dynabeads, Invitrogen) for one hour at 4°C, immunocomplexes were washed and eluted. The cross-linking was reversed by heating at 65°C for 4 h. Chromatin-associated proteins were digested with proteinase K and DNA was purified by QIAGEN PCR purification kit following manufacturer's protocol. Immunoprecipitated DNA was quantified by real-time quantitative PCR using SYBR Green Master Mix (Applied Biosystems) and primer pair spanning the human IL-8 promoter region from −121 to +61: sense 5′-GGGCCATCAGTTGCAAATC -3′ and antisense 5′-TTCCTTCCGGTGGTTTCTTC-3′. Primers targeting the genomic region from −1042 to −826 of the IL-8 gene were used as negative control region: sense 5′-AACAGTGGCTGAACCAGAG-3′ and antisense 5′-AGGAGGGCTTCAATAGAGG-3′. Data were shown as mean values with standard deviation (SD). Differences between experimental groups were determined by a two-tailed Student t test using GraphPad Prism 5 software.

10.1371/journal.pcbi.1000194

The Inactivation Principle: Mathematical Solutions Minimizing the Absolute Work and Biological Implications for the Planning of Arm Movements

An important question in the literature focusing on motor control is to determine which laws drive biological limb movements. This question has prompted numerous investigations analyzing arm movements in both humans and monkeys. Many theories assume that among all possible movements the one actually performed satisfies an optimality criterion. In the framework of optimal control theory, a first approach is to choose a cost function and test whether the proposed model fits with experimental data. A second approach (generally considered as the more difficult) is to infer the cost function from behavioral data. The cost proposed here includes a term called the absolute work of forces, reflecting the mechanical energy expenditure. Contrary to most investigations studying optimality principles of arm movements, this model has the particularity of using a cost function that is not smooth. First, a mathematical theory related to both direct and inverse optimal control approaches is presented. The first theoretical result is the Inactivation Principle, according to which minimizing a term similar to the absolute work implies simultaneous inactivation of agonistic and antagonistic muscles acting on a single joint, near the time of peak velocity. The second theoretical result is that, conversely, the presence of non-smoothness in the cost function is a necessary condition for the existence of such inactivation. Second, during an experimental study, participants were asked to perform fast vertical arm movements with one, two, and three degrees of freedom. Observed trajectories, velocity profiles, and final postures were accurately simulated by the model. In accordance, electromyographic signals showed brief simultaneous inactivation of opposing muscles during movements. Thus, assuming that human movements are optimal with respect to a certain integral cost, the minimization of an absolute-work-like cost is supported by experimental observations. Such types of optimality criteria may be applied to a large range of biological movements.

When performing reaching and grasping movements, the brain has to choose one trajectory among an infinite set of possibilities. Nevertheless, because human and animal movements provide highly stereotyped features, motor strategies used by the brain were assumed to be optimal according to certain optimality criteria. In this study, we propose a theoretical model for motor planning of arm movements that minimizes a compromise between the absolute work exerted by the muscles and the integral of the squared acceleration. We demonstrate that under these assumptions agonistic and antagonistic muscles are inactivated during overlapping periods of time for quick enough movements. Moreover, it is shown that only this type of criterion can predict these inactivation periods. Finally, experimental evidence is in agreement with the predictions of the model. Indeed, we report the existence of simultaneous inactivation of opposing muscles during fast vertical arm movements. Therefore, this study suggests that biological movements partly optimize the energy expenditure, integrating both inertial and gravitational forces during the motor planning process.

In order to perform accurate goal-directed movements, the Central Nervous System (CNS) has to compute neural commands according to the initial state of the body, the location of the target, and the external forces acting on the limbs. Arm movement planning requires solving redundancy problems related to angular displacements, joint torques, muscular patterns, and neural inputs [1]. Experimental studies reported stereotypical kinematic features during pointing and reaching arm movements (e.g., quasi-straight finger paths, bell-shaped finger velocity profiles [2]–[4]). These features were found to be robust despite changes in mass, initial/final positions, amplitudes, and speeds of displacements [5]–[9]. Therefore, many studies have attempted to identify the principles of motion planning and control, hypothesizing that movements were optimal with respect to some criteria. The present article addresses the question whether motor planning is optimal according to an identifiable criterion. A promising approach to answer this question, called inverse optimal control, is to record experimental data and try to infer a cost function with regard to which the observed behavior is optimal [10]. In the theory of linear-quadratic control, the question of which quadratic cost is minimized in order to control a linear system along certain trajectories was already raised by R. Kalman [11]. Some methods allowed deducing cost functions from optimal behavior in system and control theory (linear matrix inequalities, [12]) and in Markov decision processes (inverse reinforcement learning, [13]). In the field of sensorimotor control and learning, some authors suggested that motor learning results from the optimization of some “loss function” related to the task (e.g., pointing accuracy) providing, therefore, a technique allowing to measure such function from experimental data [14]. Nevertheless, in most optimal control studies focusing on arm movements, a cost function is chosen and used in a mathematical model to check its validity a posteriori by comparing the theoretical predictions to the experimental observations. Kinematic models include minimum hand acceleration [15] and minimum hand jerk criteria [16]. These models produce horizontal arm movements that globally fit well with experimental data, providing smooth symmetric velocity profiles and straight trajectories in space. Dynamic models include minimum torque-change [17] and minimum commanded torque-change [18] criteria. They also accurately reproduce certain types of movements (point-to-point and via-point movements performed in the horizontal plane) but in several cases provide non-realistic double-peaked speed profiles (see for instance Figure 11 in [19]). In the Riemannian geometry framework, a model used geodesics to separately determine the geometrical and temporal movement features, allowing therefore a unification of previous computational models [19]. Specifically, the geodesic model accurately predicts the spatiotemporal features of three dimensional arm movements. However it results in hand paths that are excessively curved for planar movements. Additional criteria have also been considered, such as energy-like criteria [20]–[25] and effort related criteria [26], which minimize the peak value of the work, the metabolic energy expenditure, or the amount of neural control signals necessary to drive the arm. These models quantitatively reproduce some specific features of reaching and grasping, such as trajectories, velocity profiles, or final postures. Stochastic models, which are grounded on the hypothesis that noise in the nervous system corrupts command signals, have also been proposed. The minimum variance model was aimed at minimizing endpoint errors and provides not only accurate simulated trajectories of both eye saccades and arm pointing movements in the horizontal plane, but also the speed-accuracy trade-off described by Fitt's law [27]. In the optimal feedback control theory, noise is assumed to induce movement inaccuracy. If errors interfere with task goals, then the controller corrects deviations from the average trajectory. Otherwise the errors are ignored and, thus, variability in task-irrelevant dimensions is allowed [28]–[30]. Despite extensive literature concerning direct optimal control of arm movements, the hypotheses seem too restrictive in some models. For instance, in several models [19],[26], the static (gravity-related) and dynamic (speed-related) torques are calculated separately; therefore their predictions are independent from the gravity field. This assumption partly relies on the physiological observations that muscle activity patterns show two components: a tonic one (gravity-related) and a phasic one (speed-related) [31],[32]. Nevertheless, some authors reported difficulties in solving optimal control problems while taking into account gravitational forces in the optimization process [33],[34]. Thus, this assumption was also aimed at simplifying computations. Furthermore, the models previously cited are generally not consistent with the observation that the kinematics of arm movements performed in the sagittal plane depends on the direction with respect to gravity (i.e., upward versus downward movements) [35]–[38] whereas such a directional difference is significantly attenuated in microgravity [39]. A possible explanation of these findings would be that the CNS uses the gravity to move the limbs efficiently, rather than simply offset it at each instant. This idea guided the development of the theoretical model presented here. During a movement, the energetic consumption is related to the work of muscular forces. However, work is a signed physical quantity that may cancel itself out, even though both active and resistive forces consume energy in muscles. Therefore, work has to be always counted positive in order to express the energy expenditure of a movement: this is the absolute work of forces. The problem of minimizing this absolute work was never solved previously, despite its apparent simplicity and its potential interest for neurophysiologists. A reason might be the mathematical difficulty due to the non-differentiability of the cost function (induced by the absolute value function). Thus, while most existing models deal with smooth cost functions (i.e., functions that have continuous derivatives up to some desired order), this study relies on this non-smoothness property. The cost chosen here includes two terms: the first represents the absolute work and the second is proportional to the integral of the squared acceleration. In this article, two theoretical results are reported. Firstly, an “Inactivation Principle” states that minimizing a cost similar to the absolute work implies the presence of simultaneous inactivation of both agonistic and antagonistic muscles acting on a joint during fast movements. Secondly, a reciprocal result is that the presence of such inactivation along optimal trajectories implies the non-smoothness of the cost function. Therefore, by using transversality arguments from Thom's Differential Topology [40], Pontryagin's Maximum Principle [41], and Non-smooth Analysis [42], an equivalence between the non-smoothness of the cost function and the presence of simultaneous inactivation of both agonistic and antagonistic muscles is established. The proposed model permits to simulate accurately the kinematics of fast vertical arm movements with one, two, and three degrees of freedom. Moreover, experimental observations actually show simultaneous silent periods on the electromyographic (EMG) signals of opposing muscles during fast arm movements. The main results of this study are presented in the next two subsections. The theoretical analysis is exposed in the first subsection. In order to check the model, features of human arm movements were measured and are compared with the model predictions in the second subsection. The current subsection summarizes the mathematical theory which is more fully presented in the Materials and Methods Section. The reader who may not be interested in the full mathematical development of the model may read this subsection only, as a general survey. Although human vertical arm movements are studied here, the above theoretical results may apply to locomotion, whole-body reaching, and more generally to any mechanical system described in the Mathematical Theory Subsection. Firstly, we show that minimizing the compromise Aw/Ae is consistent with temporal and spatial features of biological arm movements. Secondly, we report simultaneous inactivation of agonistic and antagonistic muscles during arm movements. This suggests that the proposed criterion is also relevant at the muscular level and gives insights concerning the cost minimized during fast arm movements. Limb movement planning theory, presented in this study, focuses on fast, open-loop, vertical arm movements, and is based upon the assumption that such movements are optimal with respect to a certain integral cost. Within this framework, the question was to characterize possible cost functions. A model that minimizes a cost based upon the absolute work (i.e., an energetic optimality criterion) has been shown to allow simulating plausible arm movements in the sagittal plane. This was checked by means of three relevant kinematic features: fingertip path curvature, asymmetry of fingertip velocity profiles, and final arm posture. Since this cost function is non-smooth, the Inactivation Principle can be stated: for a large class of non-smooth cost functions, the net torque acting on a joint is zero during a short period occurring around the mid-path movements that are sufficiently rapid. This principle is also valid if a pair of agonistic-antagonistic actuators is considered, exerting opposite torques. Each of the torques is zero during an inactivation period which still appears if the biomechanics of the muscles is considered, when response times are brief (a few tens of milliseconds). For longer response times, complete inactivation is progressively replaced by low-levels of muscular activities. Such quiet periods in the EMGs of opposing muscles were observed during fast arm movements (see Figures 4, 5, and 6), which suggests that this optimality criterion is suitable. The suitability of a similar non-smooth cost function was also found for animals in a recent study [46]. The author concludes that the locomotor pattern of legged animals is optimized with respect to an energetic cost based upon the “positive work” of forces. However, the direct optimal control approach does not prove that the motor planning process actually minimizes energy expenditure. It just shows that such a criterion is plausible because it provides realistic behavior. Indeed, several other cost functions or theories may lead to similar results. For instance, muscle inactivation was also interpreted as a consequence of the Equilibrium Point hypothesis [47]. According to this interpretation, the threshold position control and the principle of minimal interaction would, together, determine the “Global EMG minima” which appear simultaneously in all muscles during rhythmic movements, near the point of direction reversals. Nevertheless, in the theory proposed here, inactivation is somewhat different: it appears near the time of peak velocity, and the precise interval of inactivity may be different at different joints. Moreover, inactivation is still predicted even if biomechanics of muscles, inertia and external forces are taken into account, which is not the case in Equilibrium Point theory [47]. Alternatively, it could be also considered that the CNS simply activates and deactivates the muscles, explicitly determining inactivation phases. However, this would be an argument against our main assumption that the brain tries to minimize some costs. Here, under this assumption, inactivation provides information on the cost function. The theoretical results also allow us to characterize the non-smoothness of the cost function once the simultaneous inactivation of opposing muscles is measured in practice, during movements presumed as optimal. Using mathematical transversality arguments from differential topology we proved that the minimization of an absolute-work-like cost during arm movements is a necessary condition to obtain inactivation phases along optimal trajectories. In other words, assuming that human movements are optimal with respect to a certain integral cost, the simultaneous inactivation of muscles that we observed provides evidence for an absolute-work-like cost. Notably, this simultaneous inactivation of opposing muscles, which is a singular phenomenon, cannot be predicted by models using smooth cost functions, such as the minimum endpoint variance [27], the minimum jerk [16], or the minimum torque-change [17]. Those models would predict deviations from “zero torque”, whereas singularity analysis proves the existence of an exact inactivation period. Simultaneous inactivation periods also appeared on intra-muscular EMG traces recorded from monkeys when performing horizontal arm movements (see Figure 5 in [48]). These findings suggest that the minimization of the energy expenditure may be a basic motor principle for both humans and animals. It should be emphasized that such an equivalence between specific movement features and well-identified properties of the cost function is not common in studies using optimal control approach for movement planning. The simulated movements replicated the experimental records accurately, except, obviously, for the bang-bang command signals which provide non-zero accelerations at the beginning and end of the movement (see Figure 5). The patterns of motor command are actually smoothed by the biomechanical characteristics (low-pass filters) of the muscles. As pointed out by several authors some models have been rejected hastily due to the lack of biological validity of their optimal solutions (bang-bang behaviors) [15],[49]. This problem was also discussed in a study where the authors used a similar non-smooth cost function based upon the “positive work” of forces [23]. They noticed that the abrupt velocity profiles predicted by their model were non-realistic but might actually be smoothed by modeling muscles dynamics. In fact, depending on the precision of modeling, different conclusions may be drawn. This is illustrated in Figure 1 where gradient constraints on the torques lead to smoother motor patterns whereas Figure 10 shows solutions in a simpler case of torque control. In the first case the acceleration is continuous while in the second case the acceleration jumps at the initial and final times (to make the transition between posture and movement). Nevertheless, in both cases, inactivation is present and fingertip velocity profiles reproduce the experimental directional asymmetries. Thus, these relevant features of movements are not affected by such changes in modeling. The reason for not systematically considering more precise levels of modeling is twofold. Firstly, it causes important additional computational difficulties, and secondly, many more parameters, which are not always well-known, appear in the model. Here, the model depends on a few parameters. Firstly, the maximum torque that can be developed by each muscle is finite. In particular, this determines the shortest possible movement duration in order to complete the pointing task. Nevertheless these maximum torques did not seem to be reached in practice (at least during the movements tested here) so that their precise values were not important for the present study. Secondly, the weighting parameters that appear in the cost could depend on the individual and the task goal. However, they are not critical with respect to the qualitative behavior of the optimal solutions and, although their values could be discussed, the simulations obtained using this model were accurate for a large range of these parameters. Importantly, the whole theory holds without precise constraints on these parameters. A first example is given by the strongly consistent kinematic difference in the 1-dof case for movements performed in the upward versus the downward direction. For instance, for an upward movement (1-dof, 45° and 400 ms), the relative time to peak velocity (TPV) ranged between 0.43 and 0.5 for weighting parameters ranging between 0 and 10. For the corresponding downward movement, TPV ranged between 0.57 and 0.5. The classical models [16]–[18] were not able to reproduce this directional difference in the speed profiles observed in vertical arm movement executed with 1-dof [37]. Moreover, it has been found that this difference disappeared for movements performed in the horizontal plane, either in upright or reclined postures [37],[38]. This behavior is experimentally well established and can be easily verified with simulations. Interestingly, it is predicted by our optimality criterion, whatever the choice of the tuning parameters. A second example concerns the final posture selected by the model. The exact terminal limb configuration depends on these weighting parameters. However, we tested several instances of the model, for weighting parameters ranging between 0.05 and 1. In all instances, the simulated terminal postures were in the range of those measured in practice. In order to check the validity of the present model, its predictions were also compared with well-known experimental findings, without trying to fit the data. The tuning parameters used are defined in the Materials and Methods Section. Movement curvature is known to depend on movement duration [36],[50]. Here, the 2-dof model predicts a change in the fingertip path curvature (FPC) when movement duration varies. For the movements tested in Figure 2, the FPC ranged between 0.18 and 0.23 for movement durations of between 0.2 s and 1 s. Moreover, the final postures have been found to be invariant with respect to the speed of the movement [8] and to the addition of a mass of 600 g on the forearm [9]. Here, in the 3-dof case, the final posture does not significantly vary with movement duration. For instance, the final postures changed by less than 3° (maximum change at each joint) while the movement duration ranged between 0.2 s and 1 s (tested for U and D movements that appeared in the left column of Figure 5). Also, adding a mass of 600 g to the forearm did not change the simulated final limb configuration: the model predicted less than 0.5° of variation at each joint. In the proposed model, the final posture is selected as the final limb configuration that minimizes the amount of the compromise Aw/Ae necessary to bring the finger to the target. Movements directed toward a single target were tested for various starting configurations of the arm. It resulted in changes in the final posture (about 1°, 10°, and 15° of variability at the shoulder, the elbow, and the wrist levels, respectively). Thus, the final posture depends on the initial configuration of the arm, in agreement with experimental results [21]. It must be noted that the minimum torque-change and the minimum force-change models failed to predict the curvature of movements when antigravity torques were implied in the optimization process, according to Figure 3 in [33]. In contrast, the finger trajectory for a 2-dof arm predicted by our model (for the same movements of duration equal to 400 ms) was quite realistic (Figure 7A). This was also in agreement with the experimental finger paths observed in Figure 4 in [6] for other movements performed in the sagittal plane (see Figure 7B). Although the proposed model was only tested in a sagittal workspace, it appears to be well-suited for a large set of movements and may, thus, motivate future extensions of the model to 3-dimensional movements. Several investigators have proposed that the CNS optimizes inertial forces and compensates gravitational forces at each instant [19],[26]. Static and dynamic forces were assumed to be controlled separately. Although plausible, this idea is hardly compatible with several experimental results. For instance, when considering an upward movement in the sagittal plane performed with the arm fully-extended (1-dof case), according to such a viewpoint, agonistic (anti-gravitational) muscles should be active throughout the movement (corresponding to a tonic component of EMGs) [31]. In this case, a muscular activity counteracting the gravity would be necessary to continuously maintain the arm, as if it were at equilibrium at each instant, and would be noticeable in EMGs. However, EMG recordings showed that the activities of the agonistic muscles were quasi-null near the time of peak velocity suggesting, thus, that no muscle was acting against gravity at this instant. Moreover, it may explain why, after subtracting the tonic activity from rectified EMG data, some authors obtained negative phasic activities of some muscles (e.g., see [51],[52]). Rather than resulting from errors in the evaluation of the tonic component of muscles activity, the gravitational and inertial forces could just be integrated into the same motor plan, within the minimization of energy expenditure. In that case, an explicit separation between tonic and phasic activities of muscles could be impossible, at least for fast movements. It must be noted that separating static and dynamic forces is not the same as separating posture and movement. Indeed, static and dynamic forces are present during posture maintaining. Neuro-anatomical and experimental evidences for distinct controls of posture and movement were reported in [53]. Thus, the present results concerning inactivation do not contradict the hypothesis that, while maintaining posture, anti-gravity control seems to be tightly related to the muscular system's viscoelastic properties (see [54] for a study of equilibrium control during quiet standing). This problem was not addressed here since we focused on the control of the transient phase of fast movements. In conclusion, from a methodological point of view, the novelty of the present work is to introduce a hypothetical-deductive approach in studies focusing on motor planning of arm movements. The possible existence of the inactivation phenomenon was deduced from a mathematical analysis which aimed to reproduce directional asymmetries in arm movements performed in the sagittal plane. Then, the presence of these inactivation periods produced by the model was confirmed by the EMG signals obtained from experimental data. The mathematical analysis showed that this inactivation was a necessary and sufficient condition for the minimization of an absolute-work-like cost. As far as we know, this is the first time that such a condition has been proved in studies investigating optimality principles in human movement. These results suggest that, considering that inactivation is a short and quite singular phenomenon, more attention should be paid to this specific movement feature in future studies. Two major conclusions can be drawn: This section is devoted to technical details and proofs of the results presented in the Theoretical Analysis Subsection. It is organized as follows. Firstly, we present the general setting of the optimal control problem under consideration. Secondly, we present the examples that will be used to illustrate the theory. After presenting some prerequisites that may be helpful to understand the main mathematical results, we state two theorems concerning the Inactivation Principle and the necessity of non-smoothness. Then, some details on the computation of the optimal solutions using Pontryagin's Maximum Principle [41] are reported (for the 1-dof and 2-dof cases). Finally, three extensions of the model are given in the case of i) gradient constraints on the control; ii) distinct control of agonistic and antagonistic torques; and iii) modeling the dynamics of agonistic and antagonistic muscles.

10.1371/journal.pcbi.0030171

Pathway Switching Explains the Sharp Response Characteristic of Hypoxia Response Network

Hypoxia induces the expression of genes that alter metabolism through the hypoxia-inducible factor (HIF). A theoretical model based on differential equations of the hypoxia response network has been previously proposed in which a sharp response to changes in oxygen concentration was observed but not quantitatively explained. That model consisted of reactions involving 23 molecular species among which the concentrations of HIF and oxygen were linked through a complex set of reactions. In this paper, we analyze this previous model using a combination of mathematical tools to draw out the key components of the network and explain quantitatively how they contribute to the sharp oxygen response. We find that the switch-like behavior is due to pathway-switching wherein HIF degrades rapidly under normoxia in one pathway, while the other pathway accumulates HIF to trigger downstream genes under hypoxia. The analytic technique is potentially useful in studying larger biomedical networks.

A complex biomolecular network utilizes different pathways to perform different functions. However, the interactions within the network are typically so complicated that the pathway structure is usually hidden. By some mathematical techniques, the pathways can be identified and possibly decoupled, whereby the insightful details of the network can be exposed. As an example, we study in this paper the hypoxia response network that manifests a dramatic switch-like behavior for certain sets of rate constants: a slight change of the oxygen concentration close to a critical value will lead to distinct reaction patterns. By a technique called extreme pathway analysis, the network is decoupled into three major and some minor pathways. Flux distribution among these pathways can thus be measured by integrating the ordinary differential equations for any given set of rate constants. For the sets of rate constants where the switch-like behavior is observed, we found that such a behavior is due to the switching of flux between two of the three major pathways.

Molecular oxygen is the terminal electron acceptor in the mitochondrial electron transport chain. Hypoxia, or oxygen deficiency, induces a number of metabolic changes with rapid and profound consequences on cell physiology. A hypoxia-induced shortage of energy alters gene expression, energy consumption, and cellular metabolism to allow for continued energy generation despite diminished oxygen availability. A molecular interaction map of the hypoxia response network has been proposed [1–3] on the basis of analyzing conserved components between nematodes and mammals. The key element in this network, hypoxia-inducible factor (HIF), is a master regulator of oxygen-sensitive gene expression [4–6]. HIF is a heterodimeric transcription factor which consists of one of the three different members (HIF-1α, HIF-2α, and HIF-3α) and a common constitutive ARNT subunit which is also known as HIFβ. The system also includes an enzyme family: prolyl hydroxylases (PHDs), which directly sense the level of oxygen and hydroxylate HIFα by covalently modifying the HIFα subunits. It is very likely that reactive oxidative species (ROS), which are a byproduct of mitochondrial respiration, are also involved in oxygen sensing by neutralizing a necessary cofactor, Fe2+, for the hydroxylation of HIFα by a PHD [7–10]. There are three members in this enzyme family: PHD1, PHD2, and PHD3. The hydroxylated HIFα is then targeted by the von Hippel-Lindau tumor-suppressor protein (VHL) for the ubiquitination-dependent degradation. Hypoxia response element (HRE) is the promoter of the hypoxia-regulated genes, and the occupancy of HRE controls the expression levels of these genes. The cascade in Figure 1 (reproduced from Figure 2 of [1]) consists of an input (the concentration of oxygen) and an output (the activation of promoters that are under control of HREs) as the core network. The network is characterized by a switch-like behavior, namely the sharp increase of HIFα when the oxygen decreases below a critical value, followed by a sharp increase of HRE occupancy. It was observed experimentally on many cell lines including Hela cells [11] and Hep 3B cells [12] that HIFα increases exponentially as the oxygen concentration decreases. The past two decades have seen a growing body of work on the use of mathematical modeling to help uncover both general principles behind molecular networks and to provide quantitative explanations of particular network phenomena [13] that may one day have sufficient predictive power to accurately model large subnetworks of the cell. In this sense, Kohn et al. [1] have successfully modeled the switch-like response characteristics of HRE occupancy, by numerically integrating a system of ordinary differential equations (ODEs) involving a score of molecular species related to hypoxia. The large model, however, does not identify the smaller components that are actually responsible for the switch-like response and that may occur in other such networks. Furthermore, a numerical solution does not provide the type of insight that mathematical formulas can. At the same time, it is virtually impossible to solve symbolically the type of nonlinear differential equations that model reactions. In this context, methods are desirable that are both tractable, that reduce a system to its key components, and that are not solely reliant on numerical solution. Extreme pathway analysis (EPA) is one such recently developed method [14–16]. In this method, the dynamics of interactions between species are formulated as a Boolean network in which the state of a gene is represented as either transcribed or not transcribed. Upregulation and downregulation of genes are captured through an appropriate sign (plus or minus) and a scaling constant. The Boolean network is then formulated as a matrix of interaction rules that is then analyzed to help reveal key components and their contributions to the dynamic behavior [16]. The theory of matrices then allows us to look for vectors that characterize the matrix in ways that are helpful for further analysis. The EPA technique, in particular, finds vectors (extreme pathways) that correspond to the boundaries of the space of steady-state solutions to the differential equations. We note that similar methods, such as flux balance analysis (FBA) and elementary modes analysis, have been developed in other contexts [17–19]. They essentially yield the same results [18], which have been verified by ExPa [20] and CellNetAnalyzer [21,22]. These methods provide a way out of the intractable complexity of sizable molecular networks [23–26]. Our contribution is to go beyond this type of matrix approach and provide a detailed quantitative analysis that explains the observed behavior in the models. This is achieved by combining elementary pathway identification via EPA, which depends solely on the network topology, and the detailed analytical as well as numerical analysis of the governing differential equations in the model, which allows studies of the phase space spanned by the mostly unknown rate constants in the differential equations. Specifically, EPA is first used in our approach to decompose the original network into several underlying pathways. Following this, we make some reasonable approximations to facilitate analytic solution. We show that this analytic solution, in the case of the hypoxia network, explains the switch-like behavior. This explanation is confirmed by comparing the numerical output of the simplified model with the numerical output of the complete (and complex) differential equation model. A second contribution of this paper is to highlight a particular mechanism of pathway-switching or pathway branching effect [27] that appears to cause the sharp response to oxygen concentration. In particular, we examine the flux redistribution among the elementary pathways as a function of oxygen concentration. We also identify the key molecular species involved in the subcomponent of the network and show quantitatively how the response of this subcomponent exactly matches the overall response and thus is responsible for it. For hypoxia, our analysis suggests that the cycle of abundant production and efficient degradation of HIFα plays the main role in the sharp response. For consistency and ease of understanding, we use the notation and nomenclature in [1] and use their 23-species network and differential equation model as the starting point. With this background, the original network shown in Figure 2 of [1] can be further reduced in the following way. Kohn et al. [1] have shown that the feedback of mRNA is not necessary for the switch-like behavior. We therefore eliminate this feedback loop (reaction k32). Hence, Transcript intermediate 1, 2, and 3 (Species 8, 9, and 10) can also be dropped, as well as the associated reactions: k7, k8, k9, k10, and k11, because they do not affect the dynamics of the network. Species 23 is only the joint name of HIFα:ARNT:HRE (Species 7) and HIFαOH:ARNT:HRE (Species 22); therefore, it is dropped. HIFα precursor (Species 1) is a constant and is thus dropped, because its information can be simply encoded in the reaction k1. The degradation products (Species 2) are also eliminated because they are assumed to leave the network immediately after their production and do not affect the dynamics. Similarly, species 19, 20, and 21 do not contribute to the dynamics and are therefore removed. The resulting network is summarized in Tables 1 and 2, where there are 13 molecular species and 19 reactions in total. The system can be described by the following set of ODEs where [Sn] stands for the concentration of species n as tabulated in Table 1 and [O2] indicates the input cellular oxygen concentration. Table 2 shows the specific reaction each rate constants kn represents. The real values of kn are from [1]. Note that the ODE system below is typical: the terms are based on mass-action principles and, taken together, result in complex behavior not readily discernible by examining the equations. We also dropped the precursor concentration [S1] since it is set to unity. This section assumes some familiarity with linear algebra. The 13 × 19 stoichiometric matrix Φ of the reduced hypoxia response network (Tables 1 and 2) is shown in Figure 2 (for details, see Materials and Methods). The rank of Φ is computed and shown to be 9, indicating that there are only nine independent molecular species to serve as constraints for the analysis. Therefore, the dimension of the corresponding null space is 10. The linearly independent basis B vectors are generated by Matlab 6.5 and are shown in Figure 3. According to the constraint that no negative values are allowed in the basis vectors, we can uniquely transform basis B into basis P as shown in Figure 4. Both b8 and b9 have negative terms. b8 has to be transformed first, otherwise there will be no +1 to cancel out the −1 at the twelfth row of b8. Each −1 in b8 is canceled out through the operation b8 + b9. In the second step, one has to use b7 to cancel −1 at the ninth row of b9. In this way, we obtain the set of basis vectors P. The above analysis indicates that the dimension of this null space is the same as the number of edges for its corresponding convex cone [28], which is the algebraic basis for extreme pathways [14] and elementary modes [29]. The ten basis vectors of P represent ten underlying pathways of the hypoxia network. They are illustrated in Figures 5 and 6, from which one finds two distinct patterns: p1, p7, and p9 belong to the HIFα degradation pathways (Figure 5) and the others belong to the simple association–dissociation pathways (Figure 6). More specifically, through p1, HIFα is directly degraded by reaction k2, a presumably oxygen-independent degradation pathway; whereas in oxygen-dependent pathways p7 and p9, the hydroxylated HIFα is recognized by the VHL that channels it through a ubiquitin degradation component that is shown as the dotted box in Figure 5. Even though p1, p7, and p9 are all elementary modes [29], they can share certain reactions of the network. For example, the total influx for HIFα synthesis from a precursor can thus be decomposed into three parts with the overall rate constant k1 being given by γ1k1, γ2k1, and γ3k1, where γ1 + γ2 + γ3 = 1. The pathways p7, and p9 are almost the same. The only difference is that HIFα is associated with ARNT in the middle part of the pathway p9. Therefore, ARNT must be functionally very important, otherwise it would be hard to explain why the two underlying pathways, which should play significantly different roles, look so similar. Indeed, HIFα degrades differently through the two pathways. The k-sets were selected in [1] on the basis that they produced a switch-like behavior. For all the three k-sets, it was observed that HIFα has high affinity to PHD and low affinity to ARNT. The former is consistent with the usual case of high enzyme-substrate binding affinity. This implies that p9 is not the major degradation pathway because HIFα does not bind with ARNT very well. Moreover, p9 is immediately adjacent to HRE, which suggests its major role is to deliver signals to activate the promoters of hypoxia-regulated genes. As a signal transducer, the rate constant γ3k1 itself need not to be high; what the downstream genes are sensitive to is d(γ3k1)/dt. Therefore, we hypothesize that there is a negligible flux through p9 (or γ3 ≈ 0). To verify our hypothesis, we calculate the γ1, γ2, and γ3 values, as the indication of the relative importance of p1, p7, and p9 in HIFα degradation, at different oxygen levels. The results are given in Table 3. Note that [O2] = 0.1 and [O2] = 1.0 represent typical low and high oxygen levels according to [1]. One sees that the pathway p9 is always much less important than the other two as far as HIFα degradation is concerned. The majority of HIFα gets degraded via either p1 at low oxygen or p7 at high oxygen. The comparison of hypoxia response network and heat shock response network [30] as in Table 4 shows the similarity between these two networks with respect to the issue of affinity. The huge difference in the affinity can clearly separate the underlying pathways and assign different functions to them. This is also the basis for the Goldbeter-Koshland model [31]. We tested our method to all three parameter sets (k-sets 1, 2, and 3) in [1], and find that the analytical results are almost identical with those of the direct simulations of the entire network, which strongly validates our approximation. For the rest of the paper thereafter, we only report numerical results for k-set 1. The EPA method gives us a starting point from which to analyze our reaction network in greater and more revealing detail. The verification of our hypothesis implies that the pathways associated with p9 could be neglected in the first place. The following equations describe the combination p1, p4, p6, and p7, which constitute the oxygen sensing mechanism: Note the differences between Equations 1, 6, and 8, and Equations 14, 15, and 17. The p9 related elements have been omitted due to their smallness. By setting the left-hand sides of the above equations to zero, one obtains the steady state equations: The total amount of PHD is conserved: PHD is either in the form of PHD (S12) or HIFα:PHD (S13). This implies where is the initial concentration of PHD. By some derivation, the following equation is obtained: where a = k2k12, b = b1 + b2[O2], c = c1 +c2[O2], where Since c < 0, Equation 25 has one and only one reasonable root Note that none of the species and reactions in the degradation box is present in Equation 26, which indicates that the components in the degradation box are not responsible for the sharp response curve. Once is determined, the analysis of the remaining network (p2, p3, p5, p6, p8, p9, and p10) becomes straightforward, and the results are given in the section “Additional results.” In fact, these results can be further simplified (see Equations 30 and 43 for and ). Figure 7A and 7B shows the steady-state values of [HIFα] ( ) and [HIFα:ARNT:HRE] ( ) at different oxygen values. The red lines depict the simulation results obtained by the numerical integration of the ODE system (1–13) until the steady state is reached. The black lines depict the analytical solutions that are obtained by the algebraic Equations 26 and 56. To better determine the critical point of pathway switching, we calculate ∂ /∂[O2] and ∂ /∂[O2]. The results are shown in Figure 7C and 7D. One sees that both ∂ /∂[O2] and ∂ /∂[O2] change abruptly in a very narrow region of [O2], with the rest of the values almost zero. One observes that the critical point is about [O2]c = 0.65. We thus show that the sharpness of the response curve can be determined analytically, instead of exhaustively enumerating [O2] values combined with time-consuming numerical integration of a large number of ODEs at each [O2] value. EPA is a powerful, yet simple tool that can significantly reduce the complexity of the original network and thus make further analytical effort feasible. In this paper, the additional analysis explains precisely the sharp reaction to oxygen of the network as a whole. The clear separation of p7 and p9 indicates their different functions: the pathway p7 and its other associated pathways constitute the sensing of ambient molecular oxygen; in contrast, the pathway p9 and its associated other pathways are responsible for the signal transduction to form the promoters of hypoxia-regulated genes. Most importantly, the simplification allows for a complete explanation of the switch behavior and a clear presentation of the relations between and [O2], and [O2], and . The first step below explains the sharp HIFα stabilization. The second step explains the sharp HRE occupancy that is induced by the HIFα stabilization. HIFα stabilization. This involves the dissociation of pathways p1, p4, p6, and p7 from the whole network, due to the fact that the flux through p9 is always small and can be neglected. It reveals a critical value that corresponds to the switching between pathways p1 and p7. Since an abrupt change often relates to the notion of singularity in mathematics, we proceed to see if a singularity can be found. Under nomoxia, ≈ 0, and can be neglected in Equation 25, which yields One immediately finds the singularity For k-set 1 in [1], one obtains [O2]c = 0.64, which is exactly the critical value found in Figure 7. When the oxygen level decreases to a value close to [O2]c, the denominator in Equation 27 becomes very small and a can no longer be ignored. Moreover, the term c in Equation 25 can be ignored compared with the large value of . One thus has This explains the linear decrease of versus [O2] increasing in Figure 7A. In summary, one has One can check that k2 can be ignored in the upper branch of Equation 30 due to its smallness, which again demonstrates that the pathway p1 is not important under nomoxia. The very smallness of k2 reflects the importance of p1 under hypoxia, for k2 exists at the denominator of the lower branch of Equation 30. HRE occupancy. The remaining pathways reveal how HIFα stabilization triggers a sharp increase of HIFα:ARNT:HRE, namely the sigmoid curve of versus [O2]. We conclude that the magnitude of HIFα is crucial for the sharpness of the curve. To show this, we need Equations 47, 48, 53, and 54. By removing the terms that are negligible, these equations turn into Equation 31 holds because k15 and k16 are far less than k3 and k4 . Equation 33 holds because is far less than and . Equation 34 holds because , , and are far less than , , and . The validity of the above approximations can be easily checked. For example, for [O2] = 0.1, one finds k3 = k4 = 1.66, k15 = k16 = 0.005, = 1.19, = 2.69, = 0.11, = 0.89, = 0.23, = 0.0016, and = 0.0005. From Equations 31–34, one obtains where α = α1 + α2/ , α1 = + + k6/k5, and α2 = k4k6/(k3k5). has one and only one reasonable solution where x = 2 /α. Taylor expanding , yields No matter what value is, x < 2 /α1 = 0.74, so 1 − x2/2 > 0.73 and x4/8 < 0.0375. Therefore ≈1 − x2/2 and Under nomoxia is small, so α ≈ α2/ and 1/α ≈ 0. From Equation 37 one has which is also small. Under hypoxia, is large, and thus By substituting Equation 39 into Equation 37, one obtains the important relation between and : where and By substituting Equation 29 into Equation 40, one obtains where m = /β1 and λ = β2b2/a. It is well-known that Equation 41 represents a sigmoid curve with m controlling the saturation value and λ controlling the sharpness. By ignoring the small term k6/k5 in the expression of α1, one finds m is a function of and only. One also finds Here The association (disassociation) constants k3, k5 (k4, k6) exist in the numerator (denominator) of the term k3,k5/(k4,k6), which implies that the higher the affinity, the sharper the response. The third term b2/a is proportional to the HIF level. In summary, our analysis yields One sees that HIFα is the key to triggering the HIFα:ARNT:HRE response. As long as the oxygen level is greater than [O2]c, HIFα is efficiently degraded by the pathway p7 and maintains a very low level, and the HIFα:ARNT:HRE level is also low (Equation 38). When the oxygen level drops below [O2]c, the system switches to the pathway p1, and HIFα stabilizes with a large concentration (b2/a large). This triggers the sharp increase of HIFα:ARNT:HRE. The smaller k2 is, the larger b2/a, and the sharper the HRE occupancy response (see Figure 8). Also, the validity of our analytical approximation is justified by the close resemblance of Figure 8B and 8C. The three major results of Kohn [1] involve HRE occupancy as a function of the oxygen concentration. The dependence of the curve on ARNT, VHL, and PHD are obtained by both simulation (Figure 9A) and analysis (Figure 9B) and are explained as follows. ARNT dependence. We need only to analyze the pathway p9. The amount of ARNT does not affect the shape of the response curve or the location of the sharp transition, because p9 is the pathway for HRE expression, while p7 is responsible for the sharpness. HIFα:ARNT would not be generated without ARNT and there would not be expressions of HRE for any level of oxygen. At high oxygen levels, the concentrations are similar because HRE occupancy is low anyway. At low oxygen levels, low levels of ARNT will give low HIFα:ARNT and then low HRE occupancy. VHL dependence. VHL is present in both p7 and p9. At high oxygen levels, one should analyze p7 because it is the major pathway for HIFα degradation. The VHL source will affect the upstream HIFα. If VHL concentration is low, it cannot degrade HIFα fast, and the system yields high HRE occupancy. At low oxygen levels, p1 is the major pathway for HIFα degradation, which does not depend on VHL. PHD dependence. One interesting property relates to the different locations of the transition. Using the criterion identified by the alternative model without reaction k2, we can calculate the transition locations as shown in Figure 9B3 for different PHD values. The results are the same as in Figure 9A3. As a matter of fact, considering the p9 pathway only yields a much simpler analytical solution that is also accurate. This further simplification is due to the fact that the expression of HIFαOH:ARNT:HRE is always negligible. The present model of the hypoxia response network is probably an oversimplified one. Nevertheless, it serves as an important starting point, from both theoretical and experimental perspectives, before a more detailed model can be understood. The present model will be gradually expanded and analyzed, with the input of more quantitative data from future experiments. One advantage of EPA is that the method can easily incorporate mechanistic details as soon as they become available [16]. There are various molecular interactions that can be added to the model. For example, it was demonstrated that HIF influences mitochondrial function by inducing pyruvate dehydrogenase kinase 1 (PDK1) to suppress the tricarboxylic acid (TCA) cycle and thus the aerobic respiration. Then the respiration shifts to be anaerobic, whereby the oxygen resource can be preserved to promote cell survival under hypoxic environment [32–34]. Another subject we are interested in is the inclusion of ROS in the network. It has been established that ROS affects HIFα degradation through Fe2+ [7–10]. The direct hydroxylation of HIFα by oxygen requires Fe2+. Under hypoxia conditions, however, ROS increases dramatically and consistently removes Fe2+ via oxidation to Fe3+. Together with the shortage of oxygen, this makes the HIFα degradation through the VHL pathway even more difficult. Consequently, the transition would be faster and sharper. Analysis should focus on the explanation of the coexistence of two oxygen sensing components, a matter that does not appear to be settled as yet. To obtain the dynamical response when the oxgen changes continuously in time, Equations 1–13 (the full model) are integrated. Figure 10 shows the temporal changes of [HIFα] and [HIFα:ARNT:HRE] as responses to the oxygen decreasing from 1.0 to 0.1 with different rates. One sees that [S3] and [S7] increase prominently only after [O2] decreases below [O2]c. The faster [O2] decreases, the more rapid the responses are. In particular, when [O2] abruptly jumps from 1.0 to 0.1, the responses ensue promptly. However, it is worth noting that one cannot tell practically how fast the responses are because the time scale is unknown. Indeed, the model is dimensionless and no units are given. Nevertheless, the sharp curves illustrate that the responses are very sensitive to the oxygen concentration and imply that the system can provide a timely response under hypoxia. Physiologists have long been puzzled by the ceaseless HIFα cycle, characterized by both abundant generation and efficient degradation, which seems to be a highly wasteful process. Our analysis provides a reasonable explanation. To deal with a sudden environmental change from nomoxia to hypoxia, an organism must respond in time to trigger the genes necessary for adapting to the new environment. To achieve such a sharp response, a high HIFα generation potential is necessary. Since the hypoxia conditions are rare, an efficient degradation pathway has to be designed to maintain a low HIFα under nomoxia. The HIFα cycle is indeed uneconomic, but it appears useful in helping the cell respond to sudden, unpredictable changes in its environment. In summary, we have obtained an accurate analytical solution to the hypoxia response network and have provided a complete explanation of the switch-like behavior first observed and modeled in [1]. The first step of our analysis applied the EPA technique to a reduced, yet complete system that resulted in exposing ten independent pathways, allowing us to focus on analyzing the pathways relevant to HRE occupancy. The analysis showed that the sharp response of HRE occupancy is due to the switch between the pathway p7 (p1) that degrades HIFα under nomoxia (hypoxia). In following the law of mass conservation, a particular reactant through each reaction can be written in the form of homogeneous linear equations, Here Φ is an m × m stoichiometric matrix, where m is the number of metabolites and n is the number of reactions taking place within the network, with each element , where N = dim NulΦ is the dimension of the null space of Φ, vk is the basis vector that corresponds to the k-th pathway, and fk is the flux through the k-th pathway. The determination of flux fk requires numerical calculations. Note that vk does not depend on fk and can be obtained solely from Φ. That is, one can decompose the whole network into N elementary pathways without any ODE integration involved. The dimension of the null space of Φ follows the simple equation Using Matlab (http://www.mathworks.com), the rank of this 13 × 19 stoichiometric matrix Φ is found to be nine, and the dimension of the corresponding null space is thus ten. Also, a set (B) of ten independent basis vectors is generated as shown in Figure 3. However, the set is not biologically reasonable due to the negative entries therein. By simple linear transformation, another set (P) of ten vectors (Figure 4) is obtained whose entries are all non-negative (either 1 or 0) and are thus biologically feasible. The remaining network can be solved analytically with HIFα( ) already determined. The ODE description is as follows: Together with the three constraints of , , and , the steady state equations are expressed as follows: where and have already been determined from the analysis of p7. By some reasoning, a quartic equation is derived, from which can be obtained. and can then be expressed as functions of : where Protein accession numbers as listed in Table 1 are from http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?DB=protein.

10.1371/journal.pgen.1005520

FLCN and AMPK Confer Resistance to Hyperosmotic Stress via Remodeling of Glycogen Stores

Mechanisms of adaptation to environmental changes in osmolarity are fundamental for cellular and organismal survival. Here we identify a novel osmotic stress resistance pathway in Caenorhabditis elegans (C. elegans), which is dependent on the metabolic master regulator 5’-AMP-activated protein kinase (AMPK) and its negative regulator Folliculin (FLCN). FLCN-1 is the nematode ortholog of the tumor suppressor FLCN, responsible for the Birt-Hogg-Dubé (BHD) tumor syndrome. We show that flcn-1 mutants exhibit increased resistance to hyperosmotic stress via constitutive AMPK-dependent accumulation of glycogen reserves. Upon hyperosmotic stress exposure, glycogen stores are rapidly degraded, leading to a significant accumulation of the organic osmolyte glycerol through transcriptional upregulation of glycerol-3-phosphate dehydrogenase enzymes (gpdh-1 and gpdh-2). Importantly, the hyperosmotic stress resistance in flcn-1 mutant and wild-type animals is strongly suppressed by loss of AMPK, glycogen synthase, glycogen phosphorylase, or simultaneous loss of gpdh-1 and gpdh-2 enzymes. Our studies show for the first time that animals normally exhibit AMPK-dependent glycogen stores, which can be utilized for rapid adaptation to either energy stress or hyperosmotic stress. Importantly, we show that glycogen accumulates in kidneys from mice lacking FLCN and in renal tumors from a BHD patient. Our findings suggest a dual role for glycogen, acting as a reservoir for energy supply and osmolyte production, and both processes might be supporting tumorigenesis.

The ability of an organism to adapt to sudden changes in environmental osmolarity is critical to ensure growth, propagation, and survival. The synthesis of organic osmolytes is a common adaptive strategy to survive hyperosmotic stress. However, it was not well understood, which biosynthetic pathways and storage strategies were used by organisms to rapidly generate osmolytes upon acute hyperosmotic stress. Here, we demonstrate that glycogen is an essential reservoir that is used upon acute hyperosmotic stress to generate the organic osmolyte glycerol promoting fast and efficient protection. Importantly, we show that this pathway is regulated by FLCN-1, an ortholog of the human tumor suppressor Folliculin responsible for the Birt-Hogg-Dubé cancer syndrome, and by AMPK, the master regulator of energy homeostasis.

Water is a fundamental molecule for life and the ability of an organism to adapt to changes in water content is essential to ensure survival. Hyperosmotic stress promotes water efflux, causing cellular shrinkage, protein and DNA damage, cell cycle arrest and cell death. All living organisms encounter hyperosmotic environments [1,2]. In humans, both renal and non renal tissues are exposed to hyperosmotic stress, a condition that is regarded as a major cause for many chronic and fatal human diseases including diabetes, inflammatory bowel disease, hypernatremia, dry eye syndrome, and cancer [1]. Cells/tissues/organisms have evolved adaptive strategies to cope with threatening hyperosmotic environments [1,2]. Among adaptive strategies, the synthesis of compatible organic osmolytes, which keeps cellular osmotic pressure equal to that of the external environment, is widely used by all organisms [3]. In yeast and C. elegans, hyperosmotic stress triggers glycerol production via transcriptional upregulation of glycerol-3-phosphate dehydrogenase-1 (gpdh-1), a rate-limiting enzyme in glycerol synthesis [4,5]. Moreover, several osmotic stress resistance mutants of divergent signaling pathways exhibit a constitutive transcriptional upregulation of gpdh-1, leading to increased glycerol content [6–10]. Here we define a novel hyperosmotic stress resistance pathway mediated by the 5' AMP-activated protein kinase (AMPK), a key regulator of cellular energy balance [11], which is chronically inactivated by the worm ortholog of the renal tumor suppressor Folliculin (FLCN-1). In humans, FLCN is a tumor suppressor gene responsible for the BHD disease, an autosomal dominantly-inherited syndrome associated with increased susceptibility to the development of several cancerous and non cancerous lesions including kidney cancer, pulmonary, renal, pancreatic and hepatic cysts and skin fibrofolliculomas [12–25]. FLCN has been shown to bind AMPK via the scaffold FLCN-interacting proteins FNIP1 and FNIP2 [26,27]. We have recently demonstrated that FLCN negatively regulates AMPK signaling in the nematode C. elegans and in mammalian cells [28,29]. Moreover, loss of FLCN increased ATP levels via heightened flux of glycolysis, oxidative phosphorylation, and autophagy, which resulted in an AMPK-dependent resistance to several metabolic stresses in C. elegans and mammalian cells [28,29]. Here we identify a pathway involved in the physiological response to hyperosmotic stress resistance in C. elegans mediated by FLCN-1 and AMPK. We demonstrate that glycogen is an essential reservoir that is used upon acute hyperosmotic stress to generate glycerol and promote fast and efficient adaptation to prevent water loss and ensure survival. We show that in flcn-1(ok975) mutant animals, this phenotype is significantly enhanced, due to the robust AMPK-mediated accumulation of glycogen, which is rapidly converted to the osmolyte glycerol upon salt stress. Our results also suggest that the FLCN/AMPK pathway might be an evolutionarily conserved key regulator of glycogen metabolism and stress resistance. Since we have previously observed that loss of flcn-1 in C. elegans increases AMPK-dependent resistance to energy stresses including oxidative stress, heat, and anoxia [28], we asked whether it would also increase resistance to hyperosmotic stress. We measured the survival of wt and flcn-1(ok975) animals (S1A Fig) on plates supplemented with 400mM and 500mM NaCl. Loss of flcn-1 conferred a significant increase in resistance to hyperosmotic stress (Fig 1A and 1B and S1 Table). Although NaCl treatment severely reduced the survival of both wt and flcn-1(ok975) animals as compared to untreated animals (Figs 1A, 1B and S1B), the mean survival of flcn-1(ok975) animals increased by ~2 and ~3 fold upon treatment with 400mM and 500mM NaCl respectively, as compared to wt animals (Fig 1A–1C). Moreover, we did not observe a significant difference in lifespan between untreated wt and flcn-1(ok975) animals, as reported previously [28] (S1B Fig and S1 Table). Importantly, NaCl treatment led to shrinkage and paralysis in both wt and flcn-1(ok975) animals. However, flcn-1(ok975) mutant nematodes recover significantly faster than wt animals after 2 hours of NaCl treatment suggesting that the mechanism of adaptation to salt is more robust upon loss flcn-1 (Fig 1D). We also observed a significantly greater number of wt animals with more than 30% reduction of body size as compared to flcn-1 suggesting that loss of flcn-1 activates pathways that favor body size recovery after hyperosmotic stress (Fig 1E). Importantly, the hyperosmotic stress resistance phenotype can be rescued by transgenic re-expression of C. elegans flcn-1 (S1 Table and Figs 1F and S1A). In addition, we used Agilent whole genome C. elegans microarrays to determine transcriptional profile differences between wt and flcn-1(ok975) mutant animals [30]. Key genes that were differentially expressed were validated by qRT-PCR (S1C Fig). We compared our data to published transcriptional profiles and found a significant overlap between genes upregulated in untreated flcn-1(ok975) animals versus genes upregulated in wt animals treated with NaCl or osmotic stress resistant strains including osm-7 and osm-11 [8] (S1D, S1E and S1F Fig and S2, S3 and S4 Tables). Altogether, these data suggest that flcn-1 is involved in a mechanism of regulating the resistance to hyperosmotic stress. To determine how loss of flcn-1 increases resistance to hyperosmotic stress, we assessed the morphological differences between wt and flcn-1(ok975) using electron microscopy with or without NaCl treatment. Interestingly, we observed an increase in the size and number of glycogen stores in adult (Fig 2Ai and 2Aii) and L4 (S2Ai, S2Aii, S2Ci, and S2Cii Fig) flcn-1(ok975) mutant worms as compared to wt. Specifically, our transmission electron data indicate a strong accumulation of glycogen in the hypodermis, muscle, and intestine of flcn-1(ok975) animals as compared to wt (S2C Fig). Glycogen has been previously shown to accumulate in these tissues in C. elegans [31]. Importantly, glycogen stores were barely detectable in wt and flcn-1(ok975) animals after NaCl treatment, suggesting that glycogen degradation is used to protect the animals from hyperosmotic stress (Fig 2Aiii, 2Aiv). Furthermore, we found that the prominent accumulation and salt stress-dependent degradation of glycogen in flcn-1(ok975) adult animals occurs in the hypodermis (Figs 2A, S2A and S2C). We validated and quantified the increase in glycogen levels conferred by loss of flcn-1 using iodine staining which has been previously shown to specifically stain glycogen in C. elegans [32–34] (Fig 2B and 2C). In accordance with the electron microscopy results, glycogen levels were significantly increased in untreated flcn-1(ok975) animals as compared to wt, and NaCl treatment severely reduced glycogen content in both wt and flcn-1(ok975) animals (Fig 2B and 2C). We then asked whether glycogen is used to protect wt and flcn-1(ok975) animals from damage during hyperosmotic stress. Glycogen synthase (gsy-1) is responsible for the synthesis of glycogen from UDP-glucose molecules and glycogen phosphorylase (pygl-1) catalyzes glycogen breakdown to form glucose-1-phosphate [35]. Importantly, the inhibition of glycogen synthesis or degradation using RNAi against gsy-1 and pygl-1 respectively, strongly reduced the survival in both wt and flcn-1(ok975) animals to an equal level, suggesting that the accumulation of glycogen and its degradation are both required for the resistance of wt and flcn-1(ok975) mutant animals to hyperosmotic stress (Fig 2D and 2E and S1 Table). Additionally, transcript levels of gsy-1 and pygl-1 with or without 2 hours of 400mM NaCl stress remained unchanged suggesting allosteric regulation of glycogen metabolism (Fig 2F). Altogether, these results demonstrate that the accumulation of glycogen stores and the degradation of glycogen are essential to survive hyperosmotic stress in wt and flcn-1(ok975) mutant animals. Since we have previously reported that the flcn-1-dependent resistance to energy stresses requires aak-2, the worm ortholog of the AMPKα subunit, we wondered whether the hyperosmotic stress resistance phenotype conferred by loss of flcn-1 is also mediated by AMPK [28]. AMPK is activated by hyperosmotic stress in mammalian systems [36] and its deletion confers sensitivity to NaCl stress in yeast [37]. C. elegans nematodes have two catalytic α subunits aak-1 and aak-2. Loss of aak-2 was shown to mediate lifespan extension and resistance to various stresses including oxidative stress, anoxia, nutrient deprivation, and dietary restriction [38–42]. To determine whether AMPK is involved in the increased resistance of flcn-1(ok975) animals to stress, we crossed aak-2(ok524 and gt33) [39,43] and aak-1(tm1944) [43] loss of function mutants with flcn-1(ok975) animals. Interestingly, loss of aak-2(ok524 and gt33) or aak-1(tm1044) alone conferred stress sensitivity but did not fully suppress the increased survival to hyperosmotic stress conferred by loss of flcn-1 (Fig 3A, 3B and 3C and S1 Table). To control for compensatory effects, we generated the flcn-1(ok975); aak-1(tm1944); aak-2(ok524) triple mutant and compared its survival under high salt conditions to aak-1(tm1944); aak-2(ok524) double mutant animals. Simultaneous loss of aak-1 and aak-2 completely abolished the increased osmotic stress resistance upon loss flcn-1 demonstrating that this phenotype requires both AMPK catalytic subunits (Fig 3D and S1 Table). AMPK has been shown to regulate glycogen metabolism in different organisms [44–56]. In fact, acute activation of AMPK leads to glycogen degradation [44–47], while chronic AMPK activation results in glycogen accumulation [48–50]. Since we observed an increased constitutive phosphorylation of AMPK upon loss of flcn-1 in nematodes and mammalian cells [28,29], we hypothesized that the chronic AMPK activation in flcn-1(ok975) mutants may lead to increased glycogen levels. We determined glycogen levels in aak-1(tm1944); aak-2(ok524) animals compared to flcn-1(ok975); aak-1(tm1944); aak-2(ok524) triple mutant worms and found that loss of AMPK strongly reduced glycogen levels in both strains (Fig 3E and 3F). This suggests that the chronic AMPK activation in flcn-1 animals is leading to glycogen accumulation. Interestingly, the survival and glycogen accumulation in aak-1(tm1944); aak-2(ok524) mutant animals was also severely reduced as compared to wt (Fig 3E and 3F), suggesting an important role for AMPK in maintaining glycogen stores, which are used for hyperosmotic stress resistance. Autophagy is a biological survival process through which cellular components and damaged organelles are degraded to produce energy upon starvation [57]. We reported previously that autophagy was essential for the energy stress resistance of flcn-1(ok975) mutant animals [28]. Therefore, we asked whether autophagy plays a role in osmotic stress resistance. Interestingly, atg-18(gk378) mutant animals were hypersensitive to high salt concentrations suggesting that autophagy is a process involved in the resistance to hyperosmotic stress. However, loss of flcn-1 significantly increased the resistance of atg-18(gk378) animals suggesting that flcn-1-dependent hyperosmotic stress resistance does not require autophagy, which is different from what we observed before during energy stress [28] (S3 Fig and S1 Table). Degradation of glycogen polymers leads to the formation of glucose-1-phosphate which is converted to glucose-6-phosphate, an important metabolite used in multiple pathways including glycolysis, pentose phosphate pathway, and glycerol production (Fig 4A) [35]. We hypothesized that glycogen degradation may lead to heightened glycerol levels that could protect the animals from hyperosmotic stress. To address this, we measured the mRNA levels of gpdh-1 and gpdh-2. Interestingly, we observed a significant 2-fold increase in gpdh-1 but not gpdh-2 at unstressed conditions in flcn-1(ok975) mutant animals compared to wt, which was consistent with our microarray results (Fig 4B and 4C and S2 Table). Strikingly, after 2 hour treatment with 400mM NaCl, we detected a strong induction of gpdh-1 and gpdh-2 mRNA levels in wt and flcn-1(ok975) mutant animals, which was significantly enhanced in the latter (Fig 4B and 4C). Accordingly, flcn-1(ok975) mutant animals exhibit higher glycerol content at basal level as compared to wt animals which was further increased upon NaCl treatment (Fig 4D). To determine the importance of glycerol in the protection against hyperosmotic stress, we inhibited gpdh-1 and gpdh-2 using RNAi and using mutant strains. Importantly, treatment of flcn-1(ok975) animals with RNAi against either gpdh-1 or gpdh-2 alone did not fully suppress the increased resistance of flcn-1(ok975) animals to hyperosmotic stress (S4A and S4B Fig). We then compared the resistance of flcn-1(ok975); gpdh-1(kb24); gpdh-2(kb33) triple mutant animals to gpdh-1(kb24); gpdh-2(kb33) mutant nematodes. Simultaneous loss of gpdh-1 and gpdh-2 strongly reduced the survival of flcn-1(ok975) mutant animals demonstrating an important role for the osmolyte glycerol in the survival of flcn-1(ok975) and wt animals (Fig 4E and S1 Table). Altogether, these data suggest that upon hyperosmotic stress glycogen stores are metabolized into the osmolyte glycerol via enhanced transcriptional upregulation of gpdh enzymes. This glycerol mediated osmo-protective phenotype is significantly enhanced upon loss of flcn-1 in nematodes. HOG/p38/PMK-1 MAP kinase signaling is widely known to control adaptation to hypertonic stresses in multiple organisms [4,9,58]. As expected, pmk-1(km25) mutant worms were highly sensitive to osmotic stress. However, loss of pmk-1 in flcn-1(ok975) mutant animals reduced but did not fully suppress the increased resistance conferred by flcn-1 depletion (S5A Fig and S1 Table). Supporting this result, the expression of gpdh-1 is ~2-fold higher in flcn-1(ok975); pmk-1(km25) mutant animals as compared to pmk-1(km25) alone (S5B Fig). Altogether, this suggests that pmk-1 is not involved in the transcriptional upregulation of gpdh-1 upon loss of flcn-1 and that it acts in parallel to flcn-1 and aak-1/2. Glycogen is linked to the progression and the aggressiveness of multiple cancer types in humans [59,60]. To determine whether loss of FLCN also leads to the accumulation of glycogen in mammalian systems, we used the Flcnflox/flox/Pax8-Cre mouse model where Flcn is specifically deleted in the kidney and determined glycogen content using Periodic-Acid-Schiff (PAS) staining. The Flcnflox/flox/Pax8-Cre mouse was generated by mating Pax8-Cre mice with the Flcn flox/flox C57BL/6 mice. By six months of age, all mice developed visible macroscopic lesions confirmed as cysts that later developed into tumors. Strikingly, kidneys from Flcnflox/flox/Pax8-Cre mice accumulated higher glycogen levels as compared to normal kidneys from Flcnflox/flox mouse littermates (Figs 5A and S6A). Our data show a stronger glycogen accumulation in the kidney cortex, which is due to the fact that Pax8 is expressed in the epithelial cells of the proximal and distal renal tubules, loops of Henle, collecting ducts and the parietal epithelial cells of Bowman’s capsule [61]. Importantly, PAS staining of tumors from BHD patients also indicate a strong accumulation of glycogen as compared to adjacent unaffected kidneys (Figs 5B and S6B). We also compared the expression level of glycogen biosynthesis and degradation genes in 3 different subtypes of kidney cancer, kidney renal papillary cell carcinoma (KIRP), kidney renal clear cell carcinoma (KIRC), and kidney chromophobe (KICH) tumors. Strikingly, we observed a significant upregulation of genes involved in the synthesis and degradation of glycogen (Fig 5C and S5 Table). We also observed that the expression of 46% of these genes are negatively correlated with FLCN expression (Fig 5D). Overall, our data indicate that the accumulation of glycogen upon loss of FLCN is be conserved from nematodes to mammals, and that it might play a role in tumorigenesis. A common mechanism to survive osmotic stress is the synthesis of compatible osmolytes [3]. In yeast and in C. elegans, the rapid accumulation of glycerol after hyperosmotic stress has been demonstrated [4,5]. However, it is not clear what fuels glycerol production upon acute hyperosmotic stress. Here we show that animals have evolved an interesting strategy to maintain glycogen stores, which can serve as fuel for glycerol production to ensure survival to acute hyperosmotic stress (Fig 6). While storage of soluble glucose molecules in cells would lead to osmotic stress, the storage of glucose in the form of insoluble glycogen polymers ensures osmotic homeostasis. Importantly, our data uncover that glycogen stores have a dual role: they can serve as a reservoir for production of energy or osmolytes. Indeed, pretreatment of wt and flcn-1(ok975) animals with oxidative and energy stressor paraquat, depletes glycogen stores rapidly and suppresses survival upon treatment with 400mM NaCl (S2A and S2B Fig). The regulation of glycogen metabolism by AMPK has long been a paradox [44–50]. Acute activation of AMPK, by in vitro short term treatment of the AMP mimetic drug 5-Aminoimidazole-4-Carboxamide Riboside (AICAR), leads to the phosphorylation and inhibition of glycogen synthase, which favors glycogen degradation for supply of short term energy [44–47]. However, chronic AMPK activation induced by a long term AICAR treatment or by genetic manipulation of AMPK regulatory subunits, results in glycogen accumulation via glucose-6-phosphate-dependent allosteric activation of glycogen synthase, which bypasses the inhibitory effect of the AMPK-mediated phosphorylation [48–50]. In agreement, constitutive AMPK activation through transgenic expression of activating mutations in the γ2 and γ3 subunits in mice and pigs leads to substantial glycogen accumulation in cardiac and skeletal muscles [36,50–53,55,56]. In light of these results, our data indicate that chronic AMPK activation upon loss of flcn-1 leads to glycogen accumulation. Similarly to what has been shown in yeast [54], we demonstrate that AMPK-deficient strains exhibit reduced glycogen content as compared to wt. We further show that the accumulation of glycogen in wt and flcn-1(ok975) mutant animals depends on AMPK. Based on the data presented here together with our recently published reports [28,29], we propose that FLCN is a key regulatory component of AMPK. Flcn muscle-specific knockout mice and Fnip1 knockout mice exhibited increased glycogen accumulation in muscles and liver, respectively [62,63]. Here we show that loss of FLCN leads to glycogen accumulation in kidneys of mice and in the tumors of BHD patients, suggesting that this pathway is evolutionarily conserved. In agreement with the important role for glycogen in organismal survival to stress, glycogen can be used by tumor cells to survive harsh microenvironments such as hypoxia [59,64]. In fact, glycogen accumulates in many cancer types [64] and inhibition of its degradation led to induction of apoptosis and impaired in vivo growth of tumor xenografts [59]. Importantly, our data might impinge on a novel role for glycogen in tumorigenesis. In addition to its critical role as an energy supplier, we speculate that glycogen degradation might lead to higher osmolyte levels to help survive hyperosmotic tumor microenvironments. In fact, we found that taurine and sorbitol synthesis genes, CSAD and AKR1B1 respectively, are upregulated in many kidney tumors (S5 Table). Supporting this idea, recent evidence shed light on an important role of the nuclear factor of activated T cells 5 (NFAT5), a major transcription factor that regulates osmotic stress resistance genes, in promoting tumorigenesis and metastasis of several cancer types [2,65–67]. In summary, we speculate that the increased glycogen stores in tumors might lead to extended survival of cells under hyperosmotic stress, which could ultimately lead to neoplastic transformation by accumulation of DNA damage [1, 2]. C. elegans strains were obtained from the Caenorhabditis Genetics Center (S6 Table). Nematodes were maintained and synchronized using standard culture methods [68]. The RNAi feeding experiments were performed as described in [69], and bacteria transformed with empty vector were used as control. For all RNAi experiments, phenotypes were scored with the F1 generation. To measure osmotic stress resistance, synchronized 1 day adult worms were transferred to high concentration NaCl plates. Survival was measured daily. Worms that responded by movement to touch with the platinum wire were considered as alive. To measure the percentage of animals that recovered after hyperosmotic shock, 1 day adult animals were transferred to high NaCl plates. Animals shrink and paralyse shortly after exposure to NaCl. After 2 hours, animals that were able to move their entire body forward or backward in response to touch with a platinum wire were considered as “recovered”. Paralyzed animals often look straight and are unable to move. Synchronized young adult nematodes were harvested and total RNA was extracted with Trizol. Reverse transcription and qRT-PCRs were performed as previously described [28]. Transcripts were normalized to cdc-42. Synchronized young adult wt and flcn-1(ok975) animals were harvested and RNA was extracted using Trizol and purified using Qiagen RNeasy columns. Total RNA samples were then hybridized onto Agilent gene chips. Fold change values are calculated using the mean of both data sets. The overlapping genes between flcn-1(ok975) mutant animals and the specified conditions and strains [8] were performed using the “compare two lists” online tool at http://www.nemates.org/MA/progs/Compare.html. The significance of the overlap and enrichment scores were determined via hypergeometric distribution method using http://nemates.org/MA/progs/overlap_stats.html. The number of genes in the C. elegans genome was considered to be19,735. Synchronized 1 day adult nematodes were transferred to 400mM NaCl plates for 16 hours. Recovering animals were picked and transferred for TEM. Immersion fixation and embedding was performed according to [70]. Thin sections were cut on an RMC Powertome XL (Boeckler Instruments) using a diamond knife (DDK) and collected on Pioloform-coated copper slot grids. Grids were post-stained with 4% uranyl acetate and lead citrate and viewed using a Philips CM10 electron microscope (FEI) equipped with a Morada digital camera (Olympus) and iTEM software (Olympus SIS). Synchronized young adult animals were transferred to agarose pads. For comparisons between strains, different conditions were transferred to the same agarose pad and were exposed to iodine vapor for 30 seconds. Animals were rapidly imaged individually. Quantification of the intensity of the staining was performed using ImageJ software. For human normal kidney and BHD tumor samples, slides were rehydrated after deparaffination and treated with 1% periodic acid for 10 minutes. Periodic acid was washed off with H2O and slides were then incubated in Schiff reagent for 20 min. Slides were then rinsed with H2O, counterstained with hematoxylin and embedded in entellan. Images were taken as described in [71]. Synchronized L4/young adult animals exposed or not to 400mM NaCl for 2 hours and were harvested and washed with M9 buffer adjusted to match plate salinity. Pellets were flash frozen in liquid nitrogen. Extraction was performed according to [5]. Briefly, frozen pellets were ground using a cold mortar and pestle on dry ice. The worm powder was then resuspended in 1N perchloric acid, and solutions were transferred to 15ml conical tubes and kept on ice for 1 hour. The lysate was then centrifuged and the supernatant was neutralized with 5N KOH containing 61.5mM K2HPO4 and 38.5mM KH2PO4. Glycerol levels were determined using a glycerol determination kit (R-Biopharm, Marshall, MI). Pellets were solubilized in 0.1N NaOH and protein content was determined using BCA. Glycerol levels were normalized to protein content. TCGA data including 91 kidney chromophobe gene expression RNASeq (IlluminaHiSeq), 604 kidney renal clear cell carcinoma gene expression RNASeq (IlluminaHiSeq), and 258 kidney renal papillary cell carcinoma gene expression RNASeq (IlluminaHiSeq), were extracted from cancer Genomics Browser (https://genome-cancer.ucsc.edu/proj/site/hgHeatmap). For expression analysis, data were expressed as median fold change and the Mann-Whitney test was used to calculate the p-values between normal and tumor samples. P-values less than 0.05 were considered to be statistically significant. For correlation analysis TCGA expression data (same as expression analysis) were used to calculate the Pearson correlation coefficient, and generate a heat map, using R software 3.1.1 (http://www.r-project.org/). P-values less than 0.05 were considered to be statistically significant. Data are expressed as means ±SEM. Statistical analyses for all data were performed by student's t-test, using Excel (Microsoft, Albuquerque, NM, USA). For hyperosmotic stress survival curve comparisons we used the Log-rank Mantel Cox test using GraphPad software. Statistical significance is indicated in figures (* P<0.05, **P<0.01, ***P<0.001) or included in the supplemental tables.

10.1371/journal.pgen.1003538

SPO11-Independent DNA Repair Foci and Their Role in Meiotic Silencing

In mammalian meiotic prophase, the initial steps in repair of SPO11-induced DNA double-strand breaks (DSBs) are required to obtain stable homologous chromosome pairing and synapsis. The X and Y chromosomes pair and synapse only in the short pseudo-autosomal regions. The rest of the chromatin of the sex chromosomes remain unsynapsed, contains persistent meiotic DSBs, and the whole so-called XY body undergoes meiotic sex chromosome inactivation (MSCI). A more general mechanism, named meiotic silencing of unsynapsed chromatin (MSUC), is activated when autosomes fail to synapse. In the absence of SPO11, many chromosomal regions remain unsynapsed, but MSUC takes place only on part of the unsynapsed chromatin. We asked if spontaneous DSBs occur in meiocytes that lack a functional SPO11 protein, and if these might be involved in targeting the MSUC response to part of the unsynapsed chromatin. We generated mice carrying a point mutation that disrupts the predicted catalytic site of SPO11 (Spo11YF/YF), and blocks its DSB-inducing activity. Interestingly, we observed foci of proteins involved in the processing of DNA damage, such as RAD51, DMC1, and RPA, both in Spo11YF/YF and Spo11 knockout meiocytes. These foci preferentially localized to the areas that undergo MSUC and form the so-called pseudo XY body. In SPO11-deficient oocytes, the number of repair foci increased during oocyte development, indicating the induction of S phase-independent, de novo DNA damage. In wild type pachytene oocytes we observed meiotic silencing in two types of pseudo XY bodies, one type containing DMC1 and RAD51 foci on unsynapsed axes, and another type containing only RAD51 foci, mainly on synapsed axes. Taken together, our results indicate that in addition to asynapsis, persistent SPO11-induced DSBs are important for the initiation of MSCI and MSUC, and that SPO11-independent DNA repair foci contribute to the MSUC response in oocytes.

Meiosis is a special cell division that generates genetically divergent haploid germ cells. At the very beginning of this process, during meiotic prophase, the enzyme SPO11 generates hundreds of DNA double-strand breaks (DSBs). Meiotic DSBs are repaired via a mechanism that requires the presence of an intact homologous template. This repair process stimulates homologous chromosome pairing, and the formation of a protein complex that connects the paired chromosome axes, reaching a state called synapsis. Male mammals carry a pair of largely heterologous sex chromosomes, the X and Y, which show delayed DSB repair and extensive asynapsis. In addition, the X and Y chromosomes are transcriptionally silenced by a mechanism named Meiotic Sex Chromosome Inactivation (MSCI). This mechanism is a specialization of a more general silencing mechanism, named Meiotic Silencing of Unsynapsed Chromatin (MSUC), that is induced when any pairing problem between homologous chromosomes results in asynapsis, in male as well as female meiotic prophase cells. Here, we demonstrate that in addition to asynapsis, the persistent presence of DNA repair foci is a hallmark of meiotic silencing. In addition, we show that SPO11-independent DNA repair foci form during normal oocyte development. We propose that these foci represent sites of unrepaired DSBs that are capable of inducing transcriptional silencing, irrespective of synapsis.

During meiotic prophase in yeast and mammals, the induction of DNA double-strand breaks (DSBs) by the transesterase SPO11 precedes stable pairing and synapsis of homologous chromosomes [1], [2]. Synapsis between chromosomes is achieved by the formation of a specific protein complex, consisting of lateral elements along the chromosomal axes that contain SYCP2, SYCP3 [3], [4], different components of the cohesin complex [5], [6], and (before synapsis is achieved, on axial elements) the HORMA-domain proteins HORMAD1 and HORMAD2 [7], [8]. Lateral elements are connected by a central element containing SYCP1 [9] and several other meiosis-specific proteins, including SYCE1, SYCE2 [10] and TEX12 [11]; reviewed by Yang and Wang [12]. Parallel to synaptonemal complex formation, meiotic DSBs are repaired, thereby facilitating homologous chromosomes interaction and the achievement of complete synapsis. In male mammals, the X and Y chromosomes form a very special pair; they can synapse only in their short pseudoautosomal regions, and localize to the periphery of the nucleus. Furthermore, the XY chromatin is silenced, forming the XY body, by a process named meiotic sex chromosome inactivation (MSCI). This requires the expression of the histone variant H2AX [13]. The checkpoint kinase ATR phosphorylates H2AX at S139, generating γH2AX [14]. γH2AX is the earliest known marker of MSCI. This specific histone modification is also found in somatic cells, usually at sites of DNA DSB repair [15]. Interestingly, H2AX phosphorylation in response to DNA damage has been coupled to reduced levels of RNA polymerase II activity in somatic cells [16]. MSCI is considered a specialized form of a more general mechanism termed meiotic silencing of unsynapsed chromatin (MSUC), which silences unsynapsed chromatin in male and female meiotic prophase cells [17]–[19]. The exact cascade of events that leads to this transcriptional silencing is not known, but it has been established that there is a tight correlation between the presence of unsynapsed chromosomal axes coated by HORMAD1 and HORMAD2 (the two mammalian orthologs, of the yeast protein Hop1 [7], [20], [21]), the accumulation of ATR along these axes, the formation of γH2AX, and the transcriptional silencing. Indeed, it was recently reported that efficient accumulation of ATR on the XY body requires the HORMAD1 and HORMAD2 proteins [22], [23]. Many DNA repair proteins accumulate at the XY body, together with histone modifications such as specific methylation, sumoylation and ubiquitylation (reviewed by Inagaki et al. [24]). The accumulation of DSB repair proteins may be caused by delayed or stalled DSB repair in regions that fail to synapse. Persistent meiotic DSBs can indeed be observed on the X, but not on the Y chromosome, via immunocytochemical detection of the homologous recombination proteins RAD51 and its meiosis-specific paralogue DMC1 [25]–[28]. RAD51 and DMC1 have DNA-dependent ATPase activity and form filaments on single-stranded resected DNA-ends at DSB repair sites, and are essential for homologous recombination repair in mitotic and meiotic cells, respectively [29]–[32]. Evidence for a relationship between meiotic DSBs and homologous synapsis is provided by the observation that synapsis is severely affected in the absence of SPO11-induced meiotic DSBs [33], [34]. Some heterologous synapsis can occur in Spo11 knockout meiocytes, but both spermatocytes and oocytes do not proceed beyond a zygotene-like stage [33], [34]. In Spo11 knockout spermatocytes, a pseudo XY body is formed, which most often does not localize to the X and Y chromosomes, but to part of the unsynapsed chromatin [35], [36]. It has been defined as a condensed chromatin structure that, similar to the XY body, is marked by γH2AX and ATR, and is transcriptionally silenced [35], [37]. Based upon these characteristics, it has been proposed that the pseudo XY body is a manifestation of MSUC [37]. However, in Spo11 knockout spermatocytes, HORMAD1 and HORMAD2 coat all unsynapsed axes, but the pseudo XY body forms only on part of the unsynapsed chromatin, indicating that somehow the MSUC response is not complete [7], [8] In addition, although more than 60% of the spermatocyte nuclei in Spo11 knockout testes contain a pseudo XY body, only 11% show clear accumulation of ATR along the unsynapsed axes in the pseudo XY body, compared to 100% ATR accumulation along the axes of true XY bodies in wild type spermatocytes [23]. The restriction of MSUC to only part of the unsynapsed chromatin is surprising, and raises the possibility that, apart from asynapsis, also other mechanisms may contribute to the activation of MSUC and MSCI. Since all known players in these processes function also in DNA repair we hypothesized that persistent DSBs on unsynapsed axes may contribute to the activation of MSUC and MSCI. This would then suggest that, even in the absence of SPO11, perhaps some damage-induced DSBs are frequently present, and could play a role in restricting the MSUC response to those areas that contain both unsynapsed axes and DNA damage. This notion is supported by the fact that radiation-induced DSBs in mouse leptotene cells enhance the efficiency of MSUC of a small translocation bivalent that carries a heterologous region of approximately 35–40 Mb [38]. In addition, recent data also provide a link between DSB repair, the checkpoint kinase ATM, and transcriptional silencing of surrounding chromatin in somatic cells [39]. Herein, we have generated a mouse model with a point mutation, which inactivates the catalytical site of SPO11. We used this mouse model to obtain more insight in the relation between the presence of DSBs and MSUC. As expected based on our hypothesis, we found that SPO11-independent DNA repair foci are present in spermatocytes and oocytes. Moreover, we observed de novo induction of DNA repair foci in zygotene-like SPO11-deficient oocytes. Together with the results of a thorough analysis of the relationship between the localisation of DSB repair proteins and the MSUC response, our data reveal a direct link between the presence of persistent damage and the activation of MSUC and MSCI. We used a Spo11 knock-in mouse model in which the catalytically active tyrosine (Tyr) 138 residue is replaced by a phenylalanine (Phe) (Spo11YF/YF) (Figure S1A, B). In yeast and plants, mutation of the analogous Tyr residue abolished meiotic DSB formation [40]–[42], and a similar mouse mutant was recently described [43]. Presence of the point mutation and normal expression of the mutant protein were verified by sequence analyses, RT-PCR, and Western blot analyses (Figure S1C, D, E). The amount of mutant and/or wild type SPO11 protein in the testis of +/+, +/YF and YF/YF animals was comparable. Identical to the Spo11 knockout [33], [34], male and female Spo11YF/YF mice are infertile, and leptotene and zygotene nuclei display global absence of markers of DSB formation and repair (Figure 1A, B, and C). Spermatocytes and oocytes reach a zygotene-like stage with variable degrees of heterologous synapsis (Figure S2A, B, C). We analyzed the formation of meiotic DSBs in wild type, heterozygote and homozygote Spo11YF/YF mice through immunocytochemical analysis of the formation of RAD51 foci. The number of RAD51 foci was quantified in leptotene and zygotene spermatocyte and oocyte nuclei (Figure 1). In wild type leptotene, many DSBs are formed, concomitant with the assembly of short patches of axial element along the chromosomal axes (Figure 1A, B, left panels, 1C). In zygotene, repair of meiotic DSBs occurs, parallel to the pairing of homologous chromosomes. Axial elements of paired homologous chromosomes then synapse (and are therefore termed lateral elements), through the formation of the central element of the synaptonemal complex (SC) (Figure 1A, B, left panel). The number of RAD51 foci gradually decreases, from leptotene to zygotene (Figure 1 A, C), as has been observed before [26]. It should be noted that, in mouse, male meiosis induction occurs throughout postpubertal life, whereas female meiosis is initiated only once during embryonic development (around embryonic day 13 (E13)). Oocytes progress through leptotene and zygotene in 15–20 h [44], [45]. At E17, the vast majority of oocytes has reached the pachytene stage, and around E19, oocytes enter diplotene, reaching the first meiotic arrest. Spermatocytes require a longer time span between leptotene (induction of DSBs) and early pachytene (synapsis) of approximately 48 h [46]. In Spo11+/YF leptotene spermatocyte nuclei, the number of RAD51 foci was approximately 30% lower compared to wild type (Figure 1C). However, in zygotene nuclei, no difference in the number of RAD51 foci between wild type and heterozygote nuclei was observed (Figure 1C). Similar to the males, the number of RAD51 foci was lower in Spo11+/YF leptotene oocytes, compared to the wild type, and a small difference between the wild type and heterozygote oocytes was still observed at zygotene (Figure 1C). MLH1 is mismatch repair protein that is a well-known marker of crossover sites [47], and functions in the resolution of joint molecules at the final phase of crossover formation [48]. The number of MLH1 foci was not different between wild type and Spo11+/YF spermatocytes (Figure 1D). In Spo11YF/YF animals, a few RAD51 foci were observed on the axial elements in leptotene and zygotene-like spermatocytes (average foci number 12±4.4, n = 54) and oocytes (average foci number 5±3.7, n = 50) (Figure 1A–C). Surprisingly, from E17.5 onwards, when oocytes should have reached the pachytene stage, we observed de novo RAD51 accumulation (Figure 1E), in oocytes from Spo11YF/YF mice. These RAD51 foci formed along most of the length of one or more axes (Figure 1B, lower panel, right). Such marked accumulation of RAD51 is also observed in wild type and Spo11+/YF pachytene oocytes (Figure 1B, lower panel, left), but in a relatively small proportion of the nuclei (around 20%, see also below). To confirm the specificity of this pattern of RAD51 accumulation, we also used a commercial RAD51 antibody previously reported to mark RAD51 foci in spread meiotic nuclei [49]. This antibody yielded a similar pattern of RAD51 accumulation in oocytes (Compare Figure 1B to Figure S3). To ensure that the RAD51 foci that are observed in Spo11YF/YF spermatocytes and oocytes are not caused by remnant SPO11 activity, we also analysed RAD51 localisation in Spo11 knockout meiocytes. As expected, the pattern of RAD51 foci staining in Spo11 knockout spermatocytes and oocytes was similar to what was observed in meiocytes of Spo11YF/YF animals (Figure S4). This confirms that the observed RAD51 foci in our Spo11YF/YF model are SPO11-independent. Extensive asynapsis is thought to elicit an MSUC response, which can be observed in Spo11−/− spermatocytes as a γH2AX positive domain in the nucleus [36], [37]. This domain has been termed pseudo XY body, since it does not necessarily include chromatin from the X and Y chromosomes. Similar to what has been described for Spo11 knockout mice, we observed one or two pseudo XY bodies in late zygotene-like spermatocytes from Spo11YF/YF mice (Figure S5A). In addition to γH2AX, other components of the DNA repair machinery are known to accumulate on the unsynapsed axes of the pseudo XY body (BRCA1, TOPBP1), or on the surrounding chromatin (MDC1) in Spo11 knockout spermatocytes [37], [50], and this was also observed for the pseudo XY bodies in Spo11YF/YF spermatocytes (Figure S5B–D). As recently reported, pseudo XY body-like structures can also be detected in Spo11 knockout oocytes [22], and even wild type oocytes have been reported to contain a MSUC region in a small percentage of the pachytene oocytes that fails to correctly synapse all chromosomes [51]. We also observed areas of MSUC in a minority of wild type and Spo11+/YFoocytes at E16.5 and E17.5 (Table 1). In addition, in Spo11YF/YF ovaries we observed a γH2AX-positive chromatin domain in about 14% of oocytes at E16.5 (Table 2), and in more than 80% of oocytes derived from Spo11YF/YF ovaries at E17.5 (Table 2). The transcriptional silencing in the XY body can be immunocytochemically visualized as an area that is relatively depleted of RNA polymerase II [17]. To verify that the γH2AX domain detected in SPO11-deficient spermatocytes and oocytes is a transcriptionally silenced region, we performed RNA polymerase II (RNA pol II) staining and indeed observed a depletion of this enzyme from the areas enriched for γH2AX in Spo11−/− and Spo11YF/YFspermatocytes and oocytes (Figure 2A and B). To verify the results, we quantified the relative average intensity of RNA pol II staining in the γH2AX domain in oocytes, and compared it to the relative intensity in the true XY body of wild type pachytene spermatocytes (Figure 2C). Despite the fact that we observed variable depletion levels within each of the three analysed categories, the relative average level of RNA pol II in γH2AX domains of wild type (0.77±0.16, n = 30) and Spo11YF/YF (0.76±0.18, n = 30) oocytes is similar, and also comparable to what is observed for the XY body in male wild type spermatocytes (0.69±0.14, n = 30) (Mann-Whitney, confidence interval p<0.001), indicating a significant transcriptional silencing. Based on these results, we will refer to the γH2AX domains that are observed in both Spo11YF/YF and Spo11+/+ oocytes as pseudo XY bodies. Having established that both SPO11-independent DNA repair foci and pseudo XY bodies are present in SPO11-deficient spermatocytes and oocytes, we subsequently analysed whether these foci are indeed associated with the MSUC areas. Such an association would be expected, if SPO11-independent DNA damage, present on part of the unsynapsed axes, plays a role in nucleating the formation of the pseudo XY body. To investigate this, we performed co-immunostaining experiments for RAD51 to visualize DSB repair sites, γH2AX to visualize the pseudo XY body and SYCP3 to assess the stages of the cells. Due to the severe impairment of meiotic prophase progression in Spo11YF/YF animals, spermatogenesis is arrested at stage IV, but spermatocytes never reach a true pachytene stage. We performed our analyses on a subpopulation of spermatocytes which displayed one or more areas of (heterologous) synapsis and showed no signs of SC fragmentation, in order to select healthy spermatocytes which had already entered the zygotene stage. First of all we determined the frequency of spermatocytes with RAD51 foci and with a pseudo XY body. We split our population (n = 240) in four classes (Figure 3A): 1) cells having both a pseudo XY body and RAD51 foci; 2) cells having only a pseudo XY body; 3) cells having only RAD51 foci; and 4) cells lacking both a pseudo XY body and RAD51 foci (Figure 3A, B). The results indicate that the vast majority of nuclei (78.3%) contain both a pseudo XY body as well as RAD51 foci. Although RAD51 is a well-known marker of sites of DSB repair [52], it may also accumulate on ssDNA that is formed in a different context of DNA damage, such as observed during collapse of a replication fork in S phase [53]. To obtain additional evidence for the presence of DNA damage in Spo11YF/YF spermatocytes, we performed the same analysis by staining for two more markers of DNA damage and repair: DMC1 and RPA. DMC1 is the meiosis-specific homolog of RAD51 which participates in the process of repair of meiotic DSBs via homologous recombination. Hence, we expected the results for DMC1 and RAD51 to be similar. Indeed, comparable percentages of the analyzed nuclei were found to fall in each of the four classes (Figure 3A). In addition, we observed colocalization between RAD51 and DMC1 foci in the γH2AX domains (Figure S6A). Unlike RAD51 and DMC1, RPA is not a recombinase but a single-stranded DNA (ssDNA) binding protein which takes part in many processes involving DNA metabolism (reviewed by Sakaguchi et al. [54]). At meiotic DSBs, the dynamics of RPA foci differ from those of DMC1, and although both proteins are enriched on the XY body, this occurs at different developmental time points (Figure S7). Nevertheless, similar to what was found for RAD51 and DMC1, 72.3% of the cells (n = 108) showed presence of both RPA foci and γH2AX domains (Figure 3A, lower panel). The high percentages of cells with a pseudo XY body and DNA damage markers, provided an indication for a possible correlation between the presence of DNA damage, in particular DSBs, and the formation of the pseudo XY body. To further test the hypothesis for such a correlation, we determined the colocalization between each DNA repair marker and the γH2AX domain, in the fraction of spermatocytes that was positive for both of these features. We counted similar average numbers of RAD51, DMC1 and RPA foci (5.7, 5.2 and 6.4, respectively) in the nuclei, and the percentages of colocalization with the γH2AX domain(s) ranged between 70.8% (RAD51) and 82.2% (DMC1) (Figure 3B). Furthermore, up to 89–98% of the analysed pseudo XY bodies contained at least one focus of RAD51, DMC1 or RPA (Figure 3B). To validate that the frequent localization of RAD51 in the pseudo XY body is not coincidental, we compared the relative area of the nucleus that was positive for γH2AX (pseudo XY body) to the fraction of RAD51 foci that was found inside that area. We observed that the fraction of RAD51 that localized inside the pseudo XY body (more than 70%) was much larger than the fraction of the nucleus that was taken up by this chromatin domain (20% of the total area). In addition, there was no specific correlation (Pearson linear correlation coefficient [Pcorr] = 0.0704) between the size of the pseudo XY body and the percentage of RAD51 foci that was found in the pseudo XY body (Figure 3C). In Spo11 knockout spermatocytes, a similar pattern of colocalization between RAD51, DMC1, and RPA foci and the pseudo XY body was observed (Figure S8). The localised presence of DNA repair foci in one or a few pseudo XY bodies indicates that DNA damage in spermatocytes tends to concentrate in a single, transcriptionally silenced area. To test this hypothesis, we induced exogenous DSBs in Spo11YF/YF spermatocyte nuclei by whole-body irradiation, and analysed the presence of DSB markers at different time points following the treatment. We observed approximately 120 (±5.3, n = 30) RAD51 foci and a nucleus-wide accumulation of γH2AX at 1 h following irradiation. Interestingly, 48 hours after irradiation, we still observed extensive H2AX phosphorylation emanating from the many RAD51 foci (Figure 4A). However, 120 h following irradiation, when cells that were irradiated at leptotene would have progressed to pachytene in a wild type background, a pseudo XY body was observed in about 90% (n = 70) of the analysed nuclei (Figure 4B). These pseudo XY bodies always contained RAD51 foci (25.1±1.73, n = 50), and the majority of the radiation-induced RAD51 foci that are still present at this time point (65.7%) localized in the pseudo XY body (Figure 4A). These data show that the persistent radiation-induced DSBs tend to relocalize in a specific nuclear subdomain. This phenomenon is in accordance with the colocalization of unsynapsed or partially synapsed translocation chromosomes, carrying persistent meiotic DSBs, with the XY body [38]. To confirm that the pseudo XY body in these irradiated spermatocytes is an MSUC area, as observed in non-irradiated Spo11YF/YF spermatocytes, we performed co-immunostaining for γH2AX and RNA pol II. We detected a depletion of this enzyme in the areas enriched for γH2AX, indicating that they are transcriptionally silenced (Figure 4C). Next, we asked if RAD51, DMC1, and RPA foci also preferentially localized in the pseudo XY bodies in E17.5 Spo11YF/YF oocytes. As discussed above, RAD51 was found to accumulate extensively on some chromosomal axes, often coating them completely, so that single foci could not be easily resolved. Such marked accumulation was not observed for DMC1 or RPA, which are forming fewer foci (average number of 5.6±2.3, n = 20 and 7.4±6.9, n = 30, respectively). Despite this difference in foci pattern, the percentage of oocyte nuclei that contained both a γH2AX domain and RAD51 foci (79.2%, n = 120) was similar to the percentage of oocyte nuclei with a γH2AX domain and RPA foci (83.1%, n = 89) (Figure 5A, upper and lower panel respectively). In contrast, only 25.9% of the analysed Spo11YF/YF oocytes (n = 54) displayed DMC1 foci, but all these cells also had a γH2AX domain. The rest of the nuclei had only a pseudo XY body (57.41%) or were negative for both DMC1 and γH2AX (16.67%) (Figure 5A, middle panel). In the group of nuclei that contained both RAD51 foci and a γH2AX domain, the pseudo XY body always contained RAD51 foci that coated part of the axes (Figure 5B). Also, in E17.5 Spo11YF/YF oocytes that contained a pseudo XY body and DMC1 or RPA foci, more than 90% of the pseudo XY bodies contained DMC1 or RPA foci, respectively. Conversely, the vast majority of RAD51, DMC1, and RPA foci in this subgroup of nuclei were located in the pseudo XY body, similar to what was observed for Spo11YF/YF spermatocyte nuclei. Furthermore, the DMC1 foci were found to colocalize with some of the (more abundant) RAD51 foci in the pseudo XY bodies of oocytes (Figure S6B). For comparison, these analyses were also performed on Spo11 knockout E17.5 oocytes and this provided similar results (Figure S8, right). Interestingly, also in wild type and Spo11YF/+ oocyte nuclei, RAD51 coats the axial elements in γH2AX-positive domains (Table 1). These pseudo XY bodies were observed in approximately 20% of pachytene oocytes, similar to what was previously reported by Koutznetsova et al. [51] who observed BRCA1 and ATR on unsynapsed axes in around 15% of the oocyte population from E17 wild type embryos. To analyse this further, we studied the localisation of other proteins involved in homologous recombination (DMC1 and RPA) in relation to the formation of a γH2AX domain. Again we divided the oocyte population in four subgroups, based on the detection of γH2AX and the three DNA repair markers. As expected, the majority of pachytene oocytes showed complete synapsis of all chromosomes and no clear γH2AX-positive domain. Around 20–30% of nuclei showed pseudo XY bodies, as defined by the presence of one or a few distinct γH2AX-positive domains (Figure 6A). Approximately half of the pachytene nuclei lacked both γH2AX domains and RAD51 or DMC1 foci, whereas no nuclei were found without RPA foci (Figure 6A, B). We did not observe any pseudo XY body in nuclei without RAD51 foci, but 13% of the nuclei contained a γH2AX domain but no DMC1 foci (Figure 6A). RPA is known to mark DSB repair spots after RAD51-mediated strand invasion and during homologous recombination, to protect the ssDNA regions generated during this process [55]. This explains the fact that RPA foci are always present in E17.5 oocyte nuclei which are at a mid-meiotic stage and have not yet completed the homologous recombination process at all DSB repair sites. Also, since RPA is engaged in completing recombination at synapsed autosomal sites, a relatively small fraction of the RPA foci colocalizes with pseudo XY bodies. In contrast, most DMC1 and RAD51 foci localize to γH2AX domains, similar to what was found for Spo11YF/YF oocyte nuclei (Figure 6B), although DMC1 foci are found more frequently and in higher numbers in pseudo XY bodies in Spo11+/+ compared to Spo11YF/YF oocytes. DMC1 foci colocalized with RAD51 foci when both were present in the pseudo XY body (Figure S6C). Since we observed some differences between the patterns of RAD51 and DMC1 accumulation in pseudo XY bodies of wild type oocytes, we wondered whether pseudo XY bodies that contain both DMC1 and RAD51 foci differ from those that show only RAD51 foci. First, we analysed the relation between DMC1 accumulation, formation of the pseudo XY body and synapsis, using an antibody directed against the central element protein TEX12. The results in Figure 7A and B show that DMC1 foci in oocyte pseudo XY bodies localize mainly (58.6%) on unsynapsed axes (inferred from the absence of TEX12, and placement of DMC1 foci in an axis-like pattern), and rarely (12.8%) on synapsed areas (Figure 7B). It is important to note that 28.6% of oocytes with a pseudo XY body did not show any DMC1 foci (Figure 7A, B) and that all these nuclei were also characterized by complete synapsis (based on the presence of 20 TEX12-positive bivalents) (Figure 7B). In contrast, RAD51 always coats the chromosomal axes of the pseudo XY body, irrespective of synapsis (Figure 7C). These observations prompted us to further analyse the occurrence of pseudo XY bodies in association with complete synapsis. For this, we used an antibody directed against the HORMAD1 protein, together with anti-TEX12 as well as anti-γH2AX to identify the pseudo XY body. As reported previously, HORMAD1 covered all unsynapsed axes at zygotene, and was lost once the cells reached complete synapsis at pachytene [8] (Figure 8A). Conversely, TEX12 gradually accumulated as synapsis progressed, consistent with earlier reports [11] (Figure 8A). When we analysed the pachytene population in more detail, we observed unsynapsed axes that were positive for HORMAD1 in a pseudo XY body in 9.8% of the pachytene nuclei, and another 13.1% that showed partial (5.7%) or no (7.4%) colocalisation of the pseudoXY body with HORMAD1 (Figure 8B). Whenever HORMAD1 was absent from the pseudo XY body, TEX12 was present, indicating complete synapsis. To verify that synapsis was complete in the nuclei that lacked HORMAD1 but contained a pseudo XY body, we measured the total length of synapsed axes, visualized as TEX12 stretches, in pachytene oocyte nuclei. We found that the total SC length was comparable in pachytene oocytes without any HORMAD1 staining, independent of the presence of a pseudo XY body. On the contrary, the total synapsis length was significantly lower in pachytene oocyte nuclei which showed both a pseudo XY body and HORMAD1 (Figure 8C). Finally, to confirm that these pseudo XY bodies elicit true meiotic silencing, despite the absence of asynapsis, we performed a triple staining for RNA polII, TEX12 and γH2AX. As shown in Figure 8D and 8E, RNA polII is depleted from the pseudo XY body, irrespective of synapsis. A point mutation in the Spo11 gene that results in the replacement of Tyr 138 by Phe in the catalytic site of the enzyme leads to the absence of detectable SPO11-dependent meiotic DSBs in oocytes and spermatocytes. This observation is in accordance with recent findings of Boateng et al. [43], who analysed a mouse mutant carrying a mutation in the Spo11 gene that leads to replacement of both Tyr 137 and Tyr138 by Phe. Although having half the amount of functional SPO11 is sufficient to generate a normal number of crossovers, as evidenced by the analysis of MLH1 foci in Spo11+/YF spermatocytes and oocytes, the dynamics of DSB induction was clearly altered. The lower number of RAD51 foci that was observed in leptotene Spo11+/YF oocytes and spermatocytes may indicate that fewer breaks are made. However, near normal numbers of RAD51 foci are observed in zygotene Spo11+/YF spermatocytes and oocytes. These data are consistent with the homeostatic control mechanism that has been observed in yeast [56] and mouse spermatocytes, allowing maintenance of normal crossover frequencies when the number of DSBs is reduced [57]. In addition, or alternatively, the recently identified feedback mechanism, requiring ATM activity, which regulates the number of breaks that can be formed by SPO11 [58] may ensure that a similar level of DSB formation is reached in the heterozygote, albeit with different kinetics when compared to the wild type. In the absence of SPO11, no meiotic DSBs are formed, and accumulation of RAD51, DMC1 and RPA proteins is therefore not expected. Nevertheless we observed significant numbers of RAD51, DMC1 and RPA foci in Spo11YF/YF and Spo11−/− oocytes and spermatocytes that preferentially localized in the pseudo XY body, identified on the basis of the γH2AX staining pattern. In Spo11YF/YF oocytes, we observed a clear increase in the number of RAD51 foci in oocytes at E17.5, compared to oocytes at E16.5. However, the number of DMC1 and RPA foci was much lower than the number of RAD51 foci in these nuclei. The number of DMC1 foci in particular would be expected to follow the same pattern as RAD51, because DMC1 has been reported to participate in the formation of recombination filaments [26]. Nevertheless, it has been recently shown that the dynamics of accumulation of DMC1 and RAD51 are different when extra DSBs are induced by a supplemental copy of the SPO11β-isoform [57]. Cole et al. [57] suggested that, in this situation, the extra DSBs may be more likely to engage in a mitotic pathway of HR repair, and thus less likely to recruit DMC1. In oocytes that completely lack a synaptonemal complex, DMC1 was found to be lost from persistent DSBs, whereas RAD51 foci were still observed [59]. Based on this, it was suggested that DMC1 can only stably associate with meiotic DSBs in the context of synapsed chromatin and normal progression of repair [59]. Our own observations also indicate that DMC1 is lost from SPO11-induced DSB repair sites before RAD51 (data not shown). Together, these observations are in accordance with the notion that the sites that recruit RAD51 foci in E17.5 oocytes can no longer recruit DMC1 with equal efficiency. This may be due to differences in the composition of the repair complexes at (persistent) DSBs in late compared to early pachytene oocytes, or is possibly caused by a drop in the level of DMC1 protein expression. It is important to establish if the DNA repair foci represent actual sites of DNA damage. The increase in the number of RAD51 foci in oocytes between E16.5 and E17.5 may be due to a DNA-damage independent association of RAD51 to chromosomal axes, or foci formation might be induced by the specific chromatin structure that is formed upon γH2AX formation, which would explain why the foci tend to colocalize in a single subnuclear region. However, we have observed that radiation-induced DSBs, that localize throughout the nucleus, first lead to a nucleus wide accumulation of γH2AX, and subsequently to a more concentrated presence of RAD51 foci and γH2AX in a specific subdomain of the nucleus (the pseudo XY body). In addition, it is known that in spermatocytes that carry autosomes with a pairing problem, meiotic DSBs persist on the unsynapsed regions, in association with MSUC, and these regions then also tend to colocalize with the XY body, indicating that persistent DSBs in the context of MSUC have a tendency to reside together in a single nuclear area [38]. The preferred presence of DMC1 and RPA in addition to RAD51 in the pseudo XY bodies supports the hypothesis of the presence of a DNA damage event. One particular feature of the SPO11-independent repair foci in Spo11YF/YF oocytes is their inefficient processing. In fact, in oocytes from E17.5 Spo11YF/YF mice, RAD51 appears to coat unsynapsed axial elements, so that individual foci are no longer clearly observed, indicating that the RAD51 filament formation is not regulated as in a normal homologous DSB repair event. Upon replacement of RPA by RAD51/DMC1, and subsequent persistence of a DSB without further processing to a recombination intermediate, such an axis-wide pattern for RAD51 may develop, possibly due to an abnormal regulation of the foci dynamics, compared to conventional DSB repair events. The spreading of RAD51 along axial elements may result from spreading of RAD51 onto double-stranded DNA, a phenomenon that has also been described for persistent DSBs in yeast [60]. Based upon these considerations, we favour the conclusion that the SPO11-independent DNA repair foci represent true sites of persistent DNA damage. To explain what might cause spontaneous DNA damage in Spo11YF/YF and knockout spermatocytes and oocytes, and possibly also in wild type meiocytes, different mechanisms can be proposed. First, during S phase in somatic cells, and most likely also in meiocytes, DSBs can form at stalled replication forks. In human cells, 50 endogenous DSBs have been proposed to occur in every cell cycle [61]. Most of these DSBs will be repaired before the cells enter G2, but some may persist, and the number of persisting breaks appears to vary between different cell types [62], [63]. A second mechanism that could generate endogenous DSBs is transcription-associated recombination (TAR). The causes of DSBs that form in association with ongoing gene transcription are thought to be related either to generation of stalled replication forks in association with transcription, or to increased accessibility of DNA during transcription, making it more vulnerable to DNA-damaging agents (reviewed by [64], [65]). Meiocytes are post S phase cells, and leptotene, zygotene, and early pachytene spermatocytes and oocytes display a low level of RNA synthesis, making TAR an unlikely source of RAD51 foci in these cells [66], [67]. A third possible endogenous source of DSBs is impaired topoisomerase II activity. Inhibition of topoisomerase II activity in pachytene spermatocytes has been found to result in DSB formation, indicating that topoisomerase II is indeed functional in meiocytes [68]. Fourth, endonuclease activity of ORF2, encoded by Line1 transposons, generates DSBs during the transposition of mobile elements in the genome [69]–[71]. Derepression of transposons has been shown to cause SPO11-independent DNA damage in Mael mutant spermatocytes [72]. In wild type oocytes and spermatocytes, transcription of Line1 elements is transiently derepressed at the onset of meiosis [73]. Finally, we cannot exclude that DNA damage may occur as a result of unknown environmental or endogenous factors such as reactive oxygen species (ROS). ROS generation has been described for normal rat spermatocytes [74], but it is not clear to what extent such damage also results in RAD51 foci formation. In Spo11YF/YF spermatocytes, it appears most likely that some or all of the SPO11-independent RAD51 foci result from carry-over of spontaneous DSBs that were induced in the previous S phase. In oocytes this may also occur, and the observed de novo generation of RAD51 foci post S phase in Spo11YF/YF oocytes indicates that (additional) spontaneous DSBs in oocytes may arise either from impaired topoisomerase II activity or from ORF2 mediated endonuclease activity in cells that should have progressed already to pachytene. Such SPO11-independent DNA damage may also be induced in wild type pachytene oocytes, but the close proximity of the homologous template in these oocytes may facilitate homologous recombination repair of most of the de novo induced DNA damage. In Spo11YF/YF oocytes the appropriate template for repair is not directly available due to almost complete lack of homologous chromosome pairing. This difference in homologous template availability readily explains the higher relative frequency of pseudo XY body formation in Spo11YF/YF oocytes compared to oocytes from wild type or heterozygote littermate controls. At present, it is not clear whether the persistent repair foci are resolved at some later time point, or whether the persistent presence of these foci and the associated γH2AX signaling triggers a checkpoint that induces apoptosis. Daniel et al. [22] reported increased apoptosis of oocytes in ovaries of newborn Spo11 knockout mice compared to controls. In addition, it has been reported that only 10–20% of the normal number of oocytes is present in Spo11 knockout ovaries at postnatal days 4 and day 8 [34], [75]. This percentage nicely corresponds to the 19% of oocytes that do not contain a pseudo XY body at E17.5 in our Spo11YF/YF model. However, although these data confirm that oocytes with a pseudo XY body are lost shortly after birth, cell death may also be caused by a so-called synapsis checkpoint, mediated by HORMAD proteins, rather than by a DNA repair checkpoint [22], [23], [76]. Our analyses of RAD51 and DMC1 foci in relation to MSUC and synapsis in pachytene oocytes from Spo11+/YF and wild type E17.5 embryos has shown that two different types of equally silenced pseudo XY bodies exist in wild type pachytene oocytes. Approximately two-third of the pseudo XY bodies accumulate DMC1 as well as RAD51 and form on unsynapsed chromatin (Type I), whereas one-third accumulate RAD51, but little or no DMC1, and form on synapsed chromatin (Type II). We propose that the Type I pseudo XY bodies represent sites that contain persistent SPO11-induced DSBs in areas that failed to synapse, whereas the Type II pseudo XY bodies represent sites where SPO11-independent damage has persisted that elicited a MSUC response, independent of synapsis. The percentage of cells with γH2AX accumulation in a pseudo XY body is highly reduced in Spo11−/− Hormad1−/− or Spo11−/− Hormad2−/− double mutant spermatocytes [22], [23]. This illustrates the important role of HORMAD proteins in the MSUC response. Yet the localization of HORMAD1 to all unsynapsed chromatin in Spo11 knockout spermatocytes [22], [23]), and the presence of some nuclei with a proper MSUC response in Spo11−/− Hormad1−/− spermatocytes indicate that, apart from HORMAD proteins, an additional localizing event is needed for pseudo XY body nucleation. Taken together, these and our observations support the hypothesis that both asynapsis, detected by HORMADs, and persistent SPO11-independent DNA repair foci are involved in the induction of H2AX phosphorylation and the establishment of meiotic silencing in pseudo XY bodies in Spo11YF/YF oocyte nuclei. We would like to propose that MSCI in wild type spermatocytes is then also triggered by both persistent DSBs, in this case SPO11-dependent, and the presence of unsynapsed chromatin (schematically presented in Figure 9). If RAD51 accumulation is as extensive as observed in pseudo XY bodies in oocytes, HORMADs may not even be required, and enough ATR may be recruited by the DNA repair machinery itself, to elicit the MSUC response, as indicated by the existence of pseudo XY bodies that lack HORMAD1 in oocytes. Despite the more prominent RAD51 accumulation on axes of the pseudo XY body in oocytes as compared to spermatocytes, we propose that the mechanism of pseudo XY body formation in Spo11YF/YF spermatocytes occurs in a similar fashion. The differences in the pattern of RAD51 accumulation may be caused by the fact that Spo11YF/YF spermatocytes are eliminated at stage IV of the spermatogenic cycle, whereas Spo11YF/YF oocytes appear to proceed normally throughout the stage that should correspond to pachytene and are eliminated later [34]. Perhaps, the few spontaneous DSBs in Spo11YF/YF spermatocytes modulate the MSUC response in a slightly different way, compared to the responses elicited by the more extensive accumulation of endogenous DSBs in Spo11YF/YF oocytes. Still, the MSUC response in both Spo11YF/YF spermatocytes and oocytes is characterized by the same intense γH2AX accumulation and by the presence of RAD51/DMC1 and RPA foci. It is interesting to note that such foci can also be observed on the unsynapsed axes of the X chromosome in wild type spermatocytes, as a hallmark of persistent DSBs. HORMAD proteins may be instrumental to spread the MSUC response along the chromosomal axes into areas that lack persistent DSBs, such as the Y chromosome. In somatic cells, formation of γH2AX chromatin domains has also been coupled to transcriptional silencing, in the context of radiation-induced damage [16]. More recently, Shanbhag et al. [39] analysed the effect of persistence of an endonuclease-dependent DSB on transcriptional activity in the neighbouring genes. They observed that H2AX phosphorylation spreads along the DNA surrounding the DSB, and that the accumulation of this histone modification correlated with reduction of RNA polymerase II activity. Persistent DSBs were shown to trigger the silencing of neighbouring genes, and the mechanism was termed DSB-induced silencing in cis (DISC) [39]. This mechanism, that occurs in somatic cells, might have some aspects in common with MSUC and MSCI in meiocytes. In conclusion, this study has revealed the presence of SPO11-independent DNA repair foci in oocytes and spermatocytes. In addition, we show that unrepaired DSBs most likely are the initial trigger of both MSCI and MSUC in spermatocytes and oocytes. For wild type oocytes, the possible presence of de novo induced DNA damage in a substantial part of the oocyte population may contribute to the massive loss of such oocytes around birth. For spermatocytes, the few SPO11-independent breaks that are present will most likely be rapidly repaired once homologous chromosome pairing is obtained with the help of the 200 or more SPO11-induced DSBs. The MSUC and MSCI response may be less unique than previously thought, and actually represent an extreme and adapted form of DISC. Therefore, knowledge about the molecular basis of meiotic silencing may also be relevant for our understanding of DNA damage-induced chromatin modifications in somatic cells. All animal experiments were approved by the local animal experiments committee DEC Consult. All animals were housed in IVC cages under supervision of the Animal Welfare Officer. Any discomfort of animals was daily scored by the animal caretakers. No more than mild or moderate discomfort of animals was expected from the treatments, and no unexpected discomfort was observed. All animal experiments were approved by the animal experiments committee DEC-Consult. Spo11 mutant mice were generated through a two-step recombination strategy as described by Soukharev et al., [77]. First, two heterospecific lox sites flanking the selectable marker hygromycin, replacing exons 4–8, were placed in the Spo11 gene, in ES cells by homologous recombination. Next, a targeting vector containing the same heterospecific lox sites flanking exon 4–8 of Spo11 with the point mutation generating Y138F at exon 4 was used to replace the selection marker by a site-specific double cross-over event (Figure S1A). The final modified Spo11 locus carries a loxP site between exons 3 and 4, the point mutation generating Y138F at exon 4, and a lox511 site between exons 8 and 9. ES cells carrying a single modified Spo11 allele were used for blastocyst injection to generate chimeras, and heterozygotes were obtained upon germ line transmission of the mutated allele. Correct targeting was verified using Southern blotting with 5′and 3′probes outside the targeted region (Figure S1B), and sequencing (Figure S1C). This Spo11 allele has been registered at Mouse Genome Informatics (MGI) as Spo11<tm1Bdm> (Allele Accession ID: MGI:5432496). Wild type, heterozygote and homozygote Spo11 mutant mice were kept on a FVB background. To genotype the animals, the following primers were used: forward, 5′CTGGTCGATGCAGATCCCTACGG3′; reversed, 5′TAGATGCACATTATCTCGATGCC3′ (Figure S1B) Spo11 knockout mice carried the Spo11tm1M allele described in [34]. For the analysis of radiation-induced DSBs in spermatocytes, Spo11YF/YF male adult mice were exposed to 5Gy whole body radiation and sacrificed 1 h, 48 h, and 120 h after the treatment to collect the testes. For primary antibodies, we used mouse monoclonal antibodies anti-phosphorylated H2AX, anti-BRCA1, anti-TOPBP1, anti-MDC1, anti-phospho H2AX (all from Upstate), anti-DMC1 (DMC1-specific), anti-RAD51, anti-RNA Polymerase II (all from Abcam); rabbit polyclonal antibodies anti-RAD51 (recognizing both DMC1 and RAD51) [78], anti-RPA (gift from P. De Boer, described in Schaarmidt et al., ([79]), anti-SYCP3 (gift from C. Heyting), anti-HORMAD1 (gift from A. Tóth) and anti-phosphorylated H2AX (Upstate); rat polyclonal anti-SYCP3 [80]; guinea pig anti-TEX12 (gift from Christer Höög). SPO11 antibody (Spo11L56S9) was raised from rabbits immunized with GST-Spo11α produced by the service of recombinant protein of CRBM (UMR5237-CNRS). For secondary antibodies, we used a goat anti-rabbit IgG alexa 405/488/546/633, goat anti-mouse alexa IgG 350/488/546/633, goat anti-rat IgG alexa 546, goat anti-guinea pig 405/555 (Molecular Probes). RNA was extracted and reverse transcribed according to standard procedures. PCR amplifications were performed with forward primer 5′AATAGTCGAGAAGGATGCAACA3′and reversed primer 5′TAGATGCACATTATCTCGATGC3′ Immunoprecipitations were carried out with rabbit polyclonal anti-SPO11 antibody, followed by western blot detection with the same primary antibody and Trueblot secondary antibody (eBioscience). Testes were fixed and stained with hematoxilin and eosin using standard histological methods. Testis tissues were processed to obtain spread nuclei for immunocytochemistry as described by Peters et al. (1997) [81]. Spread nuclei of spermatocytes were stained with antibodies mentioned above. Before incubation with antibodies, slides were washed in PBS (3×10 min), and non-specific sites were blocked with 0.5% w/v BSA and 0.5% w/v milk powder in PBS. Primary antibodies were diluted in 10% w/v BSA in PBS, and incubations were overnight at room temperature in a humid chamber. Subsequently, slides were washed (3×10 min) in PBS, blocked in 10% v/v normal goat serum (Sigma) in blocking buffer (supernatant of 5% w/v milk powder in PBS centrifuged at 14,000 rpm for 10 min), and incubated with secondary antibodies in 10% normal goat serum in blocking buffer at room temperature for 2 hours. Finally, slides were washed (3×10 min) in PBS (in the dark) and embedded in Prolong Gold with or without DAPI (invitrogen). Fluorescent images were observed by using a fluorescence microscope (Axioplan 2; Carl Zeiss) equipped with a digital camera (Coolsnap-Pro; Photometrics). To distinguish zygotenes from aberrant pachytenes, we used specific parameters defined in Figure S9. Aberrant pachytene oocytes,have also been described in previous publications [7], [51], and are characterized by the presence of one to three chromosome pairs lacking synapsis. We also included rare nuclei in which some chromosomes are entangled and not fully synapsed. Normal (late) zygotene nuclei are characterized by a higher proportion of homologs that have not completed synapsis, compared to what is observed in the aberrant pachytenes, and SYCP1/TEX12 patches can be observed which have not yet converged to become a single complete central element. In addition to specific characteristics of the SC, the labelling patterns of the repair associated recombinase RAD51 and phosphorylated H2AX are also helpful to distinguish late zygotenes from aberrant pachytenes. Single, isolated RAD51 foci are observed in zygotene nuclei, whereas multiple closely adjacent foci are present in aberrant pachytenes. H2AX phosphorylation,occurs in a nucleus-wide pattern at zygotene. In contrast, aberrant pachytene oocytes have one to three bright and defined γH2AX domains. Fluorescent images were taken under identical conditions for all slides, and images were analyzed using the ImageJ (Fiji) software (Rasband, W.S., ImageJ, U.S. National Institutes of Health, Bethesda, Maryland, USA [http://rsb.info.nih.gov/ij/]). Confocal imaging was performed on a Zeiss LSM700 microscope (Carl Zeiss, Jena): we used 63× oil immersion objective lens (N.A. 1.4), pinhole 1AU. DAPI was excited at 405 nm and imaged with a short pass filter (SP) 490 nm; Alexa 488 was excited at 490 nm and imaged SP 555 nm; Alexa 546 was excited at 555 nm and imaged SP 640 nm; Alexa 633 was excited at 639 nm and for the imaging no filter was required. Imaging of nuclei immunostained for RAD51 or DMC1 or RPA and SYCP3 was performed with the same exposure time for each nucleus. Images were analysed without any manipulation of brightness and contrast. Foci were subsequently counted using Image J software, including the Fiji plug-in. We used the analyze particles function and set the threshold manually, in order to include the smallest visible focus in the analysis. The average area of one RAD51 focus was assessed to be 40–50 pixels, therefore foci with an area larger than 100 pixels were counted as multiple foci to allow approximate quantification of RAD51 foci also when it was observed as a continuous signal along the axial elements. Measurement of synaptonemal complex length was performed using a homemade ImageJ macro. The macro generates a skeletonized image of the original picture and measures the length of that skeleton. Relative quantification of the RNA polII levels in the (pseudo) XY body was performed comparing the average intensity per pixel area in the γH2AX domain with the average intensity in a non-heterochromatic nuclear area of the same shape and size.

10.1371/journal.pcbi.1006360

ggsashimi: Sashimi plot revised for browser- and annotation-independent splicing visualization

We present ggsashimi, a command-line tool for the visualization of splicing events across multiple samples. Given a specified genomic region, ggsashimi creates sashimi plots for individual RNA-seq experiments as well as aggregated plots for groups of experiments, a feature unique to this software. Compared to the existing versions of programs generating sashimi plots, it uses popular bioinformatics file formats, it is annotation-independent, and allows the visualization of splicing events even for large genomic regions by scaling down the genomic segments between splice sites. ggsashimi is freely available at https://github.com/guigolab/ggsashimi. It is implemented in python, and internally generates R code for plotting.

Efficient visualization of splicing events from RNA sequencing data is key for the computational analysis of alternative splicing. This process, that results in a single gene giving rise to multiple transcripts, is usually illustrated through sashimi plots: a representation of the read coverage and the support of each splicing junction in the region of interest. However, available tools for this purpose present several limitations that significantly hinder their applicability. Among them, dependence on event annotation, visualization difficulties with moderate sample sizes and long introns, use of non-standard file formats or inefficient implementations. With ggsashimi, we comprehensively overcome these flaws and provide the user with a fast, stand-alone application that generates publication-ready sashimi plots. Furthermore, ggsashimi supports the display of aggregated experiments, a crucial feature in order to explore alternative splicing in the era of large RNA sequencing projects.

Alternative splicing is the process through which different combinations of exons of the same gene are selected to produce a variety of mature coding and non-coding transcripts [1]. The genome-wide landscape of alternative splicing can be easily profiled by RNA sequencing (RNA-seq) and tens of thousands of different RNA-seq experiments are now publicly available. While visualization of RNA-seq data is crucial for exploratory data analysis, visualization of splicing events is currently not dynamically integrated in common genome browsers, and stand-alone software are annotation-dependent. Visualizing splicing events is particularly challenging because such events usually occur between two regions, known as splice sites, that are not contiguous on the genome sequence, and can be as distant as tens or even hundreds of kilobases in linear space. The representation of a splicing event implies drawing a connective element that illustrates the presence of a splice junction between two splice sites. The sashimi plot [2] is a very effective and established diagram which combines the information of read coverage along a gene –a signal track– with curves connecting splice sites supported by RNA-seq data. A tool for drawing sashimi plots was initially developed as part of the MISO suite [3], a software that quantifies and compares alternative splicing from RNA-seq experiments. Current popular implementations include a stand-alone utility to create sashimi plots specifically for MISO-indexed splicing events [2] and a built-in available within the Integrative Genomics Viewer, IGV [4]. Thus, the former relies on a proper compatible annotation of the event, while the latter requires IGV installation and the time-consuming uploading of voluminous alignment files. Moreover, both of them represent splicing events for each RNA-seq experiment on a separate line, which hinders the comparison of more than a dozen samples. Like the original tool for sashimi plots [3], the data processing part of ggsashimi is implemented in python. In contrast to the original tool, ggsashimi internally generates an R script which uses the ggplot2 library [5] for the graphical rendering. To ensure reproducibility, it is distributed in a Docker image, which includes the ggsashimi python script and all the required dependencies. In its simplest usage, ggsashimi generates a publication-ready image with a read coverage histogram and curves connecting splice sites, from a single RNA-seq experiment. Curves have variable widths, proportional to the relative number of reads supporting the splice junction. In line with the most utilized bioinformatics file formats, the required input is a standard alignment BAM file (with no special aligner-dependent features), and genomic coordinates indicating the region to display. The BAM file must be coordinate-sorted and indexed in order to efficiently extract the reads from a determined genomic region. Splice junctions are inferred directly from the BAM file, and no prior knowledge of the splicing event is needed. The output of ggsashimi is available in both vector (SVG, PDF) and raster (PNG, JPEG, TIFF) formats. For the latter, the resolution in pixels per inch can be defined by the user. To allow comparisons across multiple experiments, a list of files can be specified and the signal for each experiment is plotted on a separate line. However, with increasing number of samples, visual comparison of separate plots becomes too overwhelming and some form of aggregation is essential. To this end, ggsashimi can aggregate data for hundreds of experiments and represent plots only for the aggregated measures (see Features). Finally, an annotation plot is optionally generated to visualize transcript structures in the specified region. Again, in line with current standards, a Gene Transfer Format (GTF) file is required, with no additional description of the splicing events. Because splicing events often involve short exons flanked by proportionally very large introns, the genomic regions included between two splice sites (inferred from the alignments and not from the annotation) can be optionally shrunk for better graphical representation. We observed that updating the length of the splice junctions to the original length raised to the power of 0.7 usually renders a good balance between the lengths of introns and exons. ggsashimi presents several unique features that distinguish it from its predecessors and make it a useful tool especially for large-scale projects: To illustrate how ggsashimi performs and to compare it with existing implementations, we obtained a set of 12 RNA-seq samples from the ENCODE project [6], publicly available at www.encodeproject.org. Samples were classified into three cell type groups: endothelial, epithelial and mesenchymal. We focused on a single cassette exon (chr10:27044584-27044670) with different levels of inclusion across the three cell type groups (mesenchymal > epithelial > endothelial). For comparison purposes, the genomic region containing the selected splicing event was represented both using ggsashimi and the sashimi-plot built-in available within the IGV Browser (Fig 1). In the case of ggsashimi, aggregation of samples belonging to the same group (through the --overlay option) and shrinkage of intron lengths were applied (see Features), enhancing the visualization of the event. Although the sashimi representation for splicing events is one of the standards for splicing visualization, current implementations present several limitations that narrow substantially its application. We believe that our implementation solves many of the current issues, especially we eliminated the need for specialized annotation formats and we support summarized views for hundreds of samples. Since ggsashimi uses the most popular file formats and has very few dependencies, it can be easily integrated in any splicing analysis pipeline, and can facilitate the interrogation of alternative splicing in large-scale RNA sequencing projects, such as ENCODE [6] and GTEx [7]. ggsashimi is freely available at https://github.com/guigolab/ggsashimi. Further extensions of ggsashimi include incorporating spread metrics to accompany mean and median aggregating methods, allowing the user to select which type of reads to plot (e.g. uniquely mapped) or optionally showing only the aggregated coverage.

10.1371/journal.pcbi.1004185

Machine Learning Methods Enable Predictive Modeling of Antibody Feature:Function Relationships in RV144 Vaccinees

The adaptive immune response to vaccination or infection can lead to the production of specific antibodies to neutralize the pathogen or recruit innate immune effector cells for help. The non-neutralizing role of antibodies in stimulating effector cell responses may have been a key mechanism of the protection observed in the RV144 HIV vaccine trial. In an extensive investigation of a rich set of data collected from RV144 vaccine recipients, we here employ machine learning methods to identify and model associations between antibody features (IgG subclass and antigen specificity) and effector function activities (antibody dependent cellular phagocytosis, cellular cytotoxicity, and cytokine release). We demonstrate via cross-validation that classification and regression approaches can effectively use the antibody features to robustly predict qualitative and quantitative functional outcomes. This integration of antibody feature and function data within a machine learning framework provides a new, objective approach to discovering and assessing multivariate immune correlates.

Antibodies are one of the central mechanisms that the human immune system uses to eliminate infection: an antibody can recognize a pathogen or infected cell using its Fab region while recruiting additional immune cells through its Fc that help destroy the offender. This mechanism may have been key to the reduced risk of infection observed among some of the vaccine recipients in the RV144 HIV vaccine trial. In order to gain insights into the properties of antibodies that support recruitment of effective functional responses, we developed and applied a machine learning-based framework to find and model associations among properties of antibodies and corresponding functional responses in a large set of data collected from RV144 vaccine recipients. We characterized specific important relationships between antibody properties and functional responses, and demonstrated that models trained to encapsulate relationships in some subjects were able to robustly predict the quality of the functional responses of other subjects. The ability to understand and build predictive models of these relationships is of general interest to studies of the antibody response to vaccination and infection, and may ultimately lead to the development of vaccines that will better steer the immune system to produce antibodies with beneficial activities.

Antibodies provide the correlate of protection for most vaccines [1]. This correlation is often thought to be mechanistic, as in numerous disease settings passively transferred antibodies provide protection from infection [2]. Yet, the fact that some vaccines that induce an antibody response do not provide protection indicates that beyond presence and prevalence, there are specific antibody features associated with protection: that is, not all antibodies are created equal. Efforts to develop a protective HIV vaccine may represent the setting in which the discrepancy between the generation of a robust humoral immune response and generation of protective humoral immunity has been most apparent. That this might be a more general observation is suggested by recent dengue vaccine trials, where protection was seen but did not appear to correlate with the well-established virus neutralization assay [3,4]. The significant challenges to inducing antibodies with potent anti-HIV activity have been well described [5]. Due to viral diversity, vaccine-specific antibodies may or may not recognize circulating viral strains [6]. Furthermore, beyond viral recognition, binding antibodies vary considerably in their ability to neutralize diverse viral variants (case studies in [7,8] and reviewed in [9]), with most antibodies possessing weak and/or narrow neutralization activity [10]. While generating broadly neutralizing antibodies represents a cornerstone of HIV vaccine efforts, as these antibodies clearly block infection in animal models [11], vaccines tested thus far have induced antibodies with only a limited ability to neutralize viral infectivity [12]. However, beyond this role in the direct blockade of viral entry, antibodies mediate a remarkable repertoire of protective activities through their ability to recruit the antiviral activity of innate immune effector cells. Yet, here as well, the ability of HIV-specific antibodies to act as molecular beacons to clear virus or virus-infected cells is also widely divergent [13]. Given the diversity of viral variants, the diversity of antibody binding and neutralization profiles driven by the IgG variable (Fv) domain, and the diversity of antibody effector activity driven by the IgG constant (Fc) domain, the landscape of antibody activity is perplexingly complex. While a number of structure:function relationships have been characterized in terms of virus recognition, neutralization, and innate immune recruiting capacity, our understanding of the relationship between antibody features and their protective functions remains incomplete. However, the recent development of high-throughput methods to assess properties of both antigen recognition and innate immune recognition [14] offers more fine-grained information about the antibody response, which could feed into the development of models to inform our understanding of antibody activity. The moderate success of the RV144 HIV vaccine trial, in which partial protection from infection was observed [15], presents the opportunity to study antibody structure:function relationships in the first HIV vaccine to demonstrate efficacy. Importantly, within this trial, the correlates of reduced risk of infection were binding antibodies, and, in the absence of an IgA response, antibody function, in the form of natural killer (NK) cell-mediated antibody-dependent cellular cytotoxity [16]. Subsequent analysis has supported these findings: with evidence of the impact of variable domain-specific antibodies apparent in the sequences of breakthrough infections [17], and antibodies of the IgG3 subclass associated with reduced risk of infection [18]. Because the vaccine was partially efficacious, studying the diversity of antibody responses among volunteers has the potential to help identify novel immune correlates. Thus, this trial represents a compelling opportunity to profile antibody structure:function relationships from the standpoint of relevance to protection and an excellent setting in which to apply machine learning methods to characterize the relationship between antibody features and function in a population whose response to vaccination varied in a clinically relevant way. Here, we study the relationships between biophysical data regarding HIV-specific antibodies induced by the RV144 vaccine regimen, and corresponding functional properties that have previously been correlated with better clinical outcomes in HIV infected subjects [19–21] as well as the protection observed in RV144. These effector functions are mediated by the combined ability of an antibody’s Fab to interact with the antigen and its Fc to interact with a set of FcR expressed on innate immune cells. Just as Fab variation impacts antigen recognition, Fc variation in IgG subclass dramatically influences FcR recognition, and antibody effector function is widely divergent among antibodies from different subject groups in ways that are not explained by titer, or the magnitude of the humoral response [22]. Therefore, we characterize the combination of antigen specificity and subclass in a multiplexed fashion (“antibody features”), and couple that characterization with assessments of effector activities from cell-based assays (“antibody functions”). This antibody feature and function data have previously been subjected to univariate correlation analysis, which identified associations between gp120-specific IgG3-subclass antibodies and coordinated functional responses in RV144 subjects. Conversely IgG2- and IgG4-subclass antibodies were associated with decreased activity, and subsequent depletion studies confirmed these discoveries [23]. In order to discover and model multivariate antibody feature: function relationships in data from RV144 vaccinees, we employ a representative set of different machine learning methodologies, within a cross-validation setting that assesses their ability to make predictions for subjects not used in model development. While “predict” often connotes prospective evaluation, here, as is standard in statistical machine learning, it means only that models are trained with data for some subjects and are subsequently applied to other subjects in order to forecast unknown quantities from known quantities. In particular, we show that not only are antibody features correlated with effector functions, but that computational models trained on feature: function relationships for some subjects can make predictions regarding the functional activities of other subjects based on their antibody features. Using unsupervised methods we find patterns of relationships between antibody features and effector functions as well as among features themselves. Then, using classification methods we demonstrate via cross-validation that antibody features support robust qualitative predictions of high vs. low function, and using regression methods we likewise demonstrate that the features can enable quantitative predictions of functionality across multiple, divergent activities. The various methodologies are relatively consistent in both performance and identified features, giving confidence in the general procedure and the information content in the data. This objective approach to developing predictive models based on patterns of antibody features provides a powerful new way to uncover and utilize novel structure:function relationships. To model antibody feature-function relationships we analyzed samples from 100 subjects in the RV144 trial. A set of 3 different cell-based assays was conducted to characterize the functional activity of these samples, providing data regarding the effector function of antibodies induced by RV144 including: gp120-specific antibody dependent cellular phagocytosis (ADCP) by monocytes [24], antibody dependent cellular cytotoxicity (ADCC) by primary NK cells [25], and NK cell cytokine release (namely the combination of IFNγ, MIP-1β, and CD107a) [23]. Antibody features were assessed using a customized microsphere array [14] to characterize the antibodies induced by the vaccine in terms of their antigen specificity (gp140, gp120, V1V2, gp41, and p24) and IgG subclass (IgG1, IgG2, IgG3, and IgG4). For both the array-generated antibody feature data, and cell-based assay assessment of antibody functional activity, excellent discrimination between placebo (n = 20) and vaccinated (n = 80) subjects was observed [23]. The dataset is provided as a spreadsheet (S1 Dataset). Fig 1 illustrates scaled and centered data for each antibody feature (Fig 1A) and functional measurement (Fig 1B) for the 80 vaccinated subjects. We note that the subsequent analyses all use scaled and centered feature data, as the different features are on different and somewhat arbitrary scales according to bead set and detection reagent, and this standardization enables combination of the relative feature levels across these different scales. As a linear transformation, the standardization does not affect linear models, though the additional preprocessing truncation to 6σ has an appropriate impact on outliers. The function data are only standardized for this visualization, as the assay values are meaningful for interpreting predictions. As discussed in the introduction, the data and correlation analyses have been previously presented [23]; we recapitulate the most relevant points here to lead into our machine learning approaches. We observe that the antibody features and functions are far from uniform. The relative functional responses differ by subject and by function, though a number of subjects exhibit relatively strong or weak responses in multiple functions. Likewise, relative antibody feature strength differs by subject and feature, and notably some subjects exhibit relatively strong responses across multiple antigen specificities for a given IgG subclass and/or strong responses across multiple subclasses for a given antigen specificity. Finally, there are relationships between the features and functions by subject, e.g., a group of subjects with strong ADCP and ADCC responses appear also to have strong feature characteristics. In order to better extract, assess, and utilize such observations, machine learning techniques were applied to provide models of the relationship between characteristics of HIV-specific antibodies induced by vaccination, and their functional activity. As Fig 2A illustrates, assessing antibody feature:function correlations across subjects enables the identification of several strong relationships. Consistent with their binding affinity to FcgR expressed on monocytes, IgG1 and IgG3 subclasses are most correlated with strong ADCP function, while IgG2 and IgG4 are less correlated or even mildly anticorrelated. Similarly, gp120 and V1V2 antigens tend to yield the strongest correlations, as would be expected given the direct experimental relevance of these antigens to this functional activity. For ADCC, the IgG1 correlations are weaker and the IgG3 correlations weaker still, while the IgG2 and IgG4 classes are now slightly more correlated (particularly IgG2.gp41). For the cytokines, strong IgG1 and IgG3 correlations are observed, particularly with gp120 and V1V2. The IgG4 subclass also yields some strong correlations, likely influenced by the large number of subjects with undetectable IgG4 responses (uniform colors within a column in Fig 1, no longer 0 after standardization), and rare subjects with strong IgG4 responses. A number of antibody features exhibit similar patterns of correlation with function; these can largely be explained by correlations among the features themselves. Indeed, hierarchical clustering of the feature correlation profiles (Fig 2B) reveals that the features are not independent but in fact the true dimensionality of the data is lower than the number of original columns. The figure highlights six clusters of mutually correlated features formed by bisecting the dendrogram as indicated to strike a balance between the number of clusters and their visual coherence. An array of statistical methods to determine an optimal number of clusters gave substantially different answers from each other, though the optimal partitions they identified were largely consistent how one might manually divide the dendrogram (results not shown). Some of these clusters are defined by Ab subclass (each IgG subclass dominates one cluster), while others are defined by antigen specificity (V1V2 and p24 clusters are also observed). Correlations between IgG1 and IgG3-defined clusters are also observed. The combination of the feature:feature clustering and the feature:function correlations observed suggests that different groups of subjects produce characteristically different antibody responses, yielding different functional outcomes. The strong relationships apparent among antibody features (indicating lower intrinsic dimensionality) likely result in redundancy in terms of their contributions to functional predictions. To support the supervised analysis below, a set of “filtered” feature sets was developed for each function. Filtered features were selected by choosing the feature most strongly correlated with the function within each cluster, in terms of the magnitude of the Pearson correlation coefficient (Fig 2A). Filtered features for each functional measurement are starred in Fig 2B, and span the full range of subclasses and antigen specificities. Thus, while redundancy is reduced, the ability to obtain insights into the relative contributions of each feature type to functional activities is maintained. While there are non-negligible correlations outside the clusters (and indeed between these selected features), the supervised results show that they have little impact on predictive performance. As an alternative method to account for the possible redundancy among antibody features, a principal component analysis (PCA) was also performed. PCA yields a set of principal components (PCs) that represent the main patterns of variability of the antibody features across subjects. The PCs provide a new basis for the data; i.e., each observed feature profile is a weighted combination of the PC profiles, so we can think of the PCs as “eigen-antibodies”. In contrast to the filtered features, the principal components are composites, and by inspecting their composition, we can see the patterns of concerted variation of the underlying antibody features. Fig 2C illustrates the principal components and S1 Fig provides the corresponding eigenvalue spectrum (the relative amount of variance captured by each PC). While PC1 is essentially a constant offset by which to scale the overall magnitude of a feature profile, the other leading PCs reflect many of the same relationships also observed in the clustering analysis, including both subclass relationships and antigen specificity relationships. In particular, PC2 largely contrasts IgG2/4 vs. 1/3 composition, PC3 IgG4 vs. others, and PC4 IgG3 vs. others, while PC5 focuses on the relative p24-associated contribution, PC6 that of V1V2, and PC7 apparently an even finer-grained V1V2 specificity. As these leading seven principal components are the most readily interpretable and cover a large fraction of the variance in the data (S1 Fig), they are used for supervised learning below, and trailing PCs are dropped. The unsupervised analysis suggests that there is indeed a high level of information content in the data, evidenced by the relationships among features identified by the clustering and PCA approaches, the correlations between the antibody features and the functions, and the agreement of these relationships with biological intuition. The strong relationships uncovered by these methods suggest that it might be possible to build models to predict functions from features, whether directly measured features or derived composites. We first sought to robustly classify antibody function as high or low, relative to the median. To assess how much this discrimination depends on the classification approach utilized rather than the underlying information content in the data, we employed three different representative classification techniques: penalized logistic regression (a regularized generalized linear model based on Lasso), regularized random forest (a tree-based model), and support vector machine (a kernel-based model). Furthermore, in order to assess the effect of reducing redundancy and focusing on the most interpretable feature contributions, three different sets of input features were considered: the complete set (20 features: 4 subclasses * 5 antigens), the filtered set with one feature selected from each cluster based on correlation with function (6 features), and the PC features (7 leading PCs), as illustrated in Fig 2. Separate classifiers were built for each function and each input feature set. Fig 3 summarizes the classification results for ADCP by penalized logistic regression. To assess the overall performance, we conducted 200 replicates of five-fold cross-validation. That is, for each of 200 replicates, the subjects were randomly partitioned into five equal-size sets, or “folds”, and five different models were constructed. Each model was trained using data for four of the sets of subjects, and then was used to make predictions for the fifth “held-out” set. The predictions for the held-out subjects were compared against the known (but ignored for training) values, and performance assessed accordingly. By repeating this 200 times, the impact of the random split can be factored out. Fig 3A illustrates the predictions on one replicate (combining all five of its folds, with each serving separately as test data) and Fig 3B summarizes the resulting area-under-ROC-curve (AUC) over all 200 replicates (computing AUC only on test data). This data poses a difficult classification problem as there is not a clear distinction between high and low classes, which were simply defined by the median value. Nonetheless, even with a rigorous 200-replicate five-fold cross-validation, a mean AUC of 0.83 (standard deviation of 0.10) was observed, indicating that antibody features are highly and robustly predictive of high vs. low ADCP activity. Fig 3G shows the contributions of the antibody subclass-specificity features to a classifier trained on the whole dataset; while the coefficient values varied in individual folds, the same overall trends were observed over the different splits (results not shown). Penalized logistic regression readily enables assessment of the relative importance of different features for classification. The model sums the feature values, each weighted by its specific coefficient, and then applies a logistic function to yield the predicted classification value. In order to counteract overfitting, the training process imposes a penalty relative to feature coefficients and thereby seeks a sparse model. The coefficients give the relative importance of each feature to the predictor; associated p-values indicate the confidence in those coefficient values (a large p-value indicates an unreliable estimate of the feature contribution). Thus we see, for example, that the two dominant and statistically significant (at an unadjusted 0.05 level) contributors to predicting ADCP class are IgG1.gp120 and IgG3.p24, capturing both key subclasses with two different antigen specificities. While not achieving statistically significant confidence in the coefficient value, negative contributions from IgG2 were also observed, consistent with the unsupervised analysis and the reduced ability of this subclass to bind to FcγR on phagocytes presumably due to blocking (i.e., preferred binding of antibodies with better affinity). No systematic pattern was observed among the misclassified samples; they varied over the 200 splits and were distributed over the whole range of ADCP values. They did, however, tend to be those subjects with the weakest overall feature profiles, without large contributions from features with either positive or negative coefficients. Despite penalization, a relatively large number of features contributed to the classifier, and to some extent they appeared redundant given the correlations among features observed in unsupervised analysis. To obtain a sparser and less redundant model, we trained classifiers using the filtered features from Fig 2B. Despite the reduction in data considered, Fig 3C and 3D shows that the resulting performance with the filtered feature set is comparable to that with the complete feature set, with a mean AUC of 0.84 (standard deviation 0.10). The feature contributions in Fig 3H are still driven by positive contributions of IgG1 and IgG3 with some of the same antigens, along with negative IgG2 (with gp140). Though the goal of this study was not to comprehensively and rigorously assess feature selection methods, which would require further subsampling the data, we did investigate the sensitivity of the cluster-based filtering to our use of the features within each cluster that had the highest PCC. Thus we assessed each possible combination of features taken from the six clusters in Fig 2B. We found that on average an AUC of 0.79 was obtained, with a range from 0.67 to 0.87 and a standard deviation of 0.04 (recall that the PCC-based approach obtained 0.84). This result supports the conclusion that these groups of features do contain more or less redundant information in terms of predicting function. Using the best correlated features provides a sparse model that predicts as well as the model built from the complete feature set, and carries the advantage of being less likely to perform well due to overfitting, and thus more interpretable in terms of the underlying biology. As noted above, PCA provides an alternative means commonly used to reduce redundancy. Thus we also trained classifiers using the principal components as features. Using these alternative, composite features, performance quality was maintained (Fig 3E and 3F), with a mean AUC of 0.82 (standard deviation 0.11). Inspecting the key PCs contributing to a classifier, we see that PC2 (IgG2/4 vs. 1/3) makes the biggest contribution, modulated by subclass contributions in PC3 (IgG4) and PC4 (IgG3) and antigen contributions in PC5 (p24), and PC6 (V1V2) (Fig 3I). Thus the PCA-based approach is largely consistent with the others, with subclass and antigen specificity again working in concert to predict function. Table 1 summarizes the classification performance under all three classification methods. All three machine learning techniques perform quite well, despite the difficulty of the median-split classification problem and the rigorous five-fold cross-validation assessment. The PLR model is consistently a bit better, and performance is essentially equivalent for each technique across the different feature sets (complete, filtered, or PC), suggesting that over a wide range of different modeling approaches, antibody features are indeed robustly predictive of qualitative effector function. Corresponding classifiers were also built for ADCC and cytokine profiles using each of the three different learning techniques and three different feature sets; the performance of these models is also summarized in Table 1. The cytokine classifiers perform nearly as well as the ADCP ones, and the ADCC classifiers less accurately but still strikingly well. The choice of feature set (complete, filtered, PC) did not have a substantial effect on performance. The PLR approach was generally superior, with RRF quite comparable and SVM somewhat degraded but still yielding good performance. Thus our hypothesis that antibody features enable robust, high-quality prediction of antibody function is well-supported by the summary results for each of three distinct effector functions. Furthermore, the logistic regression model enables straightforward identification of the key contributors, and points toward feature roles consistent with known IgG and innate immune cell biology. S2 Fig (ADCC) and S3 Fig (cytokines) detail the PLR results. For ADCC, the key contribution using the complete feature set is made by IgG1.gp41, consistent with ADCP in terms of subclass, but driven by a different antigen. In contrast there appears to be less contribution from IgG3 and IgG4 contributes positively (though the confidence in that coefficient is lower). Several of the selected features are gp41-specific. These trends are also largely reflected in the unsupervised feature:function correlations in Fig 2A. The cytokine feature usage is driven by IgG1 and IgG3 (with different antigens), along with an inconsistent contribution from IgG4, negative with p24 and gp140 and positive with gp41. Since these features are themselves highly correlated (Fig 2C), it appears that, despite the penalization in the PLR approach, this model is likely to be overfit. For both functions, feature filtering results in much the same relative contributions as for the complete feature set, with coefficients more strongly focused on a few key features. Notably, the inconsistent use of IgG4 features is eliminated by filtering. The ADCC response for the PC features is driven by PC6, which appears primarily to distinguish the V1V2-specificity. The PC features selected for the cytokines are more consistent with the other feature sets, with PC2 (IgG2/4 vs. 1/3) modulated by PC6 (V1V2), along with an IgG4.V1V2 down-selection via PC7. The median-based dichotomization into high and low classes allowed us to characterize which antibody features were generally associated with superior effector function, but the division between high and low was quite fuzzy, with many subjects on the border. Thus we also performed classification into the top and bottom quartiles (ignoring the middle half). While unsurprisingly, the best vs. worst classification performance was better than the better vs. worse, our focus was the features driving class assignment, which remained largely consistent (results not shown). In particular, IgG1, with a variety of antigens, was the dominating contributor, often complemented by an IgG3-based feature; in addition, IgG4 features contributed negatively to ADCP but positively to the other two functions. Given the quality of the classification results, both in predictive ability and in terms of clear and consistent use of biologically significant features, we sought to build quantitative models to predict function. Again, three representative techniques were used to broadly assess the general ability of the data to support predictive models: Lars (regularized linear regression based on Lasso), Gaussian process regression (a nonlinear model), and support vector regression (a kernel-based model). We again built separate models for each function, under each set of input features. Fig 4 summarizes the ADCP regression results from Lars across the complete feature set (Fig 4A, 4B and 4G), the filtered features (Fig 4C, 4D and 4H), and PCs (Fig 4E, 4F and 4I). While 200-replicate five-fold cross-validation was used for performance assessment, leave-one-out cross-validation (LOOCV) was used to generate representative scatterplots of experimental vs. predicted functional values, as is appropriate when viewing LOOCV as a form of jackknife. The models are clearly predictive of ADCP, obtaining a mean Pearson correlation coefficient PCC = 0.64 (standard deviation 0.15) over the 200-replicate five-fold. An example LOOCV scatterplot is illustrated in Fig 4A; the correlated trend between observed and predicted ADCP is clear. Notably, the LOOCV and five-fold PCCs (Fig 4B) were similar. As a form of linear regression, Lars enables direct inspection of the coefficients contributing to the prediction. As with penalized logistic regression, the regularization employed by Lars in training seeks to force coefficients to zero and yield a sparse model. Fig 4G depicts the coefficients and their p-values for a model trained on the entire set of features. Among the largest and most-confident coefficients, we see that IgG1.gp120 is again a strong positive contributor, joined by the related IgG1.gp41 and IgG3.p24, and IgG2.gp140 is a strong negative contributor. Despite the Lars penalization, the model incorporates offsetting positive and negative contributions from IgG4 under different antigens, though these features are highly correlated with each other (Fig 2C). In inspecting outliers, we found that the most overpredicted subjects (i.e., predicted ADCP much larger than experimental) were characterized by a relatively large number of features with large values. A possible statistical explanation for this is that the model works best when a few features are indicative of the response. A possible experimental explanation is that there are competitive effects, and indeed the contributions from multiple good antibodies are not additive in terms of recruiting effector cells. As with classification, we sought to focus on the most informative and non-redundant features in order to reduce the risk of overfitting and develop more readily interpretable models. Models learned from the filtered features from Fig 2B maintain about the same accuracy (mean PCC = 0.61 with standard deviation 0.15 for the 200-replicate five-fold (Fig 4D); an example LOOCV scatterplot is illustrated in Fig 4C). By inspecting features for a model trained on the filtered features (Fig 4H), we see that the prediction is driven primarily by IgG1.gp120 and IgG3.V1V2, with a negative contribution from IgG2.gp140. The contradictory IgG4 contribution is resolved. Similarly, PCA-based models attain mean PCC of 0.61 with standard deviation 0.15 (Fig 4E and 4F), based largely on PC2 (IgG2/4 vs. 1/3) and somewhat on PC3 (IgG4 vs. others), as can be seen in Fig 4I. The performance of all three machine learning methods using all three feature sets is summarized in Table 1. As with classification, the linear model dominates, and all methods perform similarly well with any of the input feature sets. Lars-based regression results for ADCC and cytokines are presented in S4 Fig, and S5 Fig, respectively, and summarized in Table 1. While providing the desired trend overall (with a few striking outliers), the ADCC regression with the complete feature set does not have as high a PCC (mean 0.40, standard deviation 0.18) as the ADCP one (mean 0.64, standard deviation 0.15). With a mean PCC of 0.58 and a standard deviation of 0.20, the cytokine regression is comparable to that observed in predicting ADCP, though the representative scatterplot is not as pleasing to the eye due to the density of subjects with low values. Feature filtering achieves essentially the same performance for ADCC but a degradation in the cytokine performance as assessed by PCC, though the scatterplot appears roughly as good. The switch to PC features degrades the PCC measurements for both functions, though again yielding trends that appear satisfactory visually. As for classification, ADCC prediction is driven by IgG1.gp41, with IgG1.gp140 also contributing strongly, and probably redundantly, as suggested by Fig 2B. As we saw for classification, the cytokine model has positive IgG1 and IgG3 contributions and inconsistent IgG4 contributions. For the filtered features, the ADCC model is focused on IgG1.gp41, with IgG1.gp140 replaced by the related IgG3.gp140. The feature-filtered model for cytokines retains IgG3.V1V2 and IgG1.gp120 contributions and resolves the IgG4 inconsistency, leaving a positive IgG4.gp41 contribution as observed in Fig 2A. When switching to the PCA-derived features, the ADCC regression model is driven by PC6 (V1V2), as with the classification model, while the cytokine regression model agrees with the classification model in its use of PC6 and PC7 with opposing signs, while weakening PC2 (IgG2/4 vs. 1/3) perhaps in lieu of added contributions from PC4 (IgG4) and PC3 (IgG3). Table 1 summarizes the performance for ADCC and cytokines under all machine learning techniques and feature sets. Once again the linear model dominates the nonlinear models, particularly for ADCC. With the complete feature set, this is likely directly attributable to overfitting, and an improvement of the nonlinear methods upon starting with the filtered features though not as much with the PC features, was observed. As discussed in the methods, the presented results employ a polynomial kernel for Gaussian Process Regression and a radial basis kernel for Support Vector Regression; alternative kernels did not improve the performance. While the disappointing performance of the more sophisticated methods could potentially be improved by custom feature selection methods or parameter tuning, our goal here is not to provide such a benchmark but rather to establish the general scheme of predictive modeling of antibody feature: function relationships. The overall concordance observed between different feature sets, different regression and classification methods, and across multiple, complex, antibody functional activities, subjected to cross-validation assessment, demonstrates that indeed antibody features can be used to effectively predict functional activities. We have demonstrated that the integration of antibody feature and function data via machine learning models and methods helps identify and make use of critical landmarks in the complex landscape of antibody feature:function activity. Sets of features emerge from patterns in the data, and these feature sets are able to robustly predict high/low levels of function, and are even informative enough to support quantitative predictions of functional activity. The subclass-specific contributions observed here are consistent with expectations, according to the receptors on the relevant effector cells, and the activity profiles among IgG subclasses [26]. At the same time, the approach provides a finer resolution picture of the interrelationships among antigen specificity, subclass, and effector function. In the case of RV144, it is worth noting that the vaccine included two different components, priming with canarypox ALVAC-HIV (vCP1521) and boosting with recombinant gp120 AIDSVAX B/E protein. Thus while the prime included the gp120, gp41, and p24 antigens evaluated here, the boost only included gp120. Furthermore, cell-based functional assays employed particular antigens to stimulate a response, and those studied here are gp120-specific. Thus we might expect to see differences within functional responses among subjects according to different overall specificities of their antibodies, or even within antibody specificities depending on whether they were raised in the setting of the prime or the boost. Accordingly, associations observed here, such as those between gp41-specific antibodies and functional activity in assays in which only gp120 is presented, clearly do not have mechanistic significance with respect to functional assays that characterize only gp120-specific responses. However, they may nonetheless provide useful associative markers that functionally differentiate overall antibody responses to priming and boosting or among subjects that were more finely grained than subclass and antigen-specificity alone. The machine learning approaches employed here contrast with typical univariate correlation analysis in two important ways: simultaneously combining and down-selecting features, and assessing generalization performance in a predictive setting. These approaches incorporate multiple features into a model, but do so in a way that avoids simply “memorizing” artifacts of the samples, as is easily possible with a sufficient number of features for a small sample set. Cross-validation analysis then ensures that the models are not overfit, by testing how well predictions from a model trained on one set of data match observations for another set. This predictive assessment stands in contrast to typical correlation analysis, which uses all the data and simply evaluates quality of fit. Redundancy among features confounds the interpretation of multivariate feature:function relationships. To account for redundancy, we have used representative, common approaches including feature selection within the learning algorithm (via regularization), feature filtering (via feature clustering), and feature combination (via principal components analysis). The approaches were all fairly comparable in performance for this dataset, perhaps due to the relatively small number of initial features. Larger feature sets may result in more substantial differences, and require additional techniques to reduce the number of features contributing to a model down from a highly redundant input set to a reduced but representative and robust set. For example, elastic net type approaches [27] might strike a beneficial balance between eliminating redundant features and averaging them out to improve robustness. The goal of this paper is to demonstrate that it is possible to develop models able to robustly predict the broad functional activities of antibodies from data regarding antigen specificity and Fc characteristics, with an aim ultimately in developing models that will correlate with protection or risk of infection. Several representative methods were demonstrated, though a rigorous benchmarking comparison was not performed as that would require a larger, more diverse dataset. We conclude that while there are some clear differences in performance among the methods, they all show that there is sufficient information in the features to predictively model function. The penalized generalized linear models are generally very good, and provide the added advantage of easy interpretation and relatively low model complexity; as noted in the previous paragraph, a softer regularization might be beneficial in the future. The relationships identified by machine learning methods can be used to drive prospective studies to test particular hypotheses regarding how particular antigen specificities and subclasses contribute to the stimulation of effector response. As an illustration, we note that subsequent to our modeling and characterization of feature:function relationships in the RV144 data, depletion studies confirmed a mechanistic role for antibodies associated with prediction quality. These experimental observations demonstrated that indeed IgG3 is important for a strong phagocytic response, with IgG3-depleted samples having significantly reduced ADCP activity [23]. Similarly, our models predicted that IgG4 has a negative impact on functional level, and an analogous depletion experiment did exhibit this trend across 2 different vaccine regimens, although the increase in activity in the RV144 samples when IgG4 was depleted did not meet statistical significance [23]. Due to the evident importance of innate immune recruiting for the protection observed in the RV144 trial, and given the unprecedented feature and function data available for a set of subjects from that trial, we have focused here on specific relationships within the repertoire of antibodies induced by this vaccine. However, the approach described here can also be productively applied in other settings, shedding light on relationships specific to particular cohorts, as well as different vaccination and infection contexts. By integrating diverse datasets, it may even be possible to uncover more general rules governing the ways that antibodies bridge the adaptive and innate arms, and how those rules can then be specialized in a context-dependent fashion. While the present study demonstrated the ability of antibody features to predict functional activities, the longer-term goal is to predict the impact of vaccination. To this end, an important next step is a case/control study with the potential to tease apart signatures leading to protection. Even in the context of the functions assayed here, a more complex multi-output model could be built in order to ascertain signatures of desirable polyfunctional responses. The fact that some functions were better predicted than others in the models described here, may indicate that additional antibody feature information could contribute to improved model performance. In particular, ADCC activity, the function predicted most poorly by the antigen and subclass data used here, is known to be dependent on antibody glycosylation state [22], which was not assessed in this study. Feature data could be extended to characterize a wider range of relevant antibody features, including additional antigen specificities as well as characteristics of the Fc glycan structure, or interactions with the cellular antibody receptors expressed by NK cells and phagocytes. Overall, we find that the parallel assessment of antibody function and antibody features can provide for development of models enabling quantitative predictions of functional activity across multiple, divergent antibody activities. Because these antibody functions have been associated with better clinical outcomes in HIV infected subjects, as well as the protection observed in RV144 and in many settings beyond HIV infection, but are poorly predicted by antibody titer, we anticipate that this type of predictive model can provide significant value, both in terms of permitting the substitution of high-throughput biophysical characterization for low-throughput cell-based assays, as well as for uncovering novel structure:function relationships that can inform vaccine design efforts. Plasma samples, provided by the MHRP and RV144 study group, were obtained from 100 participants in the RV144 vaccine trial [15], consisting of 20 placebo and 80 vaccinated subjects at week 26. Experimental methods used have been previously described [23]. Briefly, IgG was purified from all samples using Melon Gel according to the manufacturer’s instructions (Thermo Scientific). The functional activity of HIV-specific antibodies was determined in 3 different cell-based assays. Phagocytic activity was assessed using a monocyte-based assay in which the uptake of gp120-coated fluorescent beads is determined by flow cytometry [24]. Antibodies were tested at a concentration of 25 ug/ml MN. Similarly, the cytotoxicity profile of antibodies was tested at a concentration of 100 ug/ml in the rapid fluorescent ADCC assay, which assesses the ability of antibodies to drive primary NK cells to lyse gp120-pulsed target cells [25]. Lastly, NK cell degranulation and cytokine secretion were monitored by flow cytometry as described [23]. Surface expression of CD107a, and intracellular production of IFN-γ and MIP-1β were assessed, and the fraction of NK cells which were triple positive was determined. In order to profile antibody features, a customized antigen microsphere array was used to assess antibody specificity (gp120, gp140, V1V2, gp41, and p24) and subclass (IgG1,2,3,4) [14]. Array measurements for the vaccinees were standardized individually for each antigen.subclass feature as follows. Background signal level was derived from the values for that feature among placebos, as the placebo mean plus one standard deviation. This background was subtracted from each vaccinee. Finally, the vaccinee values for the feature were scaled and centered to a mean of 0 and a standard deviation of 1, with values truncated to 6σ. For functional assays, data was not placebo-subtracted, but was instead inspected to ensure that low activity was observed in samples from placebo subjects Antibody feature:function and feature:feature correlations were computed over the set of 80 vaccinated subjects and assessed using Pearson correlation coefficient and p-value. Features were clustered based on the profile of their correlation coefficients over the set of all features. Hierarchical clusters were generated by the Ward linkage algorithm [28], assessing pairwise similarity between profiles in terms of Pearson correlation coefficient (i.e., 1-r dissimilarity). By visual inspection, six groups were identified in the resulting dendrogram. The R package NbClust was also used to assess optimal numbers of clusters according to a number of different indices [29]. For each function and each group, the feature with the largest-magnitude feature:function correlation coefficient was identified; each such feature also had the best feature:function p-value within its group, < = 0.001. Principal component analysis was performed on the feature:subject data matrix (after preprocessing). Singular value decomposition was employed to determine a set of eigenvectors and corresponding eigenvalues, with the eigenvectors serving as a basis transformation matrix containing principal components that are linear combinations of the original features, and the eigenvalues indicating the amount of variance in the data captured by their eigenvectors. The top 7 were chosen for further use in supervised methods, by visual inspection of their components and their eigenvalues. Three different and representative classification methods were employed: L1 penalized logistic regression (PLR) [30], regularized random forest (RRF) [31], and support vector machine (SVM) [32,33]. PLR is a form of logistic regression incorporating into the model evaluation a lasso penalty term λ||β||1, where λ is a tuning parameter and ||β||1 is the L1 norm of a coefficient parameter vector,β. Thus the learning favors sparse models, as zero-valued coefficients do not contribute to the penalty term. The R package “penalized” was used for PLR. It employs a greedy search to determine the best value for λ according to nested cross-validation (i.e., given a training set, doing an internal cross-validation within it to determine the performance under possible λ choices). RRF is a decision tree-based method that generates multiple decision trees over bootstrap replicates of the data (i.e., a random forest), at each split selecting a feature from a randomly-sampled set based on an Gini index assessment of node impurity augmented with a regularization penalty to prefer a sparser set of selected features. The R package “RRF” was used for RRF-based learning. Two parameters were specified: mtry, the number of features to be randomly sampled at each split, which was set to the number of input features; and ntree, the number of trees or bootstrap samples, which was set to 2000 to obtain more reliable results. The regularization parameter is handled automatically by the method, based on the scores from a 0-penalty model. SVM is a kernel-based nonlinear classifier that finds a separating hyperplane (in a space defined by the kernel) between the classes, so as to minimize the risk of classification error. The R package “e1071”, based on the C classification method of the libsvm library [34], was used for SVM-based classification. The standard linear, polynomial, and radial basis kernels were evaluated, and results presented for the radial basis function. Default parameter values were used except where noted. Each method was trained separately for each function with each of three different feature sets: the complete preprocessed set, the filtered set from the feature:feature clustering, and the set of principal components. To study the impact of selecting different features in the cluster-based filtering, the Lars method was also applied to each possible set of features combining one from each cluster. To obtain robust characterization of classification performance, 200 replicate five-fold cross validation was employed; i.e., the data was randomly split into fifths, four used for training and one for testing, with 200 different such training/testing runs. The R package “ROCR” was used to calculate a cut-off independent evaluation of the area under the ROC curve (AUC) for each replicate. To gain insights into the features driving the PLR classification performance, a model was also built using all subjects in order to obtain the best confidence in the coefficients. In order to evaluate the impact (both prediction quality and feature usage) of median-based dichotomization, the PLR-based approach was applied in the same manner to a dataset limited to the subjects with the top and bottom quartile ADCP values. Diverse representative approaches employed for regression were Lars [35,36], Gaussian Process Regression (GP) [37], and Support Vector Regression (SVR) [38]. Lars performs penalized linear regression with the L1-norm lasso penalty discussed above for PLR. The R package “parcor” was used for Lars. As with PLR the penalty weight was selected by cross-validation. The parameter for the number of splits was set to 10 for robust fitting. GP performs nonlinear regression based on a stochastic process specified in terms of mean and covariance functions. Observed values are used to fit the functions and thereby predict unobserved ones. The R package “kernlab” was used for GP. A polynomial kernel function was used to fit the GP model, as it performed better than other kernels. SVR is based on the same theory as SVM, discussed above, but uses the kernel-based approach to fit a regression model to reduce the quantitative prediction error. The R package “kernlab” was also used for SVR. As with SVM, we evaluated the standard linear, polynomial, and radial basis kernels and presented the results for the radial basis function. Default parameter values were used except where noted. The different feature sets were tested as described in the classification section. Performance was assessed by Pearson correlation coefficient (PCC), r, between observed and predicted function value; r assesses the linear correlation (between -1 for perfectly anticorrelated and +1 for perfectly correlated), while r2 represents the fraction of the variation explained. The PCC was computed over 200-replicate five-fold cross-validation. In addition, leave-one-out cross-validation was performed in order to generate representative scatterplots. A Lars model was trained on all subjects in order to enable inspection of feature coefficients.

10.1371/journal.ppat.1002213

Significant Effects of Antiretroviral Therapy on Global Gene Expression in Brain Tissues of Patients with HIV-1-Associated Neurocognitive Disorders

Antiretroviral therapy (ART) has reduced morbidity and mortality in HIV-1 infection; however HIV-1-associated neurocognitive disorders (HAND) persist despite treatment. The reasons for the limited efficacy of ART in the brain are unknown. Here we used functional genomics to determine ART effectiveness in the brain and to identify molecular signatures of HAND under ART. We performed genome-wide microarray analysis using Affymetrix U133 Plus 2.0 Arrays, real-time PCR, and immunohistochemistry in brain tissues from seven treated and eight untreated HAND patients and six uninfected controls. We also determined brain virus burdens by real-time PCR. Treated and untreated HAND brains had distinct gene expression profiles with ART transcriptomes clustering with HIV-1-negative controls. The molecular disease profile of untreated HAND showed dysregulated expression of 1470 genes at p<0.05, with activation of antiviral and immune responses and suppression of synaptic transmission and neurogenesis. The overall brain transcriptome changes in these patients were independent of histological manifestation of HIV-1 encephalitis and brain virus burdens. Depending on treatment compliance, brain transcriptomes from patients on ART had 83% to 93% fewer dysregulated genes and significantly lower dysregulation of biological pathways compared to untreated patients, with particular improvement indicated for nervous system functions. However a core of about 100 genes remained similarly dysregulated in both treated and untreated patient brain tissues. These genes participate in adaptive immune responses, and in interferon, cell cycle, and myelin pathways. Fluctuations of cellular gene expression in the brain correlated in Pearson's formula analysis with plasma but not brain virus burden. Our results define for the first time an aberrant genome-wide brain transcriptome of untreated HAND and they suggest that antiretroviral treatment can be broadly effective in reducing pathophysiological changes in the brain associated with HAND. Aberrantly expressed transcripts common to untreated and treated HAND may contribute to neurocognitive changes defying ART.

HAND is a common complication of HIV-1 infection in the nervous system presenting a varied spectrum of clinical manifestations with cognitive, motor and behavioral symptoms. Introduction of ART has greatly reduced morbidity and mortality in HIV-1 infection; however HAND persists and its overall prevalence appears to have increased despite treatment. The effects of the treatment on neurological disease are not well understood. Here, we used genomic analysis to compare gene expression profiles in brain tissues from treated and untreated patients who died with HAND. We identified a large number of genes and biological pathways dysregulated in untreated HAND compared with uninfected controls. ART appears to be effective in mitigating aberrant gene expression in brain tissues of patients with HAND but a fraction of genes remained dysregulated under ART and the patients continued to manifest HAND in the last evaluation prior to death. Our study provides new insights into the molecular changes in brain tissues of patients with HAND and the effect of the treatment on brain transcriptome. The identification of aberrantly expressed genes common to untreated and treated HAND may contribute to understand the neurocognitive impairment observed in patients under ART.

HAND is a common complication of HIV-1 infection in the nervous system presenting a varied spectrum of clinical manifestations with cognitive, motor and behavioral symptoms. Currently three conditions of increasing severity are recognized as components of HAND: HIV-1-associated asymptomatic neurocognitive impairment, HIV-1-associated mild neurocognitive disorders (MND), and HIV-1-associated dementia (HIV-D or HAD) [1]. HAND remains prevalent in populations with access to highly active ART, despite the efficacy of these therapies in controlling viral load and ameliorating viral load-associated clinical and neuroradiologic abnormalities [2]–[6]. By some accounts, up to 50% of HIV-1-infected individuals will develop some form of HAND regardless of access to currently available ART [7]–[9]. Under ART, HAND has became milder, its course more protracted and variable in symptoms, and it now overlaps with aging processes and potentially with other neurodegenerative diseases in AIDS patients [7], [10], [11]. The etiology of persistent cognitive deficits in patients on ART remains unclear. Studies of cerebrospinal fluid from individuals with HAND have sometimes supported contradictory conclusions; in the absence of viral replication, some authors observe associations among dementia, abnormal neurometabolites, and inflammatory phenotypes; others, with non-inflammatory states or markers of neurodegeneration [12]–[16]. There have been limited studies of brain tissues from treated patients with HAND to evaluate what biologic pathways remain abnormally regulated under ART, with studies largely focused on single cell types or molecules [17]. Comprehensive analysis of the spectrum of molecular abnormalities in the brain that may underlie HAND in the presence of ART has not yet been undertaken. Here we used functional genomics to conduct comparative analysis of genome-wide gene expression profiles in brain tissues from treated and untreated patients who died with HAND. Functional genomics has been applied with success to identification of complex molecular pathways in carcinogenesis and to better detection, classification, and prognosis of some cancers [18]–[23]. This approach has also been used extensively to investigate the transcriptome correlates of HIV-1 infection in peripheral tissues from HIV-1-infected patients including lymph nodes [24]–[26], CD4+ T cells [27], [28], monocytes [29], B cells [30], and gastrointestinal mucosa [31]. Some of these studies determined the effects of ART on HIV-1 transcriptomes in patients, revealing categories of treatment-responsive genes as well as aberrantly expressed transcripts that may serve as targets for future therapies [24], [27]–[31]. In a related approach, a recent study evaluated gene expression profiles of blood monocytes as a function of ART and neuropsychological impairment of HIV-1-infected patients [32]. Interestingly in this case, there was no correlation between changes in blood monocyte transcriptomes under treatment and clinical HAND [32]. Generally, this research benefited from the ability to serially sample peripheral tissues and individual cell types in living individuals, allowing ongoing evaluation of treatment. In contrast, analyses of human brain tissues by functional genomics can only be conducted retrospectively in autopsy tissues and are complicated by the multicellular interactions underlying central nervous system diseases [33]. Nonetheless, large-scale, cross-sectional gene expression profiling of brain tissues has revealed potential pathogenic pathways in Alzheimer's disease (AD) [34], [35], Parkinson's disease [36], [37], chronic schizophrenia [38], multiple sclerosis [39], and viral encephalitis [40]. With respect to HIV-1 infection in the brain, investigators reported transcriptional changes in selected gene categories such as anion channels in the frontal cortex of patients who died with HAND [41], [42]. Another group reported aberrant expression of genes specific to HIV-1 encephalitis (HIVE) [43], established a correlation between use of methamphetamine and up-regulation of interferon genes in these patients [44], and investigated the role of microRNA in gene regulation in HIV-1-infected brain [45]. To our knowledge, these studies did not consider the effects of ART on brain gene dysregulation in HAND. The Manhattan HIV Brain Bank (MHBB; member of the National NeuroAIDS Tissue Consortium) follows a cohort of advanced-stage, HIV-1-infected individuals with a high prevalence of well-characterized cognitive dysfunction; with entry to the study, participants agree to be organ donors upon death. The antiviral treatment status of study participants is monitored while in the program. Thus, brain tissues obtained by this program provide an opportunity to examine the potentially diverse processes underlying HAND in the ART era. We report herein the gene expression profiles of individuals with HAND focusing on the impact of ART on these profiles. We assayed virus burden and cellular gene expression on archived brain tissues from 15 HIV-1-infected patients with HAND and six HIV-1-negative subjects with no neurological or neuropathological abnormalities. Information for this study group is summarized in Table 1 and Methods. All assays were performed on parallel samples from deep white matter within the anterior frontal lobe, an area implicated in HAND and HIV-1-associated neuropathologies [46], [47]. Custom consensus primers based on sequences of HIV-1 amplified from the brain were used for reliable measurement of HIV-1 in the brain by quantitative real-time PCR (QPCR) (see Methods); the results were confirmed by standard PCR and hybridization with a specific probe (Table 1 and Supplementary Figure S1). Pre-mortem plasma HIV-1 burdens are plotted for comparison (Supplementary Figure S1). With the exception of patient 30015, the HIV-1 brain burdens in untreated patients correlated with presentation of HIVE; on average, these patients had about 180-fold more viral RNA per µg total RNA than patients without HIVE. Patient 30015 had limited HIVE pathology and no detectable virus in the brain (Table 1). Patients on ART, both with and without HIVE, had lower virus burdens in the brain than untreated patients, results consistent with a previous study in a different cohort [48]. In patients with HIVE the reduction was 50% and 95% at the DNA and RNA levels respectively, while virus was undetectable in treated patients without HIVE. Notably, there was no correlation between brain and plasma viral loads and brain virus burdens were independent of patient's age, gender, ethnic background, or postmortem interval. Global gene expression profiles of patient brain tissues were determined on Affymetrix GeneChip Array Human Genome U133 Plus 2.0 Arrays. Three independent array experiments were performed, each comprising a subset of HIV-1-positive samples and controls, with most samples tested either in duplicates or repeated in independent runs and some samples tested three times. Replicate gene sets of the same samples were averaged, yielding 21 final brain tissue datasets for 21 subjects. Preliminary hierarchical cluster analysis of complete datasets from HAND patients indicated that the primary biological variable in clustering of these datasets was whether or not patients were on ART at the time of death (not shown). Interestingly, this analysis also indicated that brain transcriptomes from untreated patients with HIVE did not cluster independently of non-HIVE transcriptomes (Figure 1A), suggesting that they are statistically similar across their entire datasets. This result was surprising because untreated patients with HIVE had on average higher brain virus burdens than patients without HIVE (Table 1 and Supplementary Figure S1) and HIV-1 infection is known to induce cellular gene expression [49]. To confirm findings of cluster analysis we performed global gene set analysis using GAzer software [50] to identify biological pathways that were most significantly altered in HIVE positive and negative groups compared to uninfected controls (Figure 1B). GAzer employs parametric analysis of sets of co-regulated genes across complete microarray datasets, independent of arbitrary fold change (FC) value limits, thus increasing statistical power of detection of biological differences between datasets [50], [51]. Figure 1B depicts the eight most altered pathways in HAND and HAND/HIVE datasets versus controls and Supplementary Table S1 lists all aberrant pathways and detailed statistics of GAzer analysis including Z-score, p-and q-values, and Bonferroni correction value. Consistent with previous array studies in HIVE patients [43], [44], datasets of patients with HIVE showed greater dysregulation of immune responses and endogenous antigen presentation pathways than those without HIVE (Figure 1B), possibly reflecting effects of high virus burdens in the brain in untreated HIVE (Table 1). However, the two HAND patient groups were generally similar with respect to the ranking and extent of change (indicated by all four statistical measures) of the majority of altered gene sets (Figure 1B and Supplementary Table S1). Because of their overall similarities in both cluster and gene set analyses, we chose to pool microarray datasets from HAND patients with and without encephalitis for analysis of antiviral drug effects. Seven of the 15 patients in our cohort were treated using different antiretroviral drug combinations (Table 1) enabling investigation of the extent of HIV-1-induced changes in the brain and potential differences between treated and untreated patients at the level of brain transcriptomes. Patient microarray datasets were pooled separately into untreated (HAND, n = 8) and treated (ART, n = 7) groups and each group was compared to the uninfected control pool C (n = 6). We used a cut-off of 1.5 FC and p-value<0.05 to identify significantly dysregulated genes. For some analyses, we also established a subgroup ARTa excluding low adherence subjects ART3 and ART5. The complete lists of FC values and accompanying statistics for HAND, ART, and ARTa are provided in Supplementary Tables S2 and S3. Overall, we identified 2073 dysregulated transcripts in the untreated HAND group and 333 and 145 transcripts in the ART and ARTa groups, respectively (Table S2). The Venn diagram in Figure 2A depicts the number of genes dysregulated in brain tissues of untreated and treated patients with HAND and the overlap in the dysregulated genes among the three groups tested. Excluding multiple probes for the same gene and transcripts of undefined function at the time of this writing, the HAND group had 1470 genes with significantly altered expression compared to 260 in ART and 107 in ARTa (Figure 2A, Supplementary Tables S2 and S3). About two-thirds of dysregulated genes (947) in untreated patients were up-regulated and 95 of these had FC values of ≥3.0 and p = 10−2–10−7, among down-regulated genes, 58 had FC of ≤−3.0 and p of 10−2–10−4, suggesting that HIV-1 infection profoundly alters the brain transcriptome in untreated patients with HAND and indicating significant conformity of molecular profiles of disease in this cohort. In treated patients, down-modulated genes predominated, the FC values ranged from 3.95 to 1.5 for up-regulated genes and from −1.5 to −2.87 for down-regulated genes, and p-values were 5×10−2–7.3×10−5 (Supplementary Tables S2 and S3), indicating less uniformity of brain gene expression in this group. These results indicate marked differences between aberrant brain transcriptome profiles of untreated and treated patients, the latter showing 6–14-fold (depending on treatment compliance) fewer dysregulated genes with generally lesser dysregulation of expression than in untreated patients. To confirm this observation we conducted gene ontology analysis using GAzer to examine cellular processes affected by HIV-1 infection in the brain in our patient groups. Altered gene sets (biological pathways) were identified by comparing each patient group to HIV-1-negative controls; Supplementary Table S4 lists these pathways for HAND, ART, and ARTa in the order of their significance as determined by the Z-score, p and q values, and Bonferroni statistics [50], [51]. Figure 2B shows the eight most dysregulated biological pathways in the HAND datasets, displaying the extent of dysregulation in the same pathways in ART and ARTa datasets. Up-regulated pathways in untreated HAND patients included immune responses, inflammation, response to virus, and complement activation while synaptic transmission, neurogenesis, ion transport, cell adhesion, and signal transduction were down-regulated. The statistical significance for the eight major pathway changes reached Z-scores of 6–14 and p, q and Bonferroni values of ≥10−8 (Figure 2B and Supplementary Table S4). In contrast, samples from treated patients showed either fewer significantly altered pathways (ART and ARTa panels in Figure 2B) or lesser extent of dysregulation of the remaining pathways displayed (e.g., the ART panel in Figure 2B). The extent and kind of changes in gene ontology processes in the treatment compliant ARTa group were limited compared to changes seen in untreated patients: of the 8 most up-regulated HAND pathways only endogenous antigen presentation and processing were also up-regulated in ARTa with Z-scores less than 2.5, low significance compared to other pathways (p = 0.018 and 0.026, respectively), and no down-regulated gene sets in these categories were identified (Figure 2B and Supplementary Table S4). Notably, some pathways that were significantly down-regulated in HAND were significantly up-regulated in ARTa including neurotransmitter secretion (Z = 4.26; p = 2×10−5) and synaptic transmission (Z = 2.9; p = 0.0037) (Supplementary Table S4). Dysregulation of selected genes detected by microarrays was confirmed in adjacent tissues by QPCR. Genes were chosen by previous demonstration of their link to HAND [52]–[56]. Representative gene expression values are shown in Figure 2C and the complete list is provided in Supplementary Table S5. Consistent with microarray data, brain tissues from untreated patients showed significant up-regulation of complement component 3 (C3), macrophage antigen CD68 (CD68), and protein-tyrosine phosphatase receptor-type C PTPRC (also known as CD45R); and down-regulation of neuronal cyclin-dependent kinase 5, regulatory subunit 2 (CDK5R2) (only significant for p39), neuronal marker microtubule-associate protein 2 (MAP2) and the SNARE protein complexin 1 (CPLX1), compared to brains of patients on ART. As also noted in other studies [57], confirmatory QPCR analyses generally yielded higher FC results than parallel microarray analyses (Supplementary Table S5). Four of the changes in gene expression in the brain were also tested at the protein level by immunohistochemistry (Figure 2D). We observed increased expression of CD68, C3c and CD45R proteins in untreated HAND compared to uninfected control tissue, and an amelioration of the protein dysregulation in patients with HAND on ART. MAP2 protein was down-regulated in HAND brains compared with brains from uninfected subjects, and it was partially restored in patients under treatment for HIV-1 infection. Taken together, our findings parallel recent reports of the effects of ART on gene expression in peripheral tissues [24], [27], [29] in that treatment of HIV-1 infection by ART is also accompanied by marked reduction of the virally-induced dysregulation of gene expression in brain tissues. Treatment adherence further reduces the number of genes and biological pathways affected. In the next level of analysis, we conducted unsupervised hierarchical clustering for 2073 dysregulated transcripts implicated in HAND (Figure 2A and Supplementary Table S2) to visualize and group individual gene expression profiles from all 21 subjects in this study (Figure 3A). We used normalized Robust Microarray Average (RMA) values from duplicate samples for each subject for greater statistical power (Methods). Three main clusters of subjects were identified. With the exception of HAND5, untreated HAND patients clustered as one group (cluster 2 in Figure 3A) distinct from the other two clusters. ART-treated HAND patients clustered with uninfected controls in two co-mingled clusters, cluster 1 containing subjects C1, C2, C3 and ART1 and cluster 3 including subjects C4, C5, C6, ART2, ART3, ART4, ART5, ART6, and the remaining untreated HAND5. Compared to cluster 3, cluster 1 was more phylogenetically distant from untreated HAND, but overall clusters 1 and 3, separately or combined, were significantly different from cluster 2. The t-test values computed using RMA values for all transcripts in cluster 1 vs. cluster 2, cluster 3 vs. cluster 2, and cluster 1+cluster 3 vs. cluster 2 were 1×10−221, 7.8×10−25, and 1.96×10−24, respectively. These results demonstrate that for the 2073 transcripts tested, gene expression in brain tissues of patients on ART tends to resemble that of uninfected subjects, indicating that ART is associated with a profound reduction in the extent of dysregulation of HAND-related genes in the brain. These results also suggest that the differences related to HIV-1 disease (and treatment) override potential differences resulting from genetic heterogeneity of individual donors. The statistical similarity of ART transcriptomes with HIV-1-negative controls with respect to presumptive HIV-1-impacted genes (Figure 3A) prompted us to directly compare gene expression changes of ART brain tissue to expression in HAND brain tissue, without filtering microarray results through HIV-1-negative controls. This analysis inquires whether control of HIV-1 replication by ART removed viral perturbations to cellular gene expression in the brain, analogous to longitudinal studies of peripheral tissues from HIV-1-infected subjects pre and post-ART [24] which are not feasible for the brain. Using normalized RMA datasets for over 54,000 transcripts detected by the U133 chipset, we compared 7 treated patients to 8 untreated patients delimiting the results by FC of 1.5 and t-test value of 0.05; the complete list these differentially expressed transcripts is shown in Supplementary Table S6. In calculating the ratio of gene expression in ART to HAND, positive FC values in ART indicate an association of increased gene expression with treatment and negative FC indicate reduced cellular gene expression associated with treatment. Overall, we identified 640 significantly up-regulated and 276 down-regulated genes in this analysis (Supplementary Table S6). Samples from patients with treated HAND show increased expression (relative to untreated HAND) of many genes involved in neuronal functions including synaptoporin (SYNPR, FC 6.55 and p = 6.5×10−4), neurofilament, light polypeptide (NFEL, FC 6.15, p = 4.6×10−4), and synaptotagmin IV (SYT4, FC 4.16, p = 5.8×10−4) compared to samples from untreated patients. Conversely, genes involved in immune activation including CD74 antigen (CD74, FC −2.49, p-0.013), complement component 1, q subcomponent, C chain (C1QC, FC −2.47, p = 0.022), and interferon-induced protein with tetratricopeptide repeats 2 (IFIT2, FC −2.35, p = 0.007) were reduced in expression in treated HAND patient samples relative to samples from untreated patients. For reference, Supplementary Table S6 also lists respective ART microarray data normalized to HIV-1-negative controls. Notably, all but 7 of the genes in treated patients that increased in expression relative to untreated HAND were unchanged in the ART/HIV-negative control comparison (Supplementary Table S6), suggesting that treated patients express these genes at normal (control) levels. Similar normalization of expression (272 out of 276) was found for genes that were down-modulated in ART versus HAND. To put these findings in the context of the biological pathways potentially affected by ART, we employed GAzer to compare the complete ART microarray datasets to those from HAND. Figure 3B depicts 16 biological processes that were most significantly changed in ART relative to HAND and Supplementary Table S7 provides complete statistics for this analysis. Twelve up-regulated processes in treated versus untreated HAND, with Z-scores ranging from 9.54 to 5.02, were related to neuronal function and repair. The down-modulated pathways included, in decreasing order of significance, immune responses, inflammatory responses, apoptosis, and responses to stress. These results suggest that ART reverses many dysfunctional processes of untreated HAND represented in gene ontology analysis. To provide an alternative view of the extensive up-regulation of cellular processes in ART relative to HAND, up-regulated genes with p-values of ≤0.01 included in Supplementary Table S6 were analyzed by STRING to identify predicted gene interaction networks (Figure 4). In this analysis ART was associated with improved synaptic transmission including synaptic vesicle system, nervous system development including cytoskeleton associated proteins, and GABA neurotransmission networks. These findings strongly suggest that by reducing HIV-1 replication, ART also reduces triggers to aberrant gene expression in the brain. The HAND patients in this study share cognitive dysfunction whether they were untreated or treated with ART (Table 1). To begin to identify transcripts that may contribute to HAND development or persistence despite ART, we used t-test to compare significantly changed genes in untreated (Supplementary Table S2) and treated (Supplementary Table S3) patients with HAND; a gene whose expression was not significantly different in the HAND versus ART comparison (p>0.05) was considered similarly dysregulated in both groups of patients relative to uninfected subjects. We have identified 43 such up-regulated and 42 down-regulated genes; they are listed grouped into biological categories in Supplemental Table S8, selected genes with their expression statistics are listed in Figure 5A, and their heatmap expression profiles in all subjects in this study are shown in Figure 5B. Considering functional characterization, genes related to immune responses were up-regulated in both HAND and ART samples including complement receptor 1 (CR1), chemokine, CXC motif, ligand 2 (CXCL2), major histocompatibility complex, class II HLA-DQB1 and interferon-mediated antiviral responses including interferon-induced protein with tetratricopeptide repeats 1 (IFIT1); interferon-induced protein 44 (IFI44); myxovirus resistance 1 (MX1); 2′,5′-oligoadenylate synthetase 1 (OAS1), and signal transducer and activator of transcription 1 (STAT1). Cell cycle pathway was dysregulated in both treated and untreated HAND patients, with some sets of genes up-regulated and others down-regulated (Figure 5A and 5B). Over-expression of selected transcripts in interferon or chemokine pathways in both HAND and ART brain samples was confirmed by real-time PCR (Figure 5C), over-expression of HLA Class II alleles and proliferating cell nuclear antigen (PCNA) was also demonstrated by immunohistochemistry (Figure 5D). Of particular interest, common down-regulated genes in treated and untreated patients with HAND included myelin-related genes myelin-associated oligodendrocyte basic protein (MOBP), myelin transcription factor 1 (MYT1) and myelin basic protein (MBP). Down-regulation of MOBP and MYT1 was confirmed by real-time PCR, with the MYT1 gene being particularly suppressed in the ART group (FC = −38.38, p = 1×10−5) (Supplementary Table S5). This result is consistent with histopathological detection of myelin pallor in autopsy brain tissues from some patients with HAND [58], [59], although other explanations also exist [60]. Pearson's formula was applied to determine the correlation of the level of expression of each transcript with HIV-1 load in plasma, cerebrospinal fluid (CSF) and brain. Examples of genes correlating positively with plasma viral load are shown in Figure 6A, including histocompatibility loci HLA-B-G and F, interferon-gamma-inducible protein 30 (IFI30), OAS1, and Cathepsin S (CTSS). Transcripts with a positive (>0.5) or negative (<−0.5) correlation with viral load were analyzed using the gene ontology software Expression Analysis Significance Explorer (EASE) to identify the pathways that correlated most with viral load (Figure 6B). Pathways positively correlated with viral load were similar in the three compartments (brain, CSF, plasma), including several immune activation responses. The main difference among the three compartments was the level of significance of the changes, as represented by the EASE score. All the pathways implicated in immune response correlated better with plasma viral load than with brain viral load. Even though CSF data were available for only 9 of the 15 infected patients, the positive correlations for CSF were higher than those for brain, although less strong than those observed for plasma. The categories of antigen presentation and processing were correlated only with viral load in plasma. Pathways down-regulated in correlation with viral load differed depending on the compartment tested. No biological pathway correlated negatively with brain viral load. Pathways negatively correlated with plasma viral load included synaptic transmission, cell communication, transmission of nerve impulse, organogenesis and neurogenesis. A wide variety of metabolic pathways negatively correlated with CSF viral load. Overall, grouping genes engaged in similar biological functions indicates that gene groups induced in the brain correlated best with virus burden in the periphery and that virus burden in the brain, unlike viral load in plasma or CSF, was uncorrelated to suppression of expression of any gene group. We employed functional genomics to investigate the potential effects of antiretroviral treatment on brain pathophysiology in a cohort of patients who died with HAND. Our results suggest that ART profoundly, but not completely, alleviates aberrant gene expression in brain tissues of these patients. These findings may shed light on the molecular basis of HAND persistence despite treatment. Several points should be made about this work. The foundation of this work is a new comprehensive database of global gene expression profiles in brain tissues of patients with HAND. The profiles described here complement published datasets from previous array studies in HAND [41]–[44] with important differences. For the first time in this disease, we analyzed brain transcriptomes on the basis of the antiretroviral treatment of patients, revealing two distinct, largely non-overlapping groups of aggregate gene expression profiles termed ART and untreated HAND. The ART and untreated HAND profiles also formed separate clusters in unsupervised hierarchical cluster analysis [61] of individual patient datasets, confirming that they are phylogenetically distant from each other based on a large sets of aberrantly expressed genes used in this analysis (Figure 3A). Importantly, the hierarchical clustering distinction between treated and untreated patients was statistically more prominent than potential distinctions in gene expression related to other patient characteristics in our cohort including HIVE and intravenous drug use (Figure 3A and Supplementary Figure S2). Therefore, in most analyses in this work we considered datasets from patients with and without HIVE as one disease category, stratified only on the basis of ART. These results suggest that ART, through effects on HIV-1-associated changes in cellular gene expression [27], [62], [63], is one of the key biological variables governing the extent and pattern of gene dysregulation in molecular profiles of HIV-1 brain disease. Another important difference with previous HAND array studies concerns the histological regions of the brain tested. We evaluated frontal deep white matter, whereas most of the previous studies focused on the neighboring cortical gray matter [41]–[44] or small gene sets in both brain regions [41]. White matter is the primary site of HIV-1 infection and HIV-1-associated neuropathologies [46], [47] and frontal cortex is one of the sites of synaptic and dendritic damage consequent to this infection [64], [65]. Transcriptomes from these two areas reflect different regional physiologies and therefore different aspects of HIV-1 neuropathogenesis. With these caveats, a meta-analysis summarized in Supplementary Table S9 identified a small number of genes involved in interferon-related responses (IFIT1, IFITM1, IFI44, MX1), synaptic functions (SYN1, SYN2, GABRG2, MAP2), and cell cycle (CDC42, CDK5R1, R2) that were dysregulated in common in the present and previously published HAND datasets. These genes and biological pathways may represent features of HIV-1-associated neuropathogenesis common to white and gray matter. During the last twenty years, individual inflammatory and neurodegenerative mediators in HAND brains were demonstrated by immunocytochemistry, in situ hybridization, and other methods (reviewed in [66]–[69]). The bulk of this work was conducted with brain tissues from untreated patients, as brain autopsies after introduction of ART have become less frequent [70]. Consistent with these observations, the untreated HAND profile defined here reveals a broad and extensive dysregulation of cellular gene expression in brain tissues, with 1470 HAND-associated aberrantly expressed genes, up-regulation of immune activation, antiviral responses, and inflammation, and down-modulation of neuronal functions, neuronal repair, and cell cycle. It should be noted that our untreated HAND profile included transcriptomes from patients with and without HIVE (Figure 2 and 3). HIVE is a characteristic histopathology associated with high HIV-1 burdens in the brain [71], [72], and previous array studies documented differentially expressed transcripts and biological pathways potentially attributable to the extensive infection in the brain [43]–[45], [73]. We confirmed some of these differences in our untreated group with and without HIVE in gene ontology analysis but it is noteworthy that they differed mainly in the degree of dysregulation and not the biological pathways affected (Figure 1). Thus, at the level of a transcriptome analysis, the altered expression of some transcripts attributed to high HIV-1 burdens in HIVE [43], [44] contributes to but does not change the overall statistical characteristics of the molecular phenotype of HAND. Our results are consistent with reports of limited correlation between clinical manifestations of HAND and HIVE histopathology [58], [74], [75]. Rather, untreated HAND appears to correlate better with presence of inflammatory mediators and diffusely activated macrophages and microglial cells in the brain than with virus burdens in the tissue per se [17], [58], [75]–[78]. Independent microarray studies in other patient populations are needed to determine whether these gene expression profiles are fully representative of untreated HAND. In general, transcriptome profiles of disease can serve as a platform for verification of results obtained in studies of individual physiological processes [33], [79], [80] and as a tool for discovery of new ones. In this context, a number of “novel” (i.e., relatively new to the HAND literature) dysregulated genes in our untreated HAND database may shed light on the process of HAND pathogenesis and merit further investigation. For example, transcripts encoding apolipoprotein C-I and C-II (APOC-I and APOC-II) were among the most up-regulated in the untreated HAND dataset, with FC of 5.71 (p = 9.7×10−5) and 3.35 (p = 0.009), respectively (Supplementary Table S2). Dysregulation of lipid metabolism linked to APOE polymorphism is a marker of AD [81] and was indicated in HIV-1 dementia [82], [83]. While we could not find reports on APOC-I in HAND, this lipoprotein was found in association with beta-amyloid plaques in AD brains and expression of human APOC-I allele in native APOC-I null mice was shown to impair learning and memory [84]. Conversely, hemoglobin α-2 and hemoglobin β (HBA-2 and HBB) were among the most down-regulated transcripts in our untreated HAND dataset (FC of −5.31 and −4.22; Supplementary Tables S2 and S5). Neuronal hemoglobins are members of the globin superfamily which are predominantly expressed in neurons and may play an important role in neuroprotection [85]. Consistent with our findings, expression of neuronal hemoglobin is reduced or absent in disease affected brain regions in patients with several neurodegenerative conditions including AD and Parkinson's disease [86]. The major finding of this work is the profound difference between the extensive global gene dysregulation observed in brain tissues of untreated patients with HAND and muted gene changes in their treated counterparts. The magnitude of ART effects in the brain suggested by our results was surprising given the limited clinical outcomes of ART on HAND [7], [87], including in the cohort evaluated here (Table 1), and variable findings in the CSF of treated patients [12]–[16]. The ART effects on the brain were inferred because we could not test brain RNA in the same individuals before and after initiation of therapy, as is possible in microarray studies with peripheral tissues [24], [27], [29], [31]. However, the overall effects of ART on altered gene expression were remarkably similar in the periphery and brain. This was evident in markedly fewer dysregulated genes in treated compared to untreated patients, in our case 253 versus 1470; and in a global shift in gene expression patterns from aberrant in the absence of treatment to muted dysregulation under treatment, for example in peripheral CD4+ T cells [27], [29], lymphoid tissue [26], and brain here (Figure 3A). Importantly, both in the CD4+ T lymphocyte study [27] and in the present work, gene expression profiles of treated patients were statistically similar to those of HIV-1-negative controls, suggesting a trend toward normalization of gene expression under ART. We confirmed this trend for selected gene products in the present work by real-time PCR and immunocytochemistry (Figure 2). These results suggest that ART regimens, which generally include at least one brain penetrant antiviral compound (Table 1), are similarly effective in mitigating global molecular changes in the brain and in peripheral tissues. We noted two reciprocal effects of ART on gene dysregulation in brain tissues illuminating the systemic and brain-specific aspects of HAND pathogenesis. One is a significant and broad moderation of up-regulated genes linked to HIV-1 induced antiviral and inflammatory responses thought to drive HAND pathogenesis, including interferon-related ISG15 and IFIT3, macrophage markers CD68, CD163, and CD14, and chemokines and chemokine receptors CCL8, CCR1, and CXCR4 [67]–[69] (Figure 2 and 3). Interestingly, this effect of ART was common to diverse tissues examined by microarrays including brain (this study) and CD4+ T cells, macrophages, lymph nodes, and intestinal mucosa tested by others [24], [27], [29], [31], and thus it likely represents a system-wide response to suppression of HIV-1 replication. Although gene ontology pathways containing these genes in patients under ART were still up-regulated compared to controls (Figure 2B and Figure 5), our results suggest that ART can alleviate a surprisingly large number of deleterious responses in the brain that have been linked previously to HIV-1 infection in model systems [88]. This causal link is further strengthened by an apparent correlation in this work between treatment compliance and extent of gene dysregulation in the brain (Figure 2). The other effect of ART in the brain we observed was specific to the nervous system and it involved normalization of a large number of down-regulated genes and biological pathways linked to nervous system functions and by extension to neurocognitive disease. For example, the bioinformatics tool STRING [89] identified nervous system development, synaptic transmission, and GABA-neurotransmission pathway as the three major predicted interaction networks of genes that approached normal expression in treated patients (Figure 4). On a smaller scale, our confirmatory tests showed that products such as MAP2 and complexin-1 (CPLX1) were significantly down-regulated in untreated patients at RNA and protein levels and they were expressed at control-like levels in tissues from treated patients (Figure 2). It is conceivable that restoration of normal expression of at least some of these genes under ART would restore some aspects of normal brain physiology [7]. Although treated patients in our cohort still manifested HAND prior to death, our results suggest that they may have already shifted to a milder molecular profile in the brain that had more in common with HIV-1-negative controls than untreated patients. Of interest, the global changes in brain cell gene expression seen by microarrays correlated positively with plasma but not brain virus burdens (Figure 6). This association may be analogous to the clustering of brain transcriptomes independently of encephalitis and brain HIV-1 burden by commonly dysregulated biological pathways. In any case, the single measurement of brain virus burden at autopsy may not capture the chronic insult to brain function suffered by patients living years with HIV-1 infection, albeit with some control exerted by ART. Perhaps the most intriguing contribution of the present array analysis lies in the ability to discern patterns of abnormality that persist in cognitively-impaired patients who are on central nervous system penetrant ART, and to distinguish these patterns from HAND in the untreated state. In the bioinformatics sense, we used ART as a biological filter to reduce the overall gene expression disturbance in brain transcriptomes of patients with HAND and through that determine whether continuing dysregulation of gene expression could play a role in continuing brain disease. The results indicate that continuing up-regulation of innate and adaptive immune responses are an important part of brain abnormalities in ART-treated dementia. Over-expression of Class II MHC in the brain persists despite therapy in our study; it has been associated with many neurodegenerative diseases [90]. Defects in myelin metabolism, shown for HAND at the gene expression level here and indicated previously by neuropathological observation of myelin pallor in brains of patients with HAD [58], are also common to many other neurodegenerative diseases [38], [91]. Activation of interferon-related genes often found in symptomatic HIV-1 infection may underlie abnormalities in cell function [28]. Such changes, coupled with cell cycle perturbations, may support emerging magnetic resonance spectroscopy studies that have demonstrated persistent white matter inflammation in patients with HIV-1-related cognitive impairment [92]. However, it is unclear what particular aspects of immune activation or response are relevant to nervous system dysfunction, and how to distinguish deleterious gene products from those that may function in a neuroprotective manner. In simian immunodeficiency virus infection, innate immunity, IL-6, and interferon responses are important elements in brain viral control [56], [93], but in a recent study expression of interferon-α in the brain was conclusively linked to neuronal dysfunction in a mouse model of HIVE [94]. Careful analysis of individual genes identified here may begin to clarify the mechanism of HAND persistence under treatment. Human brain samples and clinical data were obtained from the Manhattan HIV Brain Bank (MHBB), a member of the National NeuroAIDS Tissue Consortium, under an Institutional Review Board-approved protocol at the Mount Sinai School of Medicine. Written informed consent was obtained from all subjects in this study or their primary next-of-kin. HIV-1 and gene expression analyses were conducted on de-identified brain samples under an “exempt” status approved by an Institutional Review Board of St. Luke's-Roosevelt Hospital Center. Study subject information is listed in Table 1. Fifteen HIV-1-positive subjects used in this study were chosen on the basis of having HAND and the presence or absence of HIVE as determined by neuromedical or neuropsychological evaluation and postmortem neuropathology [95], and then were further categorized as either dying on or off ART. Fourteen of HIV-1-positive patients in this study died with HAD; one patient designated ART2 in Table 1 was classified as MND based on his lack of emotive/behavioral criteria [96]. Patients who displayed HAND without HIVE histopathology at autopsy met American Academy of Neurology criteria for HAND regardless of ART status. This definition requires demonstration of cognitive and functional impairments, and the presence of emotive or motoric phenomena [96]. Except for patient ART2, patients with HIVE were similarly impaired, regardless of ART status, or had histories of HAND on medical record review. Patients felt to have cognitive impairments) due to non-HIV-1 causes (neuropsychological impairment – other, as described in [95] were excluded. All patients dying on ART had substantive treatment histories, ranging 1 to 7 years prior to demise. The ART regimens at death had a mean duration of 16 months (range, 3 months to 3 years). For five of seven patients on ART, treatment compliance estimates were made by self-reported 4 day recall [97]. HIV-1-negative subjects were chosen on the basis of normal neurological function (as determined by chart review) and normal neurohistology. All neurohistologic diagnoses were rendered by a board-certified neuropathologist (SM) and a minimum of 50 sections were examined for each brain. The following brain pathology definitions were used in the present work (Table 1): Normal: no brain pathology; HIVE: HIV encephalitis; HIVE*: HIVE was limited to the basal ganglia (pid 30015); Minimal: minimal histopathological changes including trivial microscopic abnormalities such as an isolated vermal scar (pid 10119), a venous ectasia (pid 10063), atherosclerosis and minimal perivascular inflammation sub-threshold for diagnosis (pid 10001), and minimal perivascular inflammation sub-threshold for diagnosis (pid 10015). At the time of autopsy, coronal sections of brain were snap-frozen and maintained in −85°C until sub-dissection. Effort was made to keep the post mortem interval (PMI) to a minimum; the PMI for subjects in this study are listed in Table 1. Brain samples for this analysis were obtained from the centrum semiovale (deep white matter) at the coronal level of the genu of the corpus callosum. Multiple samples from the same region were dissected for gene expression profiling, real-time PCR (QPCR), and protein assays. Equivalent regions from the contralateral hemisphere were formalin fixed and utilized for immunohistochemistry. Total DNA and RNA were isolated from human brain tissue by, respectively, DNeasy Blood and Tissue Kit and RNeasy Mini Kit (Qiagen, Valencia, CA) according to the manufacturer's protocol. RNA was quantified by spectrophotometry and RNA quality was verified by spectrophotometry and agarose gel electrophoresis. RNA was then treated with DNAse I (Fisher Healthcare, Houston, TX). cDNA was synthesized using the Superscript First-Strand Kit (Invitrogen, Carlsbad, CA) for quantitative analysis and the WT-Ovation™ RNA Amplification System (NuGEN Technologies, Inc., San Carlos, CA) for relative analysis according to the manufacturer's protocol. Viral RNA and DNA burdens in human brain were determined by quantitative real-time PCR using primers designed based on HIV-1 consensus sequences for the group of patients in this study. We first screened patient brain samples for the presence of HIV-1 content by a standard nested PCR for viral DNA or RNA (cDNA) as previously described [98]. To achieve broad detection of diverse Clade B HIV-1 species, first-round PCR was performed using custom designed primers for a conserved Clade B gag consensus sequences: SQ5 (+) 5′-CAA ATG GTA CAT CAG GCC ATA TCA CC-3′ and SQ3′ (−) 5′-CCC TGA CAT GCT GTC ATC ATT TCT TC-3′. For nested PCR step we used primers SQ5′ (above) and SK39 [99]; the PCR products were resolved on an agarose gel and detected by Southern Blot hybridization with (32P)-labeled probe SK19 [99] (Supplementary Figure S1). The HIV-1 gag nested PCR amplicons from individual HIV-1 DNA positive patients were sequenced and used to design patient consensus primers (5′ (+) QSQ5 5′-ACC CAT GTT T(T/A)C AGC ATT ATC AGA-3′ and 3′ (−) QSQ3 5′-GAT GTC CCC CCA CTG TGT TT-3′) and Taqman probe (HSQP 6FAM-AGC CAC CCC ACA AGA-MGBNFQ). For real-time PCR amplification, 2 µl of brain tissue DNA or 5 µl of cDNA were combined with 2× Universal Master Mix (Applied Biosystems, Carlsbad, CA), 900 nM consensus primer (custom synthesized by Invitrogen) and 200 nM probe (synthesized by Applied Biosystems; QPCR conditions were essentially as described [100]. Standard curve for viral DNA and cDNA quantification was constructed from graded amounts of HIV-1 NL4-3 plasmid DNA. Viral DNA burdens were normalized to total cellular DNA content by β-globin amplification and expressed as HIV-1 DNA copies/number of cells calculated from a β-globin DNA standard curve, with 2 copies of β-globin gene equaling one cell. Viral RNA burdens were normalized by tissue glyceraldehydes-3-phosphate dehydrogenase (GAPDH) content and expressed as number of viral copies in 1 µg tissue RNA [100]. Microarray experiments were conducted at the Bionomics Research and Technology Center in EOSHI University of Medicine and Dentistry of New Jersey. Total RNA were extracted from tissue samples using the RNeasy Mini Kit (Qiagen) followed by DNase I treatment. RNA qualities were assessed by electrophoresis using the Agilent Bioanalyzer 2100 and spectrophotometric analysis prior to cDNA synthesis. Fifty nanograms of total RNA from each sample were used to generate a high fidelity cDNA for array hybridization using NuGen WT-Ovation Pico RNA Amplification. Detailed protocols for sample preparation can be found at http://www.nugeninc.com. After fragmentation and biotin labeling using NuGen Encore Biotin Module, the samples were hybridized to Affymetrix Human Genome 133 plus 2.0 arrays. Washing and staining of all arrays were carried out in the Affymetrix fluidics module as per the manufacturer's protocol. The detection and quantitation of target hybridization was performed with an Affymetrix GeneChip Scanner. Data were assessed for array performance prior to analysis. The majority of patient samples were analyzed in duplicates starting from the cDNA synthesis step; in some cases second analysis was on an adjoining brain sample. The .cel data files generated by the Affymetrix microarray hybridization platform were analyzed by the ArrayAssist software (Stratagene, Santa Clara, CA). Probe level analysis was performed using the RMA algorithm. After verification of data quality by Affymetrix internal controls and signal distribution analysis as described in Stratagene ArrayAssist Protocol, data was transformed using variance stabilization and logarithm transformation with a base of 2. Fluorescence values were normalized by mean intensities of all chip samples. Means of normalized expression values were calculated for duplicate samples from each individual, and these values were either used directly in some analytical programs (see below) or employed to calculate fold change (FC) in the transcript compared to HIV-1-negative controls. Genes showing FC values above 1.5 or below −1.5 and unpaired t-test p-values of <0.05 were defined as significantly changed. In some analyses we applied a t-test cutoff of p<0.01. To test for effect of antiretroviral treatment, we calculated the FC for significantly modulated transcripts separately from untreated and treated patients, the latter subdivided into all-treated and a subset without two known low-compliant patients, versus uninfected controls. We also compared directly normalized expression values of selected transcripts from treated and untreated patients to generate the treated versus untreated FC values that were not filtered through HIV-1-negative controls. To compare brain samples of all the individuals included in the study we clustered the expression profiles using unsupervised Hierarchical Clustering (Average Linkage Clustering) and heatmap visualization software from Genesis [101] available at http://genome.tugraz.at/. This type of analysis allows us to identify the similarities and differences in the expression patterns of groups of patients and/or transcripts. For broad characterization of gene expression changes in untreated and treated patients we conducted gene set and gene ontology analysis using GAzer [50] (http://expressome.kobic.re.kr/GAzer/index.faces). GAzer is a web-based tool that identifies, by a parametric statistical analysis of complete primary normalized microarray data, over-represented sets of genes (functionally related genes) rather than individual genes. This type of analysis compensates for the fact that small changes not seen at the gene level are often detected when the gene set as a whole is examined [50], [51]. When analysis was limited to smaller sets of genes defined as differentially expressed, we performed functional categorization of gene families by EASE [102], available at the NIH web site (http://david.abcc.ncifcrf.gov/). The predicted biological pathway and network relationships among differentially expressed genes in our array datasets were identified using the Search Tool for the Retrieval if Interacting Genes/Proteins (STRING) (http://string-db.org/) [89]. Changes in expression of selected genes identified by microarrays analysis were validated in the same or adjoining brain samples by QPCR using Taqman chemistry and probes from the Universal Probe Library (Roche, Indianapolis, IN). Primers were designed using the online ProbeFinder software available at the Roche Universal Probe Library Assay Design Center (http://www.roche-applied-science.com). The QPCR reactions contained 2 µl of cDNA generated from tissue RNA obtained as described above, 10 µl of 2× Universal Master Mix (Applied Biosystems-ABI), 0.2 µl of each forward and reverse primers at 200 nM, 0.2 µl of probe at 100 nM, and RNAse/DNAase-free water. All reactions were performed in duplicate and were run in a 7500 real-time PCR system (ABI). Raw data was analyzed using the 7500 System SDS Software (ABI). Data was normalized using 2 housekeeping genes, GAPDH and ribosomal protein S18 (RPS18) to assure reproducibility. Relative quantification employed the comparative threshold cycle method (Applied Biosystems Technical Bulletin n°2). For immunohistochemical analysis, formalin fixed blocks were taken from the frontal white matter of the autopsy brains of 4 normal, 6 HIV-1-non-treated, and 4 HIV-1-ART-treated individuals. A microarray block consisting of 3 tissue punches diameter of 1 mm) from each block was constructed. 5 µM serial sections were cut and immunohistochemistry performed with an array of antibodies listed in the Table 2. Formalin-fixed, paraffin-embedded sections were deparaffinized with xylenes, hydrated in graded alcohols, and incubated with 3% H2O2 in methanol. Following washing, sections were boiled in Target retrieval solution (DAKO Corp., Carpinteria, CA), subsequently incubated in a serum free protein blocking solution and then incubated with the primary antibodies indicated in Table 2. CD45 (PTPRC) was incubated overnight at 4°C and the all the remaining antibodies for 1 h at room temperature. Primary antibodies were detected with peroxidase anti rabbit or mouse IgG ImmPRESS (DAKO Corp.) reagent and counterstained with hematoxylin. Slides were visualized in a light microscope and either one, three or six 0.03 mm2 areas of white matter staining from each case were photographed using a 40× objective and a Nikon Coolpix II digital camera attached to the microscope by a Coolpix MDC lens. The intensity of illumination and position of sub-stage condenser on the microscope were constant for all images. The number of areas taken depended on the type of quantitative analysis that followed. The percentage area occupied by cells immunoreactive for CD68, CD45 and HLA DP, DQ, DR was quantitated by analysis of six 0.03 mm2 images using a proprietary automated morphometric analysis software [103]. The mean intensity of MAP2 and STAT1 staining was analyzed on one 0.03 mm2 image, utilizing Image J software (http://rsbweb.nih.gov/ij/). Each image was converted to 8-bit grayscale and a mean gray value of the pixels in the full photographic image was measured after a background subtraction and inversion were performed. For C3c and PCNA, cell counts were done by eye using either three 0.03 mm2 (C3c) or the full punch (PCNA). Statistical analysis was performed using StatView (V.5.0.1) (Adept Scientific, Bethesda, MD). The principal statistical test used was the Analysis of Variance for single comparisons (ANOVA). Follow-up post-hoc tests were conducted when required. Significance values were set at 0.05 and below. Correlation analysis between virus loads in plasma, CSF, or brain and gene expression in the brain was performed using the Pearson's correlation formula in Microsoft Excel software (Microsoft Corporation, Redmond, WA). Transcripts showing positive (>0.5) or negative correlation (<−0.5) were categorized into biological functional pathways using EASE software. The microarray results presented here are available in the Gene Expression Omnibus database (www.ncbi.nlm.nih.gov/geo) in a MIAME compliant format under accession number GSE28160. The Gene ID numbers for the genes mentioned in the text are listed below. The ID number corresponds to the ‘National Center for Biotechnology Information database’ (http://www.ncbi.nlm.nih.gov/gene): CD4 (920), C3 (718), CD68 (968), PTPRC (5788), CDK5R2 (8941), MAP2 (4133), CPLX1 (10815), SYNPR (132204), NFEL (4747), SYT4 (6860), CD74 (972), C1QC (714), IFIT2 (3433), CR1 (1378), CXCL2 (2920), HLA-DQB1 (3119), IFIT1 (3434), IFI44 (10561), MX1 (4599), OAS1 (4938), STAT1 (6772), PCNA (5111), MOBP (4336), MYT1 (4661), MBD (4155), IFI30 (10437), CTSS (1520), IFITM1 (8519), B2M (567), CD14 (929), APOC-I (341), APOC-II (344), HBA-2 (3040), HBB (3043), IFI16 (3428), TLR7 (51284), SYN1 (6853), SYN2 (6854), GABRG2 (2566), CDC42 (998), CDK5R1 (8851), GAPDH (2597), RPS18 (6222).

10.1371/journal.ppat.1005783

TPL-2 Regulates Macrophage Lipid Metabolism and M2 Differentiation to Control TH2-Mediated Immunopathology

Persistent TH2 cytokine responses following chronic helminth infections can often lead to the development of tissue pathology and fibrotic scarring. Despite a good understanding of the cellular mechanisms involved in fibrogenesis, there are very few therapeutic options available, highlighting a significant medical need and gap in our understanding of the molecular mechanisms of TH2-mediated immunopathology. In this study, we found that the Map3 kinase, TPL-2 (Map3k8; Cot) regulated TH2-mediated intestinal, hepatic and pulmonary immunopathology following Schistosoma mansoni infection or S. mansoni egg injection. Elevated inflammation, TH2 cell responses and exacerbated fibrosis in Map3k8–/–mice was observed in mice with myeloid cell-specific (LysM) deletion of Map3k8, but not CD4 cell-specific deletion of Map3k8, indicating that TPL-2 regulated myeloid cell function to limit TH2-mediated immunopathology. Transcriptional and metabolic assays of Map3k8–/–M2 macrophages identified that TPL-2 was required for lipolysis, M2 macrophage activation and the expression of a variety of genes involved in immuno-regulatory and pro-fibrotic pathways. Taken together this study identified that TPL-2 regulated TH2-mediated inflammation by supporting lipolysis and M2 macrophage activation, preventing TH2 cell expansion and downstream immunopathology and fibrosis.

Chronic helminth infections can cause significant morbidity and organ damage in their definitive mammalian hosts. Managing this collateral damage can reduce morbidity and preserve vital tissues for normal organ function. One particular consequence of some chronic helminth infections is the deposition of fibrotic scar tissue, following immune responses directed towards helminth material. In this study we tested the role of a particular signalling kinase, TPL-2, and identified that it critically regulated the magnitude of fibrotic scarring following infection. Using several murine models with genetic deletions of TPL-2 in either all cells or specific deletion in subsets of immune cells (Map3k8–/–Map3k8fl/fl) we identified that expression of TPL-2 in myeloid cells was essential to prevent severe immune-mediated pathology. Using genome-wide analyses and metabolic assays, we discovered that TPL-2 was required for normal lipid metabolism and appropriate activation of myeloid cells / macrophages to limit fibrosis. These results revealed a previously unappreciated role for TPL-2 in preventing severe pathology following infection. Thus, activating this pathway may limit immune mediated pathology following chronic helminth infection. More broadly, this pathway is being targeted to treat inflammatory diseases and cancer [1, 2]. This study would suggest that caution should be taken to prevent untoward co-morbidities and fibrosis-related pathologies in patients when targeting TPL-2.

Immune-mediated pathologies and fibrotic scarring are a major cause of global morbidity and mortality. This is due, in part, to a shortage of available drugs and a lack of novel therapeutic targets to limit fibrogenesis, highlighting a major unmet medical need [3, 4]. Chronic infection resulting in recurring inflammation and wound repair can lead to tissue remodelling, fibrosis and ultimately organ failure. Infection with the parasitic blood fluke, Schistosoma mansoni, can cause severe intestinal and hepatic pathologies caused by fibrotic lesions surrounding trapped parasite material. Parasite eggs become lodged within vascularised tissue invoking a distinctive eosinophil and macrophage (MΦ)-rich type-2 immune-mediated granuloma [5]. TH2-cell derived IL-4 and IL-13 [6] stimulate IL-4 receptor (IL-4R)-expressing MΦ’s [7, 8] to develop an M2 or alternative activation (AA) state characterised by expression of Arginase (Arg1), Resistin-like molecule alpha (Retnla, Fizz-1) and chitinase-like molecules (Chi3l3, Chi3l4) [9]. Animal models have indicated that IL-4R-dependent M2-MΦ’s are essential to 1) prevent fatal intestinal damage and sepsis following schistosome infection [7]; 2) orchestrate tissue remodelling and fibrotic responses [10–12] and 3) regulate TH2 cell proliferation and activation [13–15]. Despite the clear and well-documented importance of M2-MΦ’s during schistosome infection and the resulting immune-mediated protection, pathology and regulation, the critical regulatory proteins that control M2-MΦ differentiation are poorly understood. The MAP3 kinase, TPL-2 (also known as COT and encoded by Map3k8) is ubiquitously expressed, phosphorylating and activating MEK1/2 following stimulation of Toll-like receptors (TLRs) and the receptors for TNF and IL-1β, leading to the activation of ERK1/2 MAP kinases [16]. TPL-2 is required for TH1 and TH17-associated inflammation and is essential for the development of autoimmunity and immunity to bacterial and protozoan pathogens [17–20]. In MΦ’s, TPL-2 is required for the synthesis and secretion of a variety of cytokines and chemokine’s following classical activation (CA) with TLR ligands [17, 20–29]. However, it is unclear whether TPL-2 controls M2-MΦ function to regulate chronic TH2-associated inflammation and immunopathology. Two distinct inflammatory pathways contribute to fibrogenic responses; classical, pro-inflammatory type-1/17 and TGF-β-mediated fibrosis [30] and type-2 inflammatory pathways leading to IL-4R-dependent fibrosis [4]. It was recently reported that TPL-2–deficient mice, or inhibition of ERK [31], protected mice from type-1/TH17 and TGFβ-mediated pulmonary fibrosis following bleomycin treatment [31] and from hepatic fibrosis following carbon tetrachloride and methionine choline-deficient diet-induced fibrosis [32]. As expected, Map3k8–deficient Kupffer cells had reduced TLR-induced IL-1β and pro-fibrotic gene expression, which the authors suggested was responsible for the reduced hepatic fibrosis in vivo. However this was not directly tested. Nevertheless, this study raised the possibility that targeting TPL-2 may forestall the progression of hepatic fibrosis. Indeed many small molecule inhibitors have been developed that block TPL-2 signalling in vitro [2], but none have yet made it in to the clinic. However, it is not known whether TPL-2 contributes to chronic type-2 inflammation and IL-4R-mediated fibrosis. In this study, we used the well-established Schistosoma mansoni infection model to test whether TPL-2 regulated chronic type-2 associated inflammation, immunopathology and fibrosis. In contrast to the reduced fibrosis observed in Map3k8–/–mice following chemical and diet-induced fibrosis [32], Map3k8–/–mice had significantly increased type-2 immune responses with concomitant elevated inflammation and fibrosis surrounding trapped parasite eggs. Using genome-wide transcriptional analysis and metabolic assays we found that TPL-2 was required for lipid oxidative metabolism and M2-MΦ activation. Specifically, TPL-2 was required for expression of immunoregulatory molecules (Retnla and Arg1) and regulated pro-fibrotic genes (Col genes and Ctgf). Consequently, myeloid cell-specific deletion of Map3k8 resulted in increased type-2 inflammation and significantly increased fibrosis in vivo, phenocopying Map3k8–/–mice. Collectively, our study identifies a novel and previously unappreciated role for TPL-2 as a molecular regulator of lipolysis in M2-MΦ’s, regulating type-2 inflammation, immunopathology and hepatic fibrosis. Following maturation and worm pairing, gravid worms release hundreds of eggs, many of which traverse the wall of the intestine and are released into the environment via the fecal route. However, many eggs do not successfully reach the intestinal lumen but instead become trapped in the intestinal wall or within vascularised organs, particularly the liver. An eosinophil and MΦ-rich fibrotic granuloma forms around trapped eggs causing significant tissue damage, orchestrated by CD4+ TH2 cells and a highly polarised type-2 immune response [5]. To test whether TPL-2 contributed to S. mansoni-associated intestinal and hepatic pathology and fibrosis, we infected Map3k8–/–mice with 50 S. mansoni cercariae. Histological analysis indicated that S. mansoni-infected Map3k8–/–mice had more fibrosis in the liver with larger hepatic granulomas (Fig 1A and 1B), despite a similar egg burden (S1 Fig) and serum LPS level as S. mansoni-infected WT mice (S1 Fig). Similarly, intestinal inflammation was also significantly increased in Map3k8–/–mice (Fig 1C and 1D). Consistent with the increased collagen staining observed in Map3k8–/–mice, collagen-synthesising genes, Col3 and Col6, were both significantly elevated in the liver and small intestine of Map3k8–/–mice, compared to WT controls (Fig 1E), with significantly more hydroxyproline in the liver of Map3k8–/–mice (Fig 1F). Map3k8–/–mice had elevated expression of Il13 in the liver, but not Il1b, Tgfb, Il17a, Ifng, Tnfa or Il6 (S1 Fig), suggesting that IL-13-driven fibrosis was exacerbated in Map3k8–/–mice [33] rather the development of other inflammatory mechanisms of fibrosis [30]. CD4+ TH2 cell-derived IL-4 and IL-13 are essential for granuloma formation [6], mobilising and activating a suite of innate immune cells, including MΦ’s and eosinophils, and promoting local collagen deposition. TH2 cell-mediated inflammatory responses are controlled by Foxp3+ regulatory T (TREG) cells [34], which restrain TH2 cell expansion. It was previously suggested that T cell intrinsic TPL-2 regulates TH2 [35] and Foxp3+ TREG cell differentiation [36]. However, these conclusions were based on in vitro experiments and were not tested in vivo. To determine whether Map3k8–/–mice had dysregulated TH2 and Foxp3+ TREG responses following S. mansoni infection, we crossed Map3k8–/–mice with Il4gfp and Foxp3rfp reporter mice, generating dual-reporter Map3k8–/–mice (Map3k8–/–Foxp3rfpIl4gfp). These reporter mice allowed us to accurately and simultaneously monitor TH2 (Il4gfp+) and Foxp3+ TREG (Foxp3rfp+) cells in Map3k8–/–mice without the requirement for re-stimulation or intra-nuclear staining. Map3k8-deficiency did not alter CD4+CD25+Foxp3rfp+ TREG cell frequencies in the spleen, mesenteric lymph node (MLN) or in the local liver tissue, indicating that TPL-2 was not required for TREG cell development or recruitment following S. mansoni infection (Fig 1G, top row). However, CD4+CD44+Il4gfp+ TH2 cells in both lymphoid tissues and the liver were significantly increased in Map3k8–/–mice compared to WT mice (Fig 1G, middle row). Map3k8-deficiency also increased the frequency of Il4gfp+Foxp3rfp+ cells in the MLN. Pharmacological inhibition of MEK1/2, a downstream target of TPL-2, protected mice from bleomycin induced fibrosis [31]. We have previously reported that bleomycin-induced fibrosis is mediated by a pro-inflammatory type-1/type-17 and TGFβ driven response, distinct from type-2 mediated pulmonary fibrosis[30]. It therefore remained unclear whether TPL-2 contributed to type-2 driven pulmonary fibrosis. To test this we treated mice intravenously with S. mansoni eggs to invoke type-2 inflammation in the lungs leading to the development of pulmonary fibrosis, as previously described [30]. Similar to responses in the liver, Map3k8–/–mice had increased collagen staining in the lung and increased hydroxyproline levels, compared to WT mice given S. mansoni eggs (S2 Fig). In the lung tissue and local draining thoracic lymph nodes (TLN), Map3k8–/–mice had increased Th2 cell frequency (S2 Fig) promoting increased Il13, Col6 and Mmp12 expression in the lung (S2 Fig). Collectively, these data indicate that TPL-2 is an important negative regulator of type-2 inflammation, immunopathology and fibrosis following S. mansoni infection or S. mansoni egg induced pulmonary fibrosis in vivo. It has previously been reported that T cell-intrinsic TPL-2 regulates TH2 cell differentiation in vitro and acute type-2 inflammation in the airways [35], however it has remained unclear whether T cell-intrinsic TPL-2 regulates TH2 cell differentiation and function in vivo. To formally test whether T cell-intrinsic TPL-2 contributed to the enhanced inflammation and fibrosis observed in Map3k8–/–mice (Fig 1) we restricted Map3k8 deficiency to T cells using Cd4CreMap3k8fl/fl mice. Deletion of Map3k8 in T cells (Cd4CreMap3k8fl/fl) had no impact on granuloma development in the liver (Fig 2A and 2B) or small intestine (Fig 2C) following S. mansoni infection. Similarly, fibrosis (Fig 2A and 2C) and expression of collagen synthesising genes, Col3 and Col6, were not affected following the deletion of Map3k8 in CD4+ cells (Fig 2D). IL-5 and IL-10 production was significantly increased in re-stimulated MLN cells from Map3k8–/–mice, compared to WT cells; however production of these cytokines was not affected when Map3k8 was deleted in T cells only (Fig 2E). IL-17 production was low and unchanged between all groups, however IFNγ secretion from lymph node cells was reduced in Map3k8–/–mice and Cd4CreMap3k8fl/fl mice, in line with a previous report [18]. To further test whether T cell intrinsic TPL-2 was required for TH2 cell differentiation, we isolated naïve T cells (TCRβ+CD4+CD44_) from WT and Map3k8–/–mice and polarised them under TH1 or TH2 conditions in vitro. Similar frequencies of IFNγ+ or IL-4+ cells were observed between WT and Map3k8–/–T cells, respectively (Fig 2F), suggesting that T cell-intrinsic TPL-2 does not contribute to TH1 or TH2 differentiation in vitro. Taken together, these data indicate that T cell-intrinsic TPL-2 is required for optimal IFNγ secretion in vivo, but does not contribute to TH2 cell differentiation in vitro or in vivo, and that T cell-intrinsic TPL-2 does not contribute to TH2 cell-mediated immunopathology following S. mansoni infection. Alternatively activated macrophages (AA or M2-MΦ) contribute significantly to inflammation, immunopathology and fibrosis following S. mansoni infection [12]. TPL-2 has a well-defined role in classically activated MΦ’s (M1 or CA-MΦ) [17, 20–29], however it is unclear whether TPL-2 contributes to M2-MΦ following S. mansoni infection. Firstly, to test whether myeloid cell-intrinsic TPL-2 contributed to the exacerbated immunopathology observed in Map3k8–/–mice, we restricted Map3k8 deletion to Lysozyme M-expressing cells using LysMCreMap3k8fl/fl mice (S3 Fig). Mice with myeloid cell-specific deletion of Map3k8 had significantly more inflammation with larger hepatic (Fig 3A and 3C) and intestinal (Fig 3B) granulomas and more severe intestinal pathology (Fig 3D), without any appreciable change in serum LPS (S3 Fig). Of note, a distinct collagen-rich fibrotic ring surrounded hepatic granulomas in LysMCreMap3k8fl/fl mice, which was absent in mice with WT myeloid cells. Increased collagen staining in the liver was supported by increased expression of collagen-synthesising genes, Col3 and Col6 (Fig 3E) and increased hydroxyproline (Fig 3F). Similar to Map3k8–/–mice, mice with myeloid cell-specific deletion of Map3k8 had elevated type-2 cytokine secretions (IL-13, IL-5 and IL-10) following lymph node re-stimulation without any appreciable change in IFNγ or IL-17A secretion (Fig 3E). Similarly, elevated expression of Il13 but not Il1b, Tgfb, Il17a, Ifng, Tnfa or Il6 was observed in LysMCreMap3k8fl/fl mice, compared to control mice (S3 Fig). These data clearly indicated that macrophage/myeloid cell intrinsic-TPL-2 contributed significantly to the regulation of TH2-mediated inflammation and fibrosis following S. mansoni infection. TH2-cell derived IL-4 and IL-13 [6] activates IL-4 receptor (IL-4R)-expressing MΦ’s [7, 8] to prevent lethal pathology following S. mansoni infection. To determine whether myeloid cell-intrinsic TPL2 contributed to M2-MΦ’s activation in vivo, we isolated MΦ’s ex vivo from infected mice. Chimeric mice were generated following the observation that Map3k8–/–mice and LysMCreMap3k8fl/fl mice had significantly elevated type-2 inflammation and fibrosis, compared to WT controls (Figs 1 and 3). Generating 50:50 chimeric mice by reconstituting lethally irradiated WT mice with 50% bone marrow from CD45.1+WT mice and 50% bone marrow from CD45.2+ Map3k8–/–mice (Fig 4A), normalised and controlled for these environmental differences allowing us to more accurately compare WT and Map3k8–/–MΦ’s ex vivo. Following 8-weeks of S. mansoni infection, CD45.1+WT or CD45.2+ Map3k8–/–CD3−CD11b+F4/80+CD11b+ MΦ’s were FACS sorted for analysis (Fig 4A). Map3k8–/–MΦ’s had significantly lower expression of Arg1, Relma and Chi3l3, compared to WT MΦ’s isolated form the same tissue (Fig 4B). In addition, Map3k8–/–MΦ’s had elevated expression of collagen synthesising genes (Col1, Col3) and connective tissue growth factor (Ctgf). These data indicate that macrophage cell-intrinsic TPL2 was required for M2-MΦ activation and regulated expression of pro-fibrotic genes. To determine how TPL-2 was regulating M2-MΦ activation we used the well-described in vitro macrophage activation assay, activating bone marrow-derived MΦ’s (BMDM) with IL-4 and IL-13. Following 6hrs of exposure to IL-4 and IL-13, Arg1, Retnla, Chi3l3 and Ear11 were all significantly reduced in Map3k8–/– M2-MΦ’s, compared to WT M2-MΦ’s (Fig 5A–5D), similar to that observed in ex vivo MΦ’s. At 24hrs, Retnla, Chi3l3 and Ear11 were still significantly reduced in Map3k8–/– M2-MΦ’s, demonstrating a non-redundant role for TPL-2 in M2-MΦ activation. The early reduction of Arg1 expression in Map3k8–/– M2-MΦ’s led to a reduction in arginase activity with reduced ornithine production, as determined by LCMS (S4 Fig). Inhibition of the kinase activity of TPL-2, using the pharmacological inhibitor, C34 [37], phenocopied Map3k8–/–MΦ’s indicating that the kinase activity of TPL-2 was responsible for the reduced Retnla expression in M2-MΦ (S4 Fig). Phosphorylated (p)STAT6, pERK, pp38α and pJNK were similar in Map3k8–/–and WT MΦ’s (Fig 5E), suggesting that TPL-2 did not regulate responsiveness of MΦ’s to IL-4 and/or IL-13 and was not required for activation of these downstream transcription factors or kinases. To determine whether TPL-2 regulated the expression of other genes, beyond the characteristic M2-MΦ-associated genes, we profiled the transcriptional landscape of WT and Map3k8–/–MΦ’s following 24hrs of IL-4 and IL-13 stimulation (Fig 6A). Pathway analysis identified increased inflammatory pathways in Map3k8–/– M2-MΦ’s, including proliferation, migration and fibrogenesis (Fig 6B). Of the significantly differentially regulated genes (P<0.05, > 2-fold, relative to un-stimulated) (Fig 6C), we identified 351 Map3k8-dependent genes (i.e. genes differentially regulated in WT only, Fig 6C and 6D) and 279 Map3k8-regulated genes (i.e. genes differentially regulated in Map3k8–/–MΦ’s only, Fig 6C and 6F). Of note, several of these elevated genes in Map3k8–/– M2-MΦ’s contribute to fibrogenesis, including Adam19 [38], Cxcr4 [39], Mmp13 [40], Cav1 [41], Itgav [42] and Vcam1 [43] (Fig 6F). Of the commonly regulated genes in both WT and Map3k8–/– M2-MΦ’s (Fig 6E and 6G) Map3k8–/– M2-MΦ’s had elevated expression of collagen-synthesising genes (Col1a1, Col3a1, Col5a2) and the connective tissue growth factor, Ctgf (Fig 6H) compared to WT MΦ’s (Fig 6G). Concurrent with increased expression of pro-fibrotic genes, characteristic M2-MΦ genes (Retnla, Arg1, Ear11 (Rnase2) and Chi3l3) were reduced, compared to WT M2-MΦ’s (Fig 6H). Taken together, these in vitro gene expression data suggested that Map3k8–/– M2-MΦ’s had both elevated pro-fibrotic properties and reduced regulatory/inhibitory functions (Arg1 and Retnla (Figs 4 and 5)). Oxidative lipid metabolism is a metabolic programme recently reported to be essential for M2-MΦ activation [44]. It also has previously been reported that TPL-2, MEK 1/2 and ERK 1/2 [45–48] can regulate lipid metabolism in a variety of different cells. We therefore hypothesised that the compromised M2-MΦ activation of Map3k8-deficient MΦ’s was due to reduced lipid metabolism. To investigate this possibility, we analysed the expression of 220 genes involved in lipid metabolism from the transcriptional data obtained from WT and Map3k8–/– M2-MΦ ‘s (S1 Table) and identified 16 TPL-2-dependent genes involved in lipid metabolism that were up regulated in WT M2-MΦ’s but not in Map3k8–/– M2-MΦ’s (Fig 7A). These genes included the LDL receptor, Olr1, which is required for lipid uptake [49] and Adipoq encoding adiponectin, which promotes lipid oxidation [50] and the alternative activation of human MΦ’s [51]. In addition, Aldh1a2 (also referred to as Raldh2) which catalyses the synthesis of the lipid metabolism-promoting metabolite, retinoic acid, from retinaldehyde [52] and the NFκB-regulated sialyltransferase, St8sia1 [53], which catalyzes the transfer of sialic acid from CMP-sialic acid to GM3 to produce gangliosides, were reduced in Map3k8–/– M2-MΦ’s, compared to WT M2-MΦ’s. Several of these genes are downstream of TPL-2 (S5 Fig), supporting our hypothesis that TPL-2 regulates lipolysis in M2-MΦ’s. Together these changes in gene expression in Map3k8-deficient M2-MΦ’s were consistent with TPL-2 signalling regulating lipolysis. To formally test whether lipid metabolism was compromised in Map3k8–/– M2-MΦ’s, we used extracellular flux analysis and measured the oxygen consumption rate (OCR) and spare respiratory capacity (SRC, the quantitative difference between maximal uncontrolled OCR, and the initial basal OCR, indicative of commitment to oxidative phosphorylation) in un-stimulated and IL-4/IL-13 mediated M2-MΦ’s. At baseline, un-stimulated WT and Map3k8–/–MΦ’s had a similar OCR and SRC (S6 Fig). However, Map3k8–/– M2-MΦ’s had significantly reduced OCR and SRC (Fig 7B and 7C), indicating that TPL-2 is required for lipid metabolism in M2-MΦ’s, and providing a mechanistic explanation for reduced M2-MΦ’s in Map3k8–/–mice. Map3k8–/–mice and LysMCreMap3k8fl/fl mice, which had compromised M2-MΦ activation, had elevated Th2 cell responses. It has previously been reported that M2-MΦ’s can directly regulate T cell responses [13–15]. We therefore hypothesised that Map3k8–/– M2-MΦ’s would not regulate Th2 cell differentiation and proliferation as well as WT M2-MΦ’s. To test this hypothesis, we co-cultured WT or Map3k8–/–BMMΦ’s with naïve cell trace violet (CTV)-labelled CD4+CD44−Il4gfp–OTII+ T cells in the presence of IL-4, IL-13 and OVA and determined the proliferation (CTV dilution) and differentiation (Il4gfp expression) of T cells. After 3 days, 24% of T cells had proliferated and differentiated when co-cultured with WT MΦ’s (Fig 7D, top row middle panel). However, co-culture of T cells with Map3k8–/–MΦ’s led to significantly more Th2 cell differentiation and proliferation (~38%, Fig 7D, bottom row middle panel). Finally, to determine whether lipid metabolism contributed to MΦ-mediated regulation of Th2 cell proliferation and differentiation we pre-treated MΦ’s for 6 hours with Orlistat, an irreversible lipase inhibitor, prior to co-culture with T cells. Orlistat treated WT MΦ’s led to more Th2 cell proliferation and differentiation (~32%, Fig 7D, top row right panel), phenocopying Map3k8–/–MΦ’s and indicating that lipid metabolism in IL-4/IL-13 activated MΦ’s was required for optimal MΦ-mediated control of Th2 cell proliferation, as previously reported [44]. Of note, Orlistat treated Map3k8–/–MΦ’s only led to a small increase in Th2 cell proliferation, suggesting that lipid metabolism was already at a minimum in Map3k8–/–MΦ’s. Taken together this study has demonstrated that TPL-2 is a critical regulator of immune-mediated pathology and fibrosis following S. mansoni infection, functioning as an important metabolic regulator in M2-MΦ activation. Liver fibrosis and cirrhosis, which is responsible for over 1.5 million fatalities per year [54], can develop following a variety of infectious insults, including chronic infection [4]. Diseases characterized by persistent TH2 cytokine responses, such as chronic helminth infections, are associated with the development of significant tissue pathology and fibrotic scarring. Although the molecular pathogenesis of fibrotic diseases are slowly emerging [4], there are very few novel therapeutic candidates progressing through clinical trials [55], highlighting a significant unmet medical need. Two distinct inflammatory axes contribute to inflammation-driven fibrosis; type-1/ TH17 mediated inflammation [30] and type-2 driven fibrosis [56]. In this study we established that the Map3 kinase, TPL-2, is an important negative regulator of chronic type-2 inflammation-driven fibrosis following schistosome infection. These data are in contrast to a previous study testing the role of TPL-2 in three models of pro-inflammatory type-1/17-associated fibrosis (carbon tetrachloride-, methionine-choline-deficient diet- and bile duct ligation-induced fibrosis)[57, 58]. In two of these three models Map3k8–/–mice had significantly reduced fibrosis [32]. These seemingly contrasting results most likely reflect the different inflammatory events contributing to the fibrogenic response. For example, it has been widely reported that TPL-2 is required for pro-inflammatory type-1/17-associated inflammation and immunity [17–29]. It therefore stands to reason that TPL-2 would be required for pro-inflammatory type-1/17-associated fibrosis, as reported by Perugorria and colleagues [32]. In contrast, TPL-2 appears to function as a negative regulator of type-2 inflammation in the lung and liver (Fig 1, S2 Fig), with increased acute [35] and chronic type-2 inflammation in Map3k8–/–mice, as presented here. In this context, the exacerbated type-2 inflammatory response in Map3k8–/–mice resulted in increased fibrosis. If these animal models reflect human disease, focused strategies targeting TPL-2 would benefit from identifying a prognostic biomarker and treating patients with type-1/17-associated fibrosis, rather than patients with type-2-associated fibrosis. Inflammation-driven fibrosis involves a co-ordinated and often dysregulated wound healing response involving a variety of migratory leukocytes activating local stroma. TPL-2 is expressed in both leukocytes and local stroma and therefore identifying where TPL-2 was regulating the fibrogenic process was essential for us to identify how TPL-2 regulated fibrosis. It has been previously suggested that increased acute TH2 responses in Map3k8–/–mice was due to a T cell-intrinsic role for TPL-2 [35], however this was not tested in vivo. Similarly, increased intestinal inflammation and tumorigenesis in Map3k8–/–mice was attributed to a reduced frequency of Foxp3+ TREG cells, [36], however again this was not specifically tested in vivo. Using Map3k8–/–Il4gfpFoxp3rfp mice and re-stimulated local lymph nodes we identified that TPL-2 negatively regulated the differentiation, expansion and/or recruitment of TH2 cells, however this was not due to a T cell-intrinsic role for TPL-2, as mice with a T cell-intrinsic deletion of Map3k8 mounted similar TH2 responses as WT mice. Furthermore, exacerbated hepatic fibrosis observed in Map3k8–/–mice was not observed in mice with a T cell-specific deletion of Map3k8, indicating that T cell-intrinsic TPL-2 had no impact on type-2–mediated inflammation or fibrosis in these systems. TPL-2 has been extensively studied in TLR-mediated classical macrophage activation (CA, M1-MΦ) [20]. However, it was unclear whether TPL-2 contributes to M2-MΦ differentiation. M2-MΦ’s are central regulators of inflammation, wound-healing and fibrosis following schistosome infection [12]. Specifically, Arginase (Arg-1) in M2-MΦ’s catalyses the cleavage of arginine to ornithine and urea, depleting extracellular arginine and depriving local leukocytes of this essential amino acid. Consequently, M2-MΦ’s limit T cell proliferation, by starvation, in an Arg-1-dependent manner [14]. Similarly, production of Retnla by M2-MΦ’s can suppress T cell responses directly [13, 15]; highlighting two mechanisms by which M2-MΦ’s regulate TH cell responses. The reduced expression of these immunoregulatory molecules (Arg1 and Retnla) in Map3k8–/– M2-MΦ’s may therefore explain the elevated TH2 cell inflammation observed in Map3k8–/–mice. Indeed, Map3k8–/– M2-MΦ’s did not control TH2 cell differentiation or proliferation (Fig 7D). M2-MΦ’s also supply proline for collagen synthesis and produce a variety of pro-fibrotic factors to promote wound healing, collagen deposition and if dysregulated, fibrotic scarring [59]. We observed increased expression of pro-fibrotic genes (Col1a1, Col3a1, Col5a2 and Ctgf) in Map3k8–/– M2-MΦ’s, which may explain the elevated fibrosis observed in Map3k8–/–mice. The mechanism of TPL-2-mediated regulation of pro-fibrotic mediators is not completely understood. It was recently reported that TPL-2 regulates hepatocyte growth factor (HGF) production in fibroblasts [60], in part, by reducing sensitivity to TGFβ. In IL-13-dependent fibrosis, TGFβ can inhibit some pro-fibrotic pathways [33]. If Map3k8-deficient M2-MΦ’s also have decreased sensitivity to TGFβ, this may explain the increased expression of several pro-fibrotic mediators in Map3k8-deficient M2-MΦ’s, however this requires further study. Together, these two observations provide some explanations for the increased TH2 cell responses and exacerbated hepatic and pulmonary fibrosis observed in Map3k8–/–mice. Indeed, specific deletion of Map3k8 in LysM-expressing cells phenocopied Map3k8–/–mice, with increased TH2 inflammation and fibrosis following S. mansoni infection or S. mansoni egg injection, compared to WT mice. The LysMCre system appears to efficiently delete floxed alleles in tissue-resident MΦ’s, but not so well in newly recruited myeloid cells [61, 62]. The exacerbated immunopathology observed in LysMCreMap3k8fl/fl mice suggests that Map3k8 may have an important function in tissue-resident MΦ’s, rather than newly recruited myeloid cells that may not have efficiently deleted Map3k8. Exactly how TPL-2 regulates IL-4R-driven M2-MΦ activation is unclear. TPL-2, and downstream kinases, MEK 1/2 and ERK 1/2 are required for lipid metabolism [45–48], a metabolic programme recently identified to be essential for M2-MΦ [44]. Joining these two independent observations, we confirmed that TPL-2 was an essential regulator of lipid metabolism in M2-MΦ’s, with extracellular flux assays identifying reduced oxygen consumption and reduced spare respiratory capacity, compared to WT M2-MΦ’s. The decreased lipid metabolism in Map3k8–/– M2-MΦ’s provides an explanation for the decreased expression of immunoregulatory molecules [44] in Map3k8–/– M2-MΦ’s. In addition, Map3k8–/– M2-MΦ’s had reduced expression of several genes encoding enzymes involved in lipid uptake, lipid metabolism and synthesis (Fig 6), providing an explanation as to how lipid metabolism may be compromised in Map3k8–/– M2-MΦ’s. Thus, we speculate that following IL-4R-signalling and STAT-6-mediated activation of downstream gene products, TPL-2 is required for the necessary metabolic re-programming [44] for optimal M2-MΦ activation. A paralleled increase in pro-fibrotic factors in Map3k8–/– M2-MΦ’s identified a novel TPL-2-regulated pro-fibrotic axis in M2-MΦ’s. Whether this was also due to compromised metabolic programme in Map3k8–/– M2-MΦ’s is currently unclear. In vivo, the increased TH2-mediated inflammation, as a result of compromised immune regulation by Map3k8–/– M2-MΦ’s, may further exacerbate pro-fibrotic pathways, providing a severely dysregulated microenvironment. In summary, this study has identified that TPL-2 is an important metabolic regulator in M2-MΦ’s in vitro and that myeloid cell intrinsic TPL-2 critically controlled chronic type-2-mediated inflammation and fibrosis in vivo. All mice were bred and maintained under specific pathogen-free conditions at The Francis Crick Institute. Strains used included: WT C57BL/6, Map3k8–/–[20], OTII [63], Il4gfp [64], Foxp3rfp[65], Cd4CreMap3k8flflR26eYFP (B6.Cd4Cre [66] crossed with Map3k8flfl and B6.R26eYFP), LysMCreMap3k8flflR26eYFP (B6.LysMCre [67] crossed with Map3k8flfl and B6.R26eYFP). Mice were infected percutaneously via the tail with 50 cercariae of a Puerto Rican strain of S. mansoni (NMRI) obtained from Biomphalaria glabrata snails, kindly provided by Dr. Quentin Bickle, LSHTM. Infection intensity was determined following perfusion and granuloma size was determined from 10–20 individual granulomas per tissue sample, measured using Image J. Tissue pathology was analysed following Masson’s trichrome (Collagen, Blue; Nuclei, black/dark blue; Muscle, cytoplasm, Red) staining of 5μm sections from paraffin embedded samples. Intestinal pathology was determined using a comprehensive scoring system taking into account the level of infiltration, disruption and severity of the intestinal architecture [68]. Hydroxyproline content was quantified in liver tissue using a hydroxyproline assay kit according to the manufacturers recommendations (Cambridge Biosciences, UK). Tissue eggs were quantified by digesting a known weight of liver tissue with collagenase and liberase and isolating eggs on a discontinuous percoll gradient, as previously described [69]. For intravenous delivery of S. mansoni eggs, eggs were washed extensively in PBS, with 5000k delivered in 200μl of sterile PBS. Bone marrow cells were plated to a density of 5 x 106 cells per 90-mm bacterial Petri dish (Sterilin) in 10ml of DMEM/F-12 Dulbecco’s Modified Eagle Medium/Nutrient Mixture F-12 with GlutaMAX supplement (Gibco) supplemented with 10% FBS, antibiotics, 20% L-cell conditioned medium, L-Glutamine (1%), HEPES (1%), Sodium Pyruvate (1%) and β-mercaptoethanol (1%). After 4 days of culture, 10ml of additional medium was added and cells were cultured for a further 3 days. Non-adherent cells were washed away and the adherent cells were collected in 5ml PBS with 5% FBS and 2.5mM EDTA. For experiments the cells were re-plated in medium with 1% FBS without L-cell supplement and were incubated overnight before stimulation. Cells were stimulated with IL-4 and IL-13 (20ng/ml) (R&D systems) or LPS (100ng/ml) (Alexis Biochemicals). In some experiments, BMDM were generated from three individual mice and co-cultured with a pool of naïve OTII CD4+CD44−Il4gfp–T cells during stimulation with OVA peptide (323–339) (Invivogen). In some of these co-culture experiments, BMDM were pre-treated for 6 hours with Orlistat (100μM; Cayman), prior to co-culture. Liver tissue was perfused and the organ was collected in gentleMACS columns (Miltenyi Biotec) in incomplete RPMI 1640 (Gibco). The tissues were dissociated in incomplete RPMI 1640 with Liberase TL (0.5mg/ml) (Roche), Collagenase (4μg/ml) (Roche) and DNAse (7.5μg/ml) using the gentleMACS dissociator (Miltenyi Biotec). The partly digested tissue fragments were incubated at 37°C for 45 min, following which the tissues were completely dissociated. The cellular fraction was run through a 100μm filter and the cells were centrifuged at 50g for 3 min, to pellet the non-parenchymal cells and the supernatant fraction was centrifuged at 320g for 5min. The cellular fractions from both steps were pooled and collected in 1X HBSS Hanks balanced salt solution and mixed with OptiPrep (Sigma) to get a 17% w/v solution and overlayed with 1X GBSS Gey’s balanced salt solution. The samples were centrifuged at 400g for 15min at room temperature with no brakes and the enriched layer of cells were collected from the interface of the GBSS and the 17% solution. The cells were stained and FACS sorted as Live/ CD45+/CD3−/CD19−/NK1.1−/Ly6G−/SiglecF−/CD11b+/F4/80+. Day 7 bone marrow-derived monocytes (BMDM) were cultured at a density of 5x105 cells in an XF24 plate (Seahorse Bioscience) over night. On day 8 cells were stimulated with 20ng/ml of recombinant IL-4 and IL-13 (R&D systems). On day 9 media was replaced with XF base medium (Seahorse Bioscience) supplemented with 100x Glutamax (Gibco), 100X Sodium Pyruvate (SIGMA) and 25mM glucose and the plate incubated for 10–30 minutes in a non-CO2 incubator at 37°C. For analysis of basal oxygen consumption rate (OCR) and extracellular acidification rate (ECAR), cells were analysed with XF-24 Extracellular Flux Analyzer (Seahorse Bioscience) under standard Seahorse running protocol. The Seahorse cartridge was loaded at a final concentration of 10μM Oligomycin (SIGMA O4876-5mg), 15μM FCCP (triflourocarbonylcyanide phenylhydrazone, SIGMA C2920-10mg), 1μM Rotenone (SIGMA R8875) plus 10μM Antimycin A (A8674-25mg) in ports A, B and C respectively. The bioenergetics profile consisted of basal OCR measurements in the absence of drugs and OCR/ECAR following the injection of drugs. All OCR/ECAR/SRC analyses were obtained from 5 replicates in 3 independent repeats. Naive CD4+ T cells (CD4+TCRβ+CD44−CD25−PI−) were FACS-purified from spleens of WT or Map3k8–/–mice. Naive T cells were cultured for 6 days in vitro with 10ng/mL IL-4 (R&D), 5ng/mL IL-2 (R&D), 10 μg/mL anti-IFNγ (XMG1.2, BioXcell), and CD3 (0.1–4.0μg) (145-2C11, BioXcell) and CD28 (10μg/ml) (37.51, BioXcell) in complete IMDM (cIMDM, 10% fetal calf serum (FCS),100 U/mL Penicillin and 100 μg/mL Streptomycin (Gibco), 8mM L-glutamine (Gibco), and 0.05mM 2-mercaptoethanol (Gibco)). For re-stimulation of lymph node cells ex-vivo, lymph nodes were disrupted into a single cell suspension with 2x105 cells cultured in a 96-well, round bottom plate with 10μg/ml of anti-CD3 (145-2C11, BioXcell). Supernatants were harvested after 3 days for analysis by ELISA. Cell sorting was performed using a FACS Aria II (BD Biosciences) cell sorter. For sorting, cell suspensions were stained for 20 minutes with antibodies in PBS with 2% fetal calf serum (FCS) and then diluted in phenol-red free IMDM (Gibco) (with 1% FCS, 2mM EDTA (Invitrogen), 100U/mL Penicillin and 100μg/mL Streptomycin (Gibco), 8mM L-glutamine (Gibco), and 0.05mM 2-mercaptoethanol (Gibco)). Propidium iodide (PI) or LIVE/DEAD fixable blue dead cell stain (Life Technologies) was used to determine cell viability. Cells were stained for surface antigens by incubation with antibodies in PBS with 2% FCS (20 minutes at 4°C). Intracellular cytokine staining was performed following 6 hours re-stimulation with 50ng/mL phorbol 12-myristate 13-acetate (PMA, Promega) and 1μg/mL ionomycin (Sigma) and BD Golgi Stop and BD Golgi Plug (diluted 1:1000, BD Biosciences). After staining for surface antigens, cells were fixed and permeabilized (Fixation/Permeabilization diluent; eBioscience), prior to incubation with cytokine antibodies in Permeabilization buffer (eBioscience) for 20 min at 4°C. Cells were analyzed using a BD LSRII flow cytometer (BD Biosciences) and data processed using FlowJo software (Version X 10.0.7r2, Treestar Inc). Antibodies used were purchased with eBioscience, Biolegend or BD Pharmingen. They included: CD45 (30-F11), CD3 (17A2), CD4 (RM4-5, GK1.5), CD11b (M1/70), CD19 (6D5, eBio1D3), CD25 (PC61), CD44 (IM7), NK1.1 (PK136), Ly6G (1A8), SiglecF (E50-2440), and F4/80 (BM8), TCR β chain (H57-597). Staining was performed in the presence of FcR Blocking Reagent (Miltenyi Biotec). RNA was isolated from tissues and cells using RNAeasy mini spin columns according to manufacturers’ instructions (Qiagen). cDNA was generated from 5ng of total RNA using WT-Ovation Pico system (version 1) RNA Amplification System followed by double stranded cDNA synthesis using WT-Ovation Exon Module. cDNA quality was determined using an Agilent BioAnalyzer and through hybridization performance on Affymetrix GeneChip mouse Genome 430A 2.0 microarray (Affymetrix) by the Systems Biology Unit at The Francis Crick Institute. Microarray data were quantile-normalized and analysed using GeneSpring software (Agilent). Differentially expressed genes were determined using ANOVA and t-tests. Genes with false discovery rate corrected p-values less than 0.1 and fold change values ≥1.5 were considered significant, and as indicated in figure legends. Three biological replicates of each subset were used. Pathways analysis was performed using Ingenuity Pathways Analysis (IPA, Ingenuity Systems, www.ingenuity.com). Data sets were compared by Mann Whitney test using GraphPad Prism (V.5.0). Differences were considered significant at *p ≤ 0.05 using one or two-tailed tests. All animal experiments were carried out following UK Home Office regulations (project license 80/2506) and were approved by The Francis Crick Institute Ethical Review Panel.

10.1371/journal.ppat.0040003

Yersinia Controls Type III Effector Delivery into Host Cells by Modulating Rho Activity

Yersinia pseudotuberculosis binds to β1 integrin receptors, and uses the type III secretion proteins YopB and YopD to introduce pores and to translocate Yop effectors directly into host cells. Y. pseudotuberculosis lacking effectors that inhibit Rho GTPases, YopE and YopT, have high pore forming activity. Here, we present evidence that Y. pseudotuberculosis selectively modulates Rho activity to induce cellular changes that control pore formation and effector translocation. Inhibition of actin polymerization decreased pore formation and YopE translocation in HeLa cells infected with Y. pseudotuberculosis. Inactivation of Rho, Rac, and Cdc42 by treatment with Clostridium difficile toxin B inhibited pore formation and YopE translocation in infected HeLa cells. Expression of a dominant negative form of Rac did not reduce the uptake of membrane impermeable dyes in HeLa cells infected with a pore forming strain YopEHJT−. Similarly, the Rac inhibitor NSC23766 did not decrease pore formation or translocation, although it efficiently hindered Rac-dependent bacterial uptake. In contrast, C. botulinum C3 potently reduced pore formation and translocation, implicating Rho A, B, and/or C in the control of the Yop delivery. An invasin mutant (Y. pseudotuberculosis invD911E) that binds to β1 integrins, but inefficiently transduces signals through the receptors, was defective for YopE translocation. Interfering with the β1 integrin signaling pathway, by inhibiting Src kinase activity, negatively affected YopE translocation. Additionally, Y. pseudotuberculosis infection activated Rho by a mechanism that was dependent on YopB and on high affinity bacteria interaction with β1 integrin receptors. We propose that Rho activation, mediated by signals triggered by the YopB/YopD translocon and from engagement of β1 integrin receptors, stimulates actin polymerization and activates the translocation process, and that once the Yops are translocated, the action of YopE or YopT terminate delivery of Yops and prevents pore formation.

The type III secretion system (TTSS) is essential for the virulence of a number of Gram-negative human pathogens of enormous clinical significance. The molecular mechanisms by which TTSS effector proteins are translocated into the host cell are not well understood. The work presented here proposes a new model in which the enteropathogen Yersinia pseudotuberculosis manipulates the host cell machinery to control effector translocation. This involves activation of the host cell Rho GTPase by the cooperative action of adhesin-mediated high affinity binding to specific cell receptor molecules known as β1 integrins, and interaction of components of the TTSS with the host cell membrane. This molecular mechanism of controlling TTSS may not be restricted to Y. pseudotuberculosis and might take place during infection of host cells with other pathogens that encode homologues of Yersinia TTSS proteins. Our findings provide a good starting point to study the molecular nature of the complex interaction between bacterial pathogens bearing TTSSs and the host cell. Importantly, components that act by modulating the TTSS are potential targets for novel antimicrobials.

A great spectrum of Gram-negative bacteria depends on a specialized secretion mechanism to establish a successful infection in the host. This machinery is known as the type III secretion system (TTSS), and is present in organisms that are pathogenic for animals or plants, as well as in symbiotic bacteria [1]. In pathogenic Yersinia species, a TTSS is encoded in a large virulence plasmid, and is required for counteracting innate and adaptive host immune defenses [2]. This is accomplished by injection of six effector proteins (YopE, YopT, YopH, YopJ, YopO, YopM) that target different host cell signaling molecules. This injection mechanism is known as Yop translocation. Two effectors relevant to this work are YopE and YopT, which target a family of Rho GTPases that control a variety of cellular functions, including regulation of the actin cytoskeleton. In turn, the activity of the Rho GTPases is tightly controlled by a number of regulators. Guanine nucleotide exchange factors (GEFs) induce activation of GTPases by inducing GDP/GTP exchange. GTPase accelerating proteins (GAPs) inactivate Rho GTPases by stimulating GTP hydrolysis. Active Rho proteins are mostly associated with cellular membranes by means of a post-translational lipid modification (prenylation) [3]. YopE inhibits RhoGTPases by acting as a GAP for RhoA, Rac1, or Cdc42 [4,5]. YopT inhibits preferably RhoA, by cleaving the isoprenyl group and removing the GTPase from the membrane [6]. Although the mechanism of translocation is not completely understood, it is thought that effectors are delivered from the bacterial cytoplasm to the outer membrane through a secretion conduit. In turn, this channel is connected to a needle–like structure that transports the effectors directly into the host cell's cytoplasm. Apart from the proteins that form the needle, three translocator proteins (YopB, YopD and LcrV) are required for the delivery of toxins into the host cell. YopB and YopD are thought to form a translocation channel at the plasma membrane [7–9]. Two recent report show that LcrV is located at the tip of the needle [10], and that it may act as an assembly platform for YopB and YopD prior to their insertion in the membrane [11]. Activation upon contact of the bacteria with the host cell is one of the hallmarks of the TTSS. Adhesion of Yersinia to host cells is mediated by surface proteins, such as invasin or YadA binding to β1 integrin host cell receptors, or by pH6 antigen interacting with glycosphingolipids [12,13]. High affinity interaction of β1 integrin receptor with invasin, or YadA (via fibronectin), stimulates a signal transduction pathway that involves activation of Src protein tyrosine kinase, tyrosine phosphorylation of focal adhesion proteins, such as FAK and Cas, and downstream activation of Rac1 and PI3-K [12,14,15]. Stimulation of this pathway results in bacterial internalization. We have previously shown that infection of epithelial cells with Y. pseudotuberculosis lacking YopE, YopT, YopJ and YopH elicits a proinflammatory signaling response that requires YopB but is independent of YopD, suggesting that this signaling event can occur in the absence of a translocation channel [16]. This proinflammatory response, characterized by activation of MAP kinases and NFκB, and production of IL-8, is blocked by the Rho GTPase inhibitory action of YopE, and to a lesser extent YopT [17]. It is therefore possible that YopB elicits activation of a signaling pathway involving Rho GTPases. Although a translocation channel composed of YopB and YopD is thought to insert into the host cell membrane, the integrity of the plasma membrane remains intact during infection with wild type Yersinia. However, infection with Yersinia mutant strains that do not produce YopE and YopT results in loss of membrane integrity, a process known as pore formation [7,18]. Interestingly, yopE,yopT mutants also induce the polymerization of an actin ring at the site of the interaction with the host cell, but the link between these “actin halos” and pore formation is not known. How YopE or YopT prevent pore formation is not fully understood, and is a controversial issue [19]. We have found that, catalytically inactive forms of YopE or YopT ([18], unpublished data) were not able to prevent pore formation, analyzed by uptake of impermeable dyes (EtdBr) or release of lactate dehydrogenase (LDH). Expression of constitutively active forms of RhoA or Rac1 prior to infection, rescued the pore forming activity of bacteria expressing YopE or YopT [18]. In addition, infection carried out in the presence of actin polymerization inhibitors dramatically reduced pore formation. Based on these results we concluded that insertion of the YopB/D translocation channel results in Rho GTPases activation, actin polymerization, and pore formation [18]. Here, we present evidence that not only pore formation but most importantly, translocation is controlled by Rho activity and actin polymerization. We also found that high affinity interaction between YadA or invasin with β1 integrin receptors is crucial for efficient translocation of Yops. Thus, we hypothesize that YopB/D signaling, in cooperation with β1 integrin signaling, activates Rho to induce changes in the host cell cytoskeleton that control the translocation process. Macrophages infected with Salmonella or Shigella species undergo a caspase-1-dependent form of cell death termed pyroptosis [20]. This death mechanism is proinflammatory, and requires Yersinia YopB homologues SipB and IpaB, from Salmonella and Shigella, respectively. A recent report shows that pyroptosis is caused by caspase-1-dependent pore formation and consequent osmotic lysis [21]. Pore formation is usually determined by the incorporation or release of membrane impermeable dyes, such as EtdBr and BCECF, respectively, by the infected cells [7,8,22]. Because pore formation is followed by osmotic lysis, an indirect method to determine pore formation involves measuring the release of the cytoplasmic enzyme lactate dehydrogenase (LDH) in supernatants of cultured cells [22]. In Yersinia-infected macrophages, caspase-1-mediated maturation and release of the proinflammatory cytokine interleukin 1β can be inhibited by YopE and YopT [23]. Because the inhibitory action of YopE and YopT on the Rho GTPases also blocks pore formation [18], we investigated whether YopB/D-mediated cell lysis in HeLa cells is a result of caspase-1 mediated cell death. We used Ac-YVAD-cmk (YVAD), a permeable peptide that specifically inhibits caspase-1, irreversibly. HeLa cells treated for 1 h with 50 μM or 100 μM of YVAD, or control untreated cells, were infected with pore forming strain yopEHJ (YP27), and the corresponding pore forming-deficient strain that lacks YopB (yopEHJB, YP29). The uptake of the impermeable dye ethidium homodimer-2 (EthD2) and the amount of LDH released in the supernatant of infected cells was tested 3 hours after infection. YVAD did not prevent LDH release (Figure 1A) or penetration of the dye (not shown) in cells infected with YP27. On the other hand, YVAD treatment dramatically inhibited YP27-induced IL-1β production in J774.1A macrophage-like cells (Figure S1), indicating that 100μM YVAD efficiently inhibits caspase-1 mediated processes. These data support the hypothesis that YopB/D-mediated loss of membrane integrity in epithelial cells does not require caspase-1 activation. Salmonella–induced pyroptosis is also inhibited by 5 mM glycine [20]. We investigated if YopB/D-induced loss of membrane integrity could be inhibited by treatment with 5 mM glycine through out the infection. As shown in Figure 1B, glycine had no effect on the amount of LDH released by YP27-infected cells. This result further suggests that in HeLa cells YopB/D-mediated LDH release occurs by a process different from pyroptosis. We therefore consider that, in our experimental system, pore formation is linked to the translocation process. We have previously found that pore formation is prevented by the catalytic activity of two Rho GTPase-inhibiting effectors, YopE and YopT [18]. To test whether inactivation of small GTPases inhibits pore formation, we incubated cells for 2 h in the presence or absence of 40ng/ml of Clostridium difficile toxin B (ToxB), an ADP-ribosylating protein that powerfully inhibits Rho, Rac and Cdc42. ToxB treatment strongly reduced the uptake of ethidium homodimer-2 (EthD-2) by cells infected with pore forming strain yopEHJ (YP27) (Figure 2A). Rho GTPase downregulation by ToxB also inhibited LDH release (Figure 2B). Thus supernatants of YP27-infected cells treated with ToxB released levels of LDH comparable to those of cells infected with the pore-forming-deficient strain yopEHJB (YP29). These data suggest that YopB/D-mediated pore formation requires activation of Rho GTPases. We have previously observed that a catalytically inactive form of YopE (YopER144A) is translocated at higher levels than wild type YopE [4,18]. Aili et al. have also reported this phenomenon recently; they showed that several YopE mutants defective for GAP activity are hypertranslocated [24,25] . Interestingly, Wong and Isberg [26] observed that overexpression of YopT inhibits YopE translocation. Altogether, these observations suggest a possible role of GTPase activation in controlling the translocation process. To study this hypothesis we tested the action of ToxB on YopE translocation using the Triton X-100 solubility assay described in Material and Methods. Pretreatment of HeLa cells with ToxB reduced the amount of YopE translocated by wild type strain YP126 by 60% (Figure 2C). As expected, only background levels of YopE were detected in the soluble fraction of cells infected with the translocation deficient YopB− mutant, YP18. The inhibitory effect of ToxB on pore formation and translocation is not likely to be a consequence of an impairment of the bacteria-host cell interaction, because the number of cell-associated bacteria did not vary with ToxB treatment (Figure S2). This led us to conclude that Yop translocation is strongly influenced by the level of Rho, Rac or Cdc42 activation. In previous experiments, we have shown that actin polymerization inhibitors, cytochalasin D (CD) and latrunculin B, inhibit pore formation [18]. Here we confirmed the effect of CD on pore formation (Figures 2A and 2B), and determined whether host actin polymerization plays a role in Yop translocation during Yersinia infection. As shown in Figure 2C, CD treatment greatly reduced the amount of translocated YopE, this inhibitory effect being comparable to the ToxB treatment. Adhesion assays showed that CD does not affect the number of cell-associated bacteria greatly (not shown). These observations appear to indicate that actin polymerization is not only required for pore formation, as we had shown previously, but it also controls Yop translocation. Y. pseudotuberculosis internalization into epithelial cells requires a signaling cascade that results from the binding of invasin or YadA to β1 integrin receptors. Bacterial uptake requires small Rho GTPases activation and actin polymerization. Thus internalization, like pore formation and translocation, is inhibited by the GAP activity of YopE, and by treatment with cytochalasin D [4,18]. With this in mind, we investigated whether invasin or YadA-mediated adhesion to β1 integrin receptors is required for efficient pore formation and translocation. We created a yopEHJ,yadA,inv mutant strain, designated YP50, and the corresponding YopB-deficient mutant YP51 (Table 1). To provide a means of adhesion, a pAY66 plasmid, constitutively expressing pH6 antigen (Table 1), was inserted into YP50 and YP51. The pH6 antigen is a fimbrial adhesin that can mediate adhesion of Yersina to epithelial cells but does not induce bacterial uptake [27]. The defect in internalization of YP50/pAY66 and YP51/pAY66 was confirmed by immunofluorescence (not shown, see below). To corroborate that pH6 ag can substitute invasin or YadA for adherence, we evaluated the binding ability of the YP50/pAY66 strain after one hour infection by immunofluorescence. We found that YP50/pAY66 adhered to HeLa cells at levels similar to yopEHJ (YP27) expressing invasin or YadA (not shown). YP50/pAY66 strain was compared to the YP27 strain for the ability to induce pore formation. Surprisingly, YP50/pAY66 caused lower levels of LDH release than YP27 (Figure 3A), and was defective for promoting uptake of EthD-2 by infected HeLa cells (not shown). As expected, infection with the corresponding yopB mutant YP51/pAY66 resulted in even lower levels of LDH release. Ectopic expression of YadA in the YP50 strain rescued LDH release, indicating that interaction with β1 integrin receptors is critical for pore formation. To rule out that the impairment of the inv,yadA, pH6 antigen-expressing mutant to cause pore formation was due to a defective activation of the TTSS, we tested the ability of YP50/pAY66 to induce IL-8 production, NFκB activation, and ERK phosphorylation. We have previously found that the ability to stimulate these pro-inflammatory signals requires YopB but is independent of pore formation [16]. As shown in Figure 3B, after 5 hours infection, IL-8 production was not considerably reduced by the absence of invasin or YadA. Similarly, YopB-dependent activation of NFκB and ERK, measured at 1 hour-post infection, did not require invasin or YadA (Figure S3), suggesting that YopB is able to stimulate cell responses whether adhesion is provided by invasin/YadA or by pH6 antigen. Collectively, these results indicate that interaction of the bacteria with β1 integrin receptors is required to stimulate pore formation. To investigate whether engagement of β1 integrin receptors is also needed for the translocation process, we tested the ability of a yadA,inv mutant to translocate YopE. To this end, we replaced the mutated yopE by the wild type yopE gene in YP50/pAY66, creating YP54/pAY66 (Table 1). In line with its reduced ability to cause pore formation, the inv,yadA, pH6 antigen-expressing mutant translocated undetectable levels of YopE (Figure 3C). Consequent with these findings, YP54/pAY66 induced cell rounding at a much slower rate than the wild type YP126 (Figure 3D, compare YP126 and YP54/pAY66 after 30 min infection). Efficient YopE translocation was restored when YadA was expressed in YP54 (Figure 3C). This suggests that the interaction of Y. pseudotuberculosis with β1 integrin receptors is required for an effective translocation process. As invasin and YadA promote both binding to β1 integrins and stimulation of signaling by this receptor, we used a mutant that is competent for binding to β1 integrins but defective in signaling, to establish which activity was important for pore formation and translocation. A single amino acid substitution, D911E, in the invasin protein retains binding to host cells, but results in low affinity interaction with β1 integrins, poor receptor clustering, and a consequent defect in signaling and internalization [28]. Thus, we assessed the ability of YP50invD911E and YP54invD911E to induce pore formation and to mediate YopE translocation, respectively. Although infection with YP50invD911E resulted in robust IL-8 production (Figure 3B), the levels of LDH release by cells infected with YP50invD911E were as low as those cells infected with the strains that adhere via pH6 antigen (Figure 3A). Similarly, YP54invD911E was impaired in YopE translocation (Figures 3C and 3D). We ruled out that the defect in translocation was a consequence of fewer YP54invD911E bacteria binding to Hela cells. Thus, immunofluorescence analyses after 1-hour infection revealed that YP50invD911E infected cells had a mean of 16.6 associated bacteria/cell, only slightly lower than the invasin-expressing strain (19.7 bacteria/cell, Figure S4A). Moreover, a two-fold increase in the multiplicity of infection of YP54/pAY66 and YP54invD911E did not result in higher levels of YopE translocation (Figure S4B). We conclude that efficient translocation and pore formation involve high affinity binding to β1 integrin receptors. To examine the binding characteristics of the inv/yadA mutant we performed transmission electron microscopy in thin section of infected HeLa cells. As expected, yopEHJ (YP27) bacteria were either internalized, or were in the process of being engulfed, and tightly attached to the host cells (Figure S5A). On the other hand, yopEHJ,yadA,inv/psaABC (YP50/pAY66) were almost exclusively extracellular and seemed to bind more loosely (Figure S5B). Adhesion mediated by invD911E differed from that imparted by wild type invasin (Figure S5A and S5C). This suggests that lack of high affinity binding to β1 integrin receptors not only impairs β1 integrin signaling, but might also affect the way the bacteria interacts with the host cell. Stimulation of signaling through β1 integrins receptor by invasin and YadA involves tyrosine phosphorylation of a series of signaling proteins. Src is a key signal-transducing protein kinase in the β1 signaling pathway leading to internalization. To determine if Src activation plays a role in Yop translocation, we tested the effect of a selective inhibitor of Src family kinases, PP2, on infected cells. Pre-treatment of cells for 1 hour with 10μM PP2 efficiently inhibited β1 integrin signaling pathway leading to bacterial internalization without decreasing bacterial adherence (not shown). Interestingly, pore formation and YopE translocation were also impaired in PP2-treated cells (Figures 4A and B). These data indicate that Src activation stimulates translocation, and point toward a role of β1 integrin signaling in the Yop translocation. Invasin triggered-Rac1 signaling pathways downstream of Tyr phosphorylation are essential for Yersinia uptake [15]. We made use of a specific Rac1 inhibitor to determine whether β1 integrin–mediated internalization was required for efficient pore formation and translocation. NSC23766 is a small chemical compound reported to specifically block the binding between Rac1 and its exclusive GEFs [29]. We tested the effect of the Rac1 inhibitor by pre-treating HeLa cells for 6h with 100μM of NSC23766 in DMEM with 5% serum. As expected, bacterial uptake was impaired by treatment with the Rac inhibitor, with the number of yopEHJ (YP27) bacteria internalized by NSC23766-treated cells being comparable to that of the uptake-deficient yopEHJ,yadA,inv (YP50/pAY66) strain (Figure 5A). NSC23766 treatment was also found to inhibit formation of phagosomes, as the number of actin cups was reduced more than 5 fold in the presence of the inhibitor (Figure S6). We further excluded any effect of NSC23766 treatment on the number of cell-associated bacteria by immunofluorescence (not shown). Transmission electron microscopy of thin sections also confirmed that NSC23766 inhibited bacterial uptake by, but not association to HeLa cells (Figure S5D). Importantly, treatment with NSC23766 did not reduce pore formation or Yop translocation (Figures 5B and 5C, respectively). These results indicate that bacterial internalization is not required for pore formation or translocation. To validate our findings using the Rac1 inhibitor, we expressed a dominant negative form of Rac1 in Hela cells. We transfected cells with a eukaryotic expression plasmid coding for a T7 tagged-RacN17 (pCGTRacN17) and we evaluated whether pore formation was impaired in transfected cells. Overexpression of Rac1N17 (green cells) did not prevent pore formation as shown by the uptake of the impermeable dye EthD-2 (Figures S7A and S7B). Altogether, these data provide evidence indicating that neither bacterial internalization, nor Rac1 activation, play a major role in the processes that govern pore formation and Yop translocation. C3 is an ADP-ribosylating protein of Clostridium botulinum that specifically inhibits Rho A, B and C. A recombinant cell-permeable form of C3 toxin (TAT-C3) was produced in E. coli and purified as described in Material and Methods. Four hours before infection, HeLa cells were treated with 10, 20, and 40μg/ml of TAT-C3 in serum-free medium, or with serum-free medium alone. C3 has been previously shown to increase Y. pseudotuberculosis uptake in COS-1 cells [30]; in our experimental model, pretreatment of cells with 20μg/ml TAT-C3 did not affect bacterial adhesion or internalization considerably (Figures S8A and S8B, respectively). Interestingly, TAT-C3 treatment of cells infected with the pore forming strain yopEHJ (YP27) inhibited LDH release in a dose-dependent manner (Figure 6A). The effect of Rho inhibition on translocation was also substantial (Figure 6B). In various experiments, treatment with different batches of purified TAT-C3 (40μg/ml) reduced YopE delivery into wild type-infected cells, by 40% to 75 %. Similar results were obtained when we tested the effect of C3 treatment on YopH translocation (Figure 6B), indicating that the requirement of Rho for translocation is not a phenomenon restricted to YopE delivery. To test whether actin polymerization required for pore formation and translocation was dependent on Rho, we analyzed the effect of C3 on the induction of actin polymerization around the bacteria [18]. We found that the number of YopB-dependent actin halos was considerably reduced in the presence of C3 (Figure 6C). To determine whether Rho is activated by infection with Y. pseudotuberculosis, we infected HeLa cells with strain yopEHJ (YP27) for 5, 10, 15 and 20 min and we analyzed the amount of active Rho (GTP-Rho) in the cell lysates by a GTP-Rho pull-down assay, as described in Material and Methods. A peak of Rho activation was detected between 10 and 15 min after infection (Figure 7A). A 15 min infection period was selected to test the levels of GTP-Rho induced by infection with yopEHJ (YP27), yopEHJB (YP29), yopEHJ,yadA,invD911E (YP50/ invD911E), and yopEHJB,yadA,invD911E (YP51/invD911E). Compared to YP27-infected cells, cells infected with YP29 have reduced amounts of GTP-Rho, indicating that Rho activation is YopB-dependent (Figure 7B). Low affinity interaction with β1 integrin receptors by infection with YP50/ invD911E cause a reduced activation of Rho. However, YopB-independent Rho activation in YP29-infected cell lysates was greater than that of cells infected with YP51invD911E. This small difference, attributed to wild type invasin or YadA interacting with β1 integrin receptors, was consistent in three independent experiments. Overall, these experiments lead us to conclude that Y. pseudotuberculosis activates Rho by a process that involves YopB and high affinity interaction with β1 integrin receptors. The TTSS-mediated translocation of bacterial effectors into host cells is an intricate mechanism that, although extensively studied, has not been completely unraveled [31]. Here we have found that Y. pseudotuberculosis engages the small GTPase Rho to control the delivery of effectors to the host cell. Activation of this signaling pathway is mediated by the YopB/YopD translocon in cooperation with the high affinity binding of invasin or YadA to β-1 integrins. It has been put forward that pore formation and translocation of effector Yops into the host cells are not related processes [19,32]. Pore formation has been recently implicated in mediating a caspase-1 dependent type of cell death in Salmonella-infected macrophages [21]. Shin and Cornelis [33] have recently reported that insertion of translocation pores in macrophages infected with a multi-effector mutant of Y. enterocolitica triggers activation of caspase-1. Here we ruled out that in our infection system, YopB/YopD-mediated pore formation induces caspase-1 dependent cell death. Thus, amounts of a specific caspase-1 inhibitor large enough to block IL-1 β production in macrophages, does not prevent LDH release in Hela cells. Also, glycine treatment that efficiently prevented cell lysis in Salmonella infected macrophages failed to inhibit LDH release in Yersinia-infected HeLa cells. Based on these findings, we sustain that in our experimental system pore formation-induced LDH release is related to the process of Yop translocation. Both pore formation and translocation require activation of small Rho GTPases, as glucosylation of Rho, Rac and Cdc42 by C. difficile toxin ToxB potently inhibits the two processes. We found that Rac activation is not likely to be involved in pore formation or translocation. Thus over-expression of a dominant negative form of Rac does not prevent uptake of membrane impermeable dyes in cells infected with the pore forming strain. In line with these results, a specific Rac inhibitor, NSC23766, that efficiently blocks Rac-mediated internalization, does not inhibit pore formation or translocation. On the other hand, we found that signaling downstream of Rho is essential for the control of Yops delivery. Treatment with C. botulinum C3 toxin, that converts endogenous Rho A, B and C into dominant negative forms [3], potently down-regulates pore formation and translocation without affecting bacterial adhesion or internalization considerably. The type of host cell processes that Rho proteins regulate to promote translocation and pore formation most likely involves actin cytoskeleton rearrangements. Thus treatment with 2μg/ml actin polymerization inhibitor CD blocks pore formation [18] and decreases the level of YopE translocation by more than 60%. In early studies aim at demonstrating that Yop translocation is mediated by extracellular bacteria, Sory et al studied the effect of 5μg/ml CD treatment on the delivery of Yop-cyclase fusion proteins by Y. enterocolitica into murine macrophages [34]. Compared to the dramatic effect on bacterial uptake (2000 fold inhibition), the authors suggest that Yop translocation was not sensitive to the action of CD. However, their results show that CD treatment decreased YopE-cyclase translocation by 32% and YopH-cyclase by 52%. Using 10 times less CD (0.5μg/ml for 30 min) and using a strain of Salmonella ectopically expressing YopE, Rosqvist et al reported that Yop translocation into HeLa cells was notably decreased [35]. The authors also reported that the same was observed when YopE was delivered by Y. pseudotuberculosis. Interestingly, our findings strongly suggest that actin polymerization required for pore formation and translocation is dependent on Rho, as inhibition of Rho A, B and/or C results in a decrease of the number of actin halos. Adhesion of bacteria to host cell is crucial for the activation of the TTSS. In Y. pseudotuberculosis two main adhesins, invasin and YadA, mediate tight binding to host cells by interaction with β1 integrin receptors. Here we show that in an inv/yadA mutant, constitutive expression of the pH6 antigen confers good adhesion properties to host cells. In spite of that, we found that such mutants are defective in pore formation and Yop translocation, suggesting that interaction with β1 integrin receptors is essential for the two processes. Mota et al. have shown that a minimal needle length is required for efficient functioning of the Yersinia injectisome, and that this length correlates with the length of the YadA protein [36]. We considered that the attachment imparted by pH6 antigen in the absence of invasin and YadA, might not provide that critical length. Our data suggest that this is not likely to be the case in our experimental system. First, a Y. pseudotuberculosis strain expressing pH6 antigen is able to stimulate a YopB-dependent proinflammatory response, including activation of NFκB and ERK, and production of IL-8. Second, a single amino acid substitution in invasin (invD911E), that is not expected to change its length, failed to mediate efficient Yop translocation. This mutant promotes adhesion without inducing receptor clustering and subsequent β1 integrin-mediated signal transduction. Altogether, these results suggest efficient translocation requires high affinity binding of β1 integrin receptors and subsequent activation of signaling. It is still conceivable that, independent of integrin signaling, tight bacterial adhesion mediated by high affinity interaction with β1 receptors preconditions effective translocation. The fact that interfering with β1 integrin signaling by the action of a Src inhibitor impairs efficient translocation, would argue against that idea. Still, we cannot discard the possibility that Src activity might also be required for YopB/D-dependent Rho activation. We predict that upon integrin clustering, RhoA could be recruited and generate a signal that polymerizes actin. It is well documented that invasin engagement of β1 integrin receptors triggers Rac1-mediated signals that induce bacterial internalization into epithelial cells [15]. This Rac1-mediated mechanism involves Arp2/3, PIP 4,5 and capping-proteins [30]. Results from our GTP-Rho pull down assays suggest that bacteria producing invasin and YadA (YP29) can also mediate Rho activation in a YopB-independent manner. There are further evidences in the literature that engagement of β1 integrin receptors can stimulate RhoA activation. Wong and Isberg have shown that RhoA is recruited at the nascent phagosome in Cos1 cells infected with a yopE yopT mutant of Y pseudotuberculosis [26]. Werner et al have reported that interaction of invasin-coated beads with α5β1 integrin in synovial fibroblast results in beads uptake by a process that is RhoA-dependent [37]. Also, activation of RhoA by engagement of α5β1 integrins by Ipa invasins has been implicated in the internalization of Shigella to HeLa cells [38,39]. Alternatively, β1 integrin may indirectly facilitate Rho activation by a focal adhesion kinase (FAK) -dependent pathway. Such a mechanism of Rho activation has been described for the regulation of microtubules stabilization at the leading edge of mouse fibroblasts [40], and involves targeting of Rho to GM1-rich domains in the plasma membrane, where it can interact with downstream effectors. We envision a model in which high affinity binding to β1 integrin receptors, in addition to stimulating Rac activation, triggers Rho activation (Figure 8). Subsequently, YopB/D insertion into the plasma membrane stimulates increased Rho activation, and the cooperative activation of Rho stimulates Yop translocation. A central question is how Rho activation regulates Yop translocation. We hypothesize that Rho signaling induces changes in the host cell, such as actin polymerization, that are required for an efficient translocation process. One possibility is that, cell molecules present in specialized membrane microdomains, such as lipid rafts, are required for efficient translocation. These membrane microdomains would be recruited at the site of bacteria-host cell contact, as a result of Rho GTPases activation and actin polymerization. More injectisomes could then interact with lipid rafts at the site of bacteria, and more effector Yops would be translocated. Once proper amounts of Yops are delivered into the host cell, the process would be shut down to avoid further cell damage caused by excessive signaling. We based our hypothesis, in part, on the fact that Salmonella and Shigella-YopB homologues bind to cholesterol [41], and that lipid raft are required for translocation in Salmonella, Shigella and EPEC [41]. Interestingly, actin polymerization and Rho GTPases activation have been shown to be involved in lipid raft clustering in B cells [42], T cells [43] and NK cells [44] . Why is Rho-dependent, but not Rac-dependent, actin polymerization required for translocation? Rho GTPases transmit signals that control the formation of distinct cytoskeletal structures through the interaction with different nucleating machineries. Cdc42 and Rac mediate nucleation of branched actin filaments through the Arp2/3 protein complex, leading to lamellipodia formation. On the other hand, Rho proteins stimulate unbranched actin filaments formation, such as those in stress fibers, via interaction with formins. It could be speculated that only F-actin structures generated by formins are important for translocation. The effect of the expression of dominant negative mutants of the formin mDia1 on translocation will be investigated in future studies. Findings from two studies that investigate translocation of TTSS effector proteins by Salmonella and Shigella in real time [45,46] indicate that effector translocation occurs right after host cell contact, with a half maximal rate of about 4 min. In our experimental model we detect the strongest Rho activation after 10 to 15 min infection with a YopEHJ bacteria. This is probably due to accumulation of GTP-Rho in the absence of the Rho inhibitors YopE and YopT. The decrease in the levels of GTP-Rho after 15 min is presumably due by the action of endogenous GAPs. We envision that during infection with wild type bacteria, the kinetics of Rho activation would be much faster. Translocation of Salmonella SipA and SopE, and Shigella IpaC were found to follow a linear kinetic [45,46]. Interestingly, however, slopes of IpaB secretion kinetics curves seemed to vary at different time points, suggesting that the speed of injection changes during the course of the translocation process resembling a slow-fast-slow type of mechanism. This type of translocation kinetic is what we would expect in our model. How does our model fit with the mechanism of Yop translocation in Y. pestis? Although closely related to Y. pseudotuberculosis, Y. pestis lacks invasin and YadA. Unless Y. pestis has yet-unidentified adhesins that interact with β1 integrin receptors, we envision that the bacteria would activate Rho only by the stimulus elicited by YopB/D. In this situation, Rho activation would be limited, and therefore, one should expect that Y. pestis would be less effective for Yop translocation. A recent report suggests that, in macrophages, Y. pestis translocate less YopJ than a Y. enterocolitica strain expressing invasin and YadA [47]. However, in this report the authors suggest that this is most likely due to a difference between the YopJ protein from the two Yersinia species. We have preliminary results suggesting that Y. pestis deliver much less YopE in HeLa cells than Y. pseudotuberculosis. It has been proposed that, because bacterial effectors are directly injected within cell cytosol, the TTSS does not need to trigger signals through cell surface receptor [48]. Our data suggest that, although not essential, signal stimulated by engagement of β1 integrin receptors greatly enhances Yop translocation. The wild-type serogroup III Y. pseudotuberculosis strain YP126 [49], and the mutants derived thereof are shown in Table 1. YP126 and its derivatives carry a naturally occurring deletion in virulence plasmid that inactivates the yopT gene and are thus devoid of YopT activity [50]. YP202/YP29 (yopEHJB,inv) was constructed by inserting the virulence plasmid of YP29 into a plasmid cured, inv::kan strain (YP202, Table 1). To create YP50 (yopEHJ,yadA,inv) and the corresponding YopB-deficient mutant (YP51), the wild type yadA gene in YP202/pYP27 (yopEHJ,inv) and in YP202/pYP29 (yopEHJB,inv), respectively, was replaced by yadA containing a frame shift deletion (yadAfs), as follows. yadAfs was constructed by amplifying yadA with primer YadA F1 (5′-CCC GGG TTT GTA GTG GGC TGA CTC CGA C-3′) and B1 (5′'-GGC TGA ACT GGC TAA ACC TTT G-3′). The yadA DNA fragment was subsequently blunt-cloned into pETBlue (Novagen). QuikChange Site-Directed mutagenesis (Stratagene) was used to create the frame-shift and generate a SphI restriction site using primers F2 (5′-CA CAA GGT CCA GAA AAA AAA GAG CAT GCA TTA GCA GAA GCA ATA C-3′), and B2 (5′-GTA TTG CTT CTG CTA ATG CAT GCT CTT TTT TTT CTG GAC CTT GTG-3′). Plasmid pETBlue-yadAfs was digested with XmaI and subcloned into the suicide plasmid pSB890 containing sacB and TetR genes [51]. pSB890yadAfs was then introduced into S17-λpir and conjugated into CamR YP202/pYV27 and YP202/pYV29. TetR CamR colonies were grown for several generations in the absence of Tet and were selected against the sacB on LB-5% sucrose. SucroseR, CamR and TetS colonies were screened for yadAfs by PCR using primers YadA F1 and B1, followed by SphI-digestion of the amplified yadA fragment. A plasmid constitutively expressing pH6 antigen fimbriae, pAY66 (LacP::psaABC, Table1), a gift from R. Isberg (Tufts University), was inserted into YP50 and YP51 by electroporation. To create YP54/pAY66 (yopHJ,inv,yadA/psaABC), we replaced yopE::kan in YP50/pAY66 by wild type yopE, by allelic exchange using suicide plasmid pSB890YopE, essentially as described above. To construct YP50invD911E (yopEHJ,yadA,invD911E) and YP51invD911E (yopEHJB,yadA,invD911E), the virulence plasmid from YP50 and YP51 were introduced into YPIII P− invD911E (Table 1, gift from R. Isberg ) by electroporation. To create YP54invD911E, we replaced yopE::kan in YP50invD911E by wild type yopE, by allelic exchange using pSB890yopE as described above. Plasmid pMMB67HEYadA (pYadA) [52], was inserted into YP50 and YP54 by electroporation. HeLa cells were cultured as previously described [16]. For experiments carried out in the presence of inhibitors, cells were pre-incubated with 50–100 μM Ac-YVAD-cmk (Calbiochem), 5mM Glycine (Roche), 40ng/ml Clostridium difficile ToxB (Calbiochem), 3.9 μM (2μg/ml) cytochalasin D (Sigma), 10 μM PP2 (Sigma), 100 μM NSC23766 (Calbiochem), 10, 20, or 40 μg/ml TAT-C3. Bacteria used for infections were grown in Luria-Bertani (LB) broth either under conditions that stimulate (low Ca2+ at 37 °C) or repress (high Ca2+ at 28 °C) Yop expression [4,51], at a multiplicity of infection of 50 to 100. The plates containing the infected cells were centrifuged for 5 min at 700 rpm and incubated at 37 °C with 5% CO2 for different periods of time to allow bacterial-host cell interaction. Cells cultured in 24-well plates with coverslips were infected for 3 h with bacteria grown under low calcium conditions. A green fluorescent membrane-permeable nucleic acid stain (SYTO10) and a red membrane-impermeable nucleic acid dye that label only cells with compromised membranes, ethidium homodimer-2 (EthD-2) were provided in the DEAD-LIVE kit (Invitrogen). After washing, a mixture of the two dyes was added to the wells and incubated in the darkness for 15 min at room temperature. Cells were then washed and fixed with 2% paraformaldheyde in PBS. Coverslips were mounted with 8 μl of ProLong mounting medium (Molecular Probes) and slides were then examined by immunofluorescence microscopy. Samples of culture media from wells containing infected cells were collected 3 h post infection. Levels of LDH were assayed using the CytoTox 96 assay kit (Promega) as previously described [16]. HeLa cells cultured in 6 cm2 dishes were infected with bacteria grown at high Ca2+ conditions. Infected cells were lysed with 0.2 ml of cold 1% Triton X-100 buffer as described [4]. Soluble and insoluble fractions were subjected to immunoblotting using an affinity purified polyclonal anti-YopE and anti YopH antibodies, as described previously [4]. Anti-β actin antibody was used as a loading control. Anti-rabbit antibodies conjugated with IR800 or IR680 were used as secondary antibodies, and the infrared signal was detected using an infrared imaging system (Odyssey, LI-COR). Quantification of a fluorescent signal is more accurate than that generated by chemiluminescence because its intensity is not time-dependent. The bands intensities were calculated using the software provided by the Odyssey system, and the values were expressed as the YopE/β-actin ratio and plotted on a graph. Supernatants of infected HeLa cells were assayed for IL-8 production by ELISA (Antigenix America) five h after infection, as previously described [16]. Values obtained from triplicate wells were assayed in duplicate and averaged. HeLa cells were seeded onto glass coverslips at 105 cells per well in a 24-well tissue culture plate 24 h before infection. Cells were infected with bacteria at a calculated MOI of 50:1. After a brief centrifugation step (5 min at 100 g), the plates were incubated for 30 min at 37 °C in a 5% CO2 incubator. A double-label immunofluorescence assay was used to differentiate between extracellular and intracellular cell-associated bacteria as previously described [4]. Coverslips containing infected cells were washed with PBS and fixed in 2% paraformaldehyde for 15 min. The washed coverslips were incubated with polyclonal anti-Yersinia antibody SB349 (diluted 1:1000) for 40 min to stain extracellular bacteria. Washed coverslips were incubated for 40 min with FITC-conjugated goat anti-rabbit IgG diluted 1:250. After washing, cells were permeabilized with 0.2% Triton X-100 for 10 min. Coverslips were washed and incubated with SB349 (1:1,000) for 40 min to label both extracellular and intracellular bacteria. Samples were then washed and incubated for 40 min with TRITC-conjugated goat anti-rabbit IgG (1:300). All antibodies were diluted in PBS containing 3% BSA, and washes were conducted three times for 5 min with PBS containing 1% BSA. Coverslips were washed with PBS before mounting and examined by immunofluorescence microscopy. The percentage uptake was calculated as the number of [intracellular bacteria (red)/total bacteria (green and red)] × 100. The effect of Rac and Rho inhibitors on the formation of actin cups was tested in Hela cells seeded on coverslips. To inhibit Rac, the cells were treated for 6 h with NSC23766 (100μM) in 5% serum-DMEM, or 5% serum-DMEM alone. To inhibit RhoA, B and C, TATC3 (40 μg/ml) was added to the cells in serum free medium for 4 h, and control cells were incubated in serum free conditions for the same time. Hela cells were then infected for 10–15 min, washed and fixed as described above for the bacterial uptake assay. Double immunofluorescence was performed as detailed above for the bacterial uptake assay, with the addition of 50 U/ml of Rhodamine Phalloidin (Molecular Probes) together with the last secondary antibody. Images were captured with a confocal laser microscope. The percentage of bacteria (extracellular and intracellular) surrounded by an “actin halo” was calculated by counting a minimum of 150 bacteria. Plasmid pTAT–C3 (a gift from Dafna Bar Sagi, Stony Brook University, NY) was introduced into E. coli (strain BL21), and His-tagged-TAT–C3 protein were expressed by IPTG induction (1 mM IPTG, 4 h). Recombinant His-TAT–C3 was extracted from E. coli BL21 strain by sonication, and purified by fast protein liquid chromatography (FPLC), as follows. The supernatant of the cell lysate was injected onto a Hi-trap Ni-column (Pharmacia Co.). The column was washed with a 5 mM imidazole buffer solution and eluted using a gradient concentration of 1M imidazole buffer. After dialysis against PBS/0.5M NaCl, the purity of each TAT–C3 preparation was determined on polyacrylamide gels stained with Coomassie blue. Cells were seeded in 10 cm dishes at 90% confluency and were left uninfected or were infected at a moi:100 for different time periods. Cells were lysed in lysis buffer (Upstate, Rho activation assay) containing 10% glycerol, and protease inhibitor (Roche). Cell lysates were clarified by centrifugation at 13,000 rpm at 4 °C for 10 min, and the supernatants were incubated with 30 μg of GST fused to the Rho binding domain of rhotekin bound to with glutathione beads, at 4 °C for 45 min. The beads were washed twice with lysis buffer and subjected to SDS-polyacrylamide gel electrophoresis on a 12% gel. Bound RhoA was detected by Western blot using a monoclonal antibody against RhoA (Santa Cruz Biotechnology). The GenBank (http://www.ncbi.nlm.nih.gov/Genbank/index.html) accession number for invasin is M17448.

10.1371/journal.ppat.1003247

Brain Inositol Is a Novel Stimulator for Promoting Cryptococcus Penetration of the Blood-Brain Barrier

Cryptococcus neoformans is the most common cause of fungal meningitis, with high mortality and morbidity. The reason for the frequent occurrence of Cryptococcus infection in the central nervous system (CNS) is poorly understood. The facts that human and animal brains contain abundant inositol and that Cryptococcus has a sophisticated system for the acquisition of inositol from the environment suggests that host inositol utilization may contribute to the development of cryptococcal meningitis. In this study, we found that inositol plays an important role in Cryptococcus traversal across the blood-brain barrier (BBB) both in an in vitro human BBB model and in in vivo animal models. The capacity of inositol to stimulate BBB crossing was dependent upon fungal inositol transporters, indicated by a 70% reduction in transmigration efficiency in mutant strains lacking two major inositol transporters, Itr1a and Itr3c. Upregulation of genes involved in the inositol catabolic pathway was evident in a microarray analysis following inositol treatment. In addition, inositol increased the production of hyaluronic acid in Cryptococcus cells, which is a ligand known to binding host CD44 receptor for their invasion. These studies suggest an inositol-dependent Cryptococcus traversal of the BBB, and support our hypothesis that utilization of host-derived inositol by Cryptococcus contributes to CNS infection.

Cryptococcus neoformans is an AIDS-associated human fungal pathogen that annually causes over 1 million cases of meningitis world-wide, and more than 600,000 attributable deaths. Cryptococcus often causes lung and brain infection and is the leading cause of fungal meningitis in immunosuppressed patients. Why Cryptococcus frequently infects the central nervous system to cause fatal meningitis is an unanswered critical question. Our previous studies revealed a sophisticated inositol acquisition system in Cryptococcus that plays a central role in utilizing environmental inositol to complete its sexual cycle. Here we further demonstrate that inositol acquisition is also important for fungal infection in the brain, where abundant inositol is available. We found that inositol promotes the traversal of Cryptococcus across the blood-brain barrier (BBB), and such stimulation is fungal inositol transporter dependent. We also identified the effects of host inositol on fungal cellular functions that contribute to the stimulation of fungal penetration of the BBB. We propose that inositol utilization is a novel virulence factor for CNS cryptococcosis. Our work lays an important foundation for understanding how fungi respond to available host inositol and indicates the impact of host inositol acquisition on the development of cryptococcal meningitis.

Cryptococcus neoformans is a basidiomycetous yeast pathogen that often causes life-threatening infections. It causes the most common fungal infection of the central nervous system (CNS) in HIV-infected persons and may present as encephalitis, meningitis, or cerebral-space-occupying lesions [1], [2], [3], [4], [5], [6]. Cryptococcal CNS infections are uniformly fatal in the absence of treatment [1], [7]. A recent survey suggests that each year there are around 1 million new cases of cryptococcal meningitis, which result in over 600,000 deaths annually [2]. Despite its medical importance and significant research efforts [3], [8], [9], [10], the molecular basis of cryptococcal CNS infection and host factors affecting disease development are poorly understood, which complicates efforts for rapid diagnosis and effective treatment. Hence, there is an urgent need to understand the molecular basis of cryptococcal CNS infection to allow the discovery and development of safer and more effective antifungal drugs. C. neoformans is a globally ubiquitous organism, which is commonly associated with certain environmental niches, including plants and soil contaminated with plant debris and bird droppings. Our previous studies revealed that this fungus can utilize inositol from plant surfaces to complete its sexual cycle [11]. Inositol is essential for cellular structure and regulation of intracellular signaling in all eukaryotes. Recent studies showed that the enzymes involved in inositol metabolism and inositol sphingolipid biosynthesis play a central role in the pathogenesis of C. neoformans [12], [13]. The inositol phosphorylceramide synthase 1 (Ipc1) protein, an enzyme of the fungal sphingolipid pathway, activates protein kinase C (PKC), which regulates the cell wall integrity of Cryptococcus and manifestation of its virulence factors [12], [14], [15]. Moreover, although it prefers to grow on media containing fermentable sugars such as glucose, C. neoformans can utilize free inositol as a sole carbon source [16], [17]. Consistent with the central importance of inositol in its development and virulence, Cryptococcus has developed a sophisticated inositol acquisition system that contains an unusually large inositol transporter gene (ITR) family with more than ten members, which contrasts with the one or two members found in most other fungi [18], [19], [20], [21]. We also demonstrated that these ITRs are required for cryptococcal infection in murine models [11], [22], [23]. In addition, Cryptococcus can utilize intracellular glucose to produce inositol in a multi-step de novo inositol biosynthetic pathway in which inositol 1-phosphate synthase (Ino1) is the rate-determining enzyme [19], [24]. Cryptococcus invasion and traversal of the blood-brain barrier (BBB) are prerequisites for CNS infection, the major cause of morbidity and mortality in people with cryptococcosis. There are evidences for both direct invasion of the endothelial cells lining the brain vasculature [25], [26] and for a “Trojan horse” mechanism whereby cryptococci enter the CNS after macrophage ingestion [27], [28], [29]. Several factors, including urease [30], [31], phospholipase B1 [32], [33], as well as host plasmin [34], have been reported to be involved in the Cryptococcus invasion of the BBB. It was recently reported that Cryptococcus interacts with lipid rafts of human brain microvascular endothelial cells (HBMECs) to promote invasion in a glycoprotein CD44-dependent manner [35]. Hyaluronic acid produced by the fungus has been found to function as a ligand for the CD44 receptor during the fungal-host cell interaction [36], [37]. However, the molecular basis for the highly frequent occurrence of cryptococcal CNS infection remains poorly understood. Human and animal brains contain high concentrations of free inositol, and inositol can be used as a carbon source for Cryptococcus. Inositol is one of the most abundant metabolites in the human brain; it is located mainly in glial cells, and functions as an osmolyte. Inositol is present in the human cerebellum (5.1 mM) at 200-fold higher concentrations than in plasma (0.02 mM) [38]. Even higher inositol concentrations (>8 mM) are detected in astrocytes that directly associate with the BBB, and inositol can be rapidly released during hyperosmolarity [38], [39]. It is believed, because of the tight interaction of astrocytes with brain microvascular endothelial cells of the BBB, that the inositol concentration around the BBB is much higher than in plasma, although the inositol level in the parenchyma around cerebral vasculature has not been precisely determined. Together with the importance of ITRs in fungal infection, we hypothesize that brain inositol is an important host factor for the development of cryptococcal meningitis. We further hypothesize that fungal inositol transporters are important for sensing and/or transporting host inositol during disease progression. In this study, we utilize an in vitro human BBB model and in vivo murine models to dissect the role of fungal inositol transporters and host inositol in the traversal of Cryptococcus across the BBB and in the development of cryptococcal meningitis. Our results showed that addition of inositol can facilitate Cryptococcus transmigration in an in vitro BBB model in an ITR-dependent manner. These observations tie inositol to the fungal infection in the brain. This work provides a framework explaining the role of inositol in enabling Cryptococcus to cross the BBB. Previous studies demonstrate that the route by which C. neoformans gains access to the CNS is through traversal across the BBB [37], [40], [41]. The high abundance of inositol in human brain is suggested to be a host factor that promotes the high rate of cryptococcal meningitis [22], [42]. To understand whether inositol plays a role in the development of cryptococcal CNS infection, we investigated the role of free inositol in C. neoformans transmigration across the BBB, using an in vitro human BBB system. Our in vitro BBB system is composed of human brain microvascular endothelial cells (HBMECs) grown on Transwell membranes to confluence, separating the top compartment (blood side) and bottom compartment (brain side) as described in Materials and Methods. We initially performed transmigration assays with C. neoformans in the presence of inositol. Inositol was added to the bottom compartment of the Transwells 30 min prior to addition of Cryptococcus cells to the top compartment to mimic the situation in the brain. As shown in Fig. 1A, the number of C. neoformans var. grubii (strain H99) cells that transmigrated across the HBMEC monolayers was not different from control transmigration (no inositol) at inositol concentrations up to 0.5 mM. However, at inositol concentrations greater than 1 mM, transmigration of C. neoformans was increased 3-fold compared to controls without inositol treatment, indicating that inositol enhances cryptococcal traversal in a dose-dependent manner. Because the binding of fungal cells to HBMEC monolayers is the first step in transmigration, association assays were carried out under the same inositol treatment. The results showed a significantly better association between cryptococcal cells and the brain endothelial cells when the bottom chamber contained 1 mM or higher concentration of inositol, indicating that inositol can promote Cryptococcus binding (Fig. 1B). Because the tissue culture medium for HBMEC is rich in nutrients including sufficient glucose (8 mM) for optimal fungal growth, all Cryptococcus strains should grow well in this medium. To address the concern that fungal cells may proliferate better in the presence of additional inositol, which might account for the apparent increase in fungal cells in the bottom compartment in the presence of high inositol, the growth rate of H99 in the bottom culture medium was determined in the presence or absence of 1 mM inositol. The results showed that H99 proliferated at the same rate with or without inositol (Fig. 1C). The average replication time for H99 in both media was about 2.2 hours. Therefore, the increased number of Cryptococcus cells in the bottom compartment at higher inositol levels could not be due to increased proliferation, and must be the result of increased transmigration. Thus, these results demonstrate that inositol stimulates traversal of cryptococcal cells across the BBB. To determine whether the inositol effect is strain specific, the transmigration assay was also carried out with C. neoformans var. neoformans strain B3501. C. neoformans var. grubii strains are in general more virulent than var. neoformans strains even though both varieties are able to cause systemic cryptococcosis and meningitis [43]. The number of transmigrated B3501 cells was comparable to that of strain H99, demonstrating that inositol enhances transmigration of C. neoformans regardless of strain origins (Fig. 1D). In addition, the inositol effect on transmigration was examined in Candida albicans, another yeast pathogen that occasionally crosses the BBB through direct transcytosis to cause CNS infection in humans [44], [45]. C. albicans contains one inositol transporter, Itr1, that is not required for fungal virulence [20]. Transmigration of C. albicans occurred at a much lower rate and was not enhanced by inositol (Fig. 1D). Subsequently, the transmigration assays were performed in the presence of another inositol isomer scyllo-inositol, or other monosaccharides such as galactose and mannose. Myo-inositol elevated the efficiency of cryptococcal transmigration exhibiting 2-fold greater number of transmigrated fungal cells than the untreated control after 3, 6 and 9 hrs of incubations (Fig. 1E). However, the cryptococcal transmigration remained unchanged in the presence of other sugars. This result indicates that myo-inositol is a specific effector for promoting cryptococcal traversal across the BBB. Taken together, our findings indicate that inositol specifically increases BBB traversal by C. neoformans. To understand how inositol in the bottom compartment affects Cryptococcus cells that are present in the top compartment, we measured the concentration of inositol in the top chamber by using an enzymatic method [46]. The medium alone contains around 0.18 mM inositol. In the absence of Cryptococcus in the top compartment, the addition of 1 mM inositol in the bottom compartment did not result in a measurable increase in the inositol levels in the top compartment after 3 hr incubation and only an increase to 0.25 mM inositol after 6 hr incubation, indicating that the HBMEC monolayer maintained high integrity and that inositol diffusion was very slow (Fig. 2A). In contrast, when 105 Cryptococcus cells were added in the top compartment, the inositol level in the top compartment reached 0.47 mM after 3 hr and 0.84 mM after 6 hr (Fig. 2A). These results demonstrated that inositol can diffuse through the HBMEC monolayer from the bottom to the top compartment at a higher rate in the presence of fungal cells in the top compartment, possibly due to increased inositol permeability caused by Cryptococcus–HBMEC interactions. Modification of tight junctions during Cryptococcus transmigration has been reported previously, which might contribute to the increased inositol permeability [40]. To test this hypothesis, we examined the integrity of tight junctions in response to Cryptococcus infection by immunofluorescence microscopy. Zona Occludens-1 (ZO-1) is a member of the tight junction protein complex and is widely used as a marker to determine the location of tight junctions [45]. The untreated HBMEC monolayer displayed continuous lining of ZO-1 staining pattern, suggesting intact tight junctions (Fig. 2B). However, Cryptococcus treatment induced dislocation of the ZO-1, results in the discontinued and scattered staining pattern between neighboring cells (Fig. 2C). These results provide evidence that the HBMEC-Cryptococcus interaction leads to the modification of tight junctions, which may contribute to the increased inositol permeability without causing major damage in the integrity of the HBMEC monolayer. To investigate whether such modification leads to the alteration of the integrity of the HBMEC monolayer, we also measured the transendothelial electrical resistance (TEER) of the monolayer. Our results showed a similar TEER readout in the absence or presence of yeast cells in the upper chamber (Fig. S1), suggesting there was no major change to the integrity of the monolayer, which is consistent with previous studies [25], [47]. Our previous studies demonstrated that C. neoformans has an unusually large inositol transporter (ITR) gene family, and established that ITRs were required for the full virulence of Cryptococcus in in vivo murine models [11], [22], [23]. Among them, Itr1a and Itr3c are two major ITRs for inositol uptake and fungal virulence [22]. To determine whether the attenuated brain infection of ITR gene deletion mutants was due to their defective ability to cross the BBB, we examined the transmigration ability of C. neoformans itr1aΔ itr3cΔ double mutants lacking these two major inositol transporter genes in our in vitro BBB model. In the absence of additional inositol, transmigration of the itr1aΔ itr3cΔ double mutant was decreased by 50% at 3 and 6 hr incubation periods compared to the wild type, indicating that ITR genes are required for cryptococcal crossing of the BBB. With inositol treatment, cryptococcal traversal was enhanced regardless of presence of ITR genes; however, a more significant defect of transmigration ability of itr1aΔ itr3cΔ mutants was evident. The number of transmigrated wild type H99 strain increased by 2.7 and 2-fold in the presence of inositol at 3 and 6 hr incubation, respectively, compared to transmigration in the absence of added inositol. In contrast, the itr1aΔ itr3cΔ double mutant exhibited approximately 1.7 and 1.3-fold increased transmigration rates after 3 and 6 hr incubation, respectively, with the result that the number of transmigrated itr double mutant cells was a third of that of the wild type strain H99 (Fig. 3A). The reduction in transmigration ability of the itr1aΔ itr3cΔ double mutant was fully restored by complementation. Because inositol can be used as the carbon source for growth, there is a logistic concern that fungal cells with intact ITRs could grow better in medium with additional inositol. To address the concern, we compared the growth rate of wild type and the itr1aΔ itr3cΔ double mutant in the medium with or without addition of inositol. The results showed that all tested strains have similar growth rates regardless of the presence of inositol or ITR genes (Fig. 1C). Thus, the data are consistent with the interpretation that Itr1a and Itr3c play an important role in responding to inositol availability and contribute to cryptococcal traversal across the HBMEC monolayer. We next pre-incubated Cryptococcus cells with 1 mM inositol (0.5, 1 and 3 hr), and then removed inositol by thorough washing with PBS before assessing transmigration (Fig. 3B). Cryptococcus cells pre-incubated for 30 min showed comparable transmigration to the untreated control, whereas longer pre-incubations of 1 or 3 hr enhanced the number of transmigrated fungal cells by 20% and 40%, respectively. However, similar enhancement in fungal transmigration was not detected with the itr1aΔ itr3cΔ double mutants following inositol pre-incubation (Fig. 3B). Association assays were then carried out with fungal cells pre-incubated with inositol to compare the ability of the wild type strain and the itr1aΔ itr3cΔ double mutant to associate with the HBMEC. The number of associated H99 cells was not changed after 30 min pre-incubation but was significantly increased after 3 hr compared to that of the untreated control (Fig. 3C). Cryptococcus cells pre-incubated for 1 hr exhibited a modest increase. Unlike the H99 wild type strain, the association rate of the itr1aΔ itr3cΔ double mutants was not changed by inositol pre-incubation (Fig. 3C), which is similar to the transmigration result shown in Fig. 3B. These results demonstrate that inositol uptake and utilization are required for efficient association and transmigration of cryptococcal cells. Our results also suggest that inositol uptake by Cryptococcus cells may lead to modification of the surface of Cryptococcus cells to enhance its association with and subsequent transmigration across the HBMEC monolayer. To further extend the results obtained from the in vitro human BBB model, we assessed infection with the itr1aΔ itr3cΔ double mutant in a murine model via intravenous injection. Infected mice were sacrificed at 1, 6, 24, 48, or 72 hr post-inoculation; brains and lungs were isolated and yeast CFUs were determined. Our results demonstrated that there was a significant difference in fungal burden in the brain between mice infected by wild type H99 or the itr1aΔ itr3cΔ double mutant after 24 hr post-infection. However, there was no significant difference in CFU at earlier time points (Fig. 4A). On the other hand, the fungal burden was similar in lungs infected either by wild type or the mutant at all time points except 72 hr (Fig. 4B). Our results thus demonstrate that inositol transporters Itr1a and Itr3c are required for fungal cells to either cross the BBB or grow in the brain after transmigration. It has been reported that although fungal crossing of the BBB occurred quite early after inoculation, the rate of traversal of Cryptococcus cells across the BBB increases dramatically 24 hr post-injection via tail vein [48]. We hypothesize there is a threshold effect with respect to time or number of cryptococcal cells accumulating on the endothelial monolayer before they can effectively penetrate the barrier. Fungal burdens at 72 hr post-injection were reduced in both brains and lungs infected by the mutant. However, the reduction is much greater in the brain (7-fold) compared to the reduction in the infected lung (4-fold) (compare 72 hr time point in Fig. 4). To investigate whether the mutant has a defect in survival in macrophage, we performed Cryptococcus-macrophage interaction assays using the macrophage-like cell line J774. Our results showed that wild type and the double mutant had similar response to phagocytosis and macrophage killing (Fig. S2). To further understand the brain fungal infection at early time points post inoculation, we examined the presence of Cryptococcus within the CNS. To perform these studies, animals were infected with H99 or the itr1aΔ itr3cΔ double mutant for 48 hr before they were sacrificed to obtain the brain. Staining and subsequent confocal microscopy of 30–50 µm brain tissue sections was performed for polysaccharide glucuronoxylomannan (GXM) antibody to visualize the transmigration of cryptococcal cells into the CNS parenchyma. Our results showed that brains infected with H99 resulted in cryptococcal cells invasion of the CNS (5 to 7 lesions for each brain, especially in the cortex area). Transmigration of Cryptococcus cells was associated with large lesions (Fig. 5). Animals infected with the itr1aΔ itr3cΔ mutant present fewer lesions and cryptococcal cells within the CNS (Fig. 5, an average of 2 lesions by brain analyzed). The outcomes from fungal CFU counts and from immunofluorescent staining further support our hypothesis that fungal ITRs play a positive role in Cryptococcus traversal of the BBB in CNS infection. All these results are consistent with our CFU counts of infected brains (Fig. 4A). To address the possibility that the double mutant may grow slower once inside the brain, thereby leading to lower CFU recovery, we tested the growth of the itr1aΔ itr3cΔ mutant on an in vitro cerebral spinal fluid (CSF) medium that has been successfully used to identify strains with a growth defect in the CNS compartment [49]. There was no significant growth defect exhibited by the itrΔ single and double mutants on CSF medium (Fig. 6A). To further confirm the in vitro growth results on CSF medium, we tested the itr1aΔ itr3cΔ double mutant in a murine intracerebral injection model of cryptococcosis. CFU from mouse brains were measured at 1, 3, and 7 days post-injection. Our results showed that at all three time points, brains (n = 4) infected by either the wild type H99 or the double mutant contain similar amount of fungal cells, indicating that there is no growth difference in vivo during brain infection (Fig. 6B). We also compared the in vivo growth of wild type and the double mutant in a rabbit CSF model of cryptococcosis via intrathecal inoculation. This model allows us to measure the yeast CFUs from the same animal at different time points post-infection. CFU from the rabbit subarachnoid space were measured 3, 7, and 10 days post-infection. All rabbits (n = 3) infected by H99 were dead before 10 days, while one rabbit infected by the mutant remained alive at 10 days post-inoculation. Overall, these results showed that comparable numbers of CFUs were isolated from rabbits infected by either wild type or the mutant strain (Fig. 6C). This result further confirms that the defect in BBB traversal is the main reason for the virulence attenuation of the itr1aΔ itr3cΔ mutant shown in the murine tail vein injection model (Fig. 4A). Inositol is a precursor for the production of phospholipids, which are essential for cellular functions in eukaryotes. Because large amounts of inositol diffused from the bottom compartment into the top in our in vitro system, Cryptococcus can utilize the inositol available in the medium for its cellular functions. One possible explanation for the inositol effect on Cryptococcus transmigration is a change in phospholipid composition. To interrogate this hypothesis, we performed 2-dimensional thin layer chromatography assays (2D-TLCs) to evaluate the production of phospholipids in Cryptococcus. Our results showed that when yeast cells were grown on synthetic medium containing 5 mM inositol, the same phospholipid species were present in both H99 and the itr1aΔ itr3cΔ double mutant, but the production of phosphatidylinositol (PI) was two-fold lower in the mutant strain than in the wild type. Production of one unidentified lipid species was also significantly reduced in the mutant strain with inositol treatment (Fig. 7). Production of other major phospholipid species, e.g. phosphatidylcholine (PC), phosphatidylserine (PS), and phosphatidylethanolamine (PE), were similar between these two strains (Fig. 7). Hence, phospholipid composition, especially the difference in PI composition, could play a role in the dramatic reduction of transmigration in the double mutant. Furthermore, when we compared the phospholipids in H99 treated or not treated with inositol, we observed a dramatic difference in overall phospholipid levels, and addition of inositol significantly induced the production of all detected phospholipid species (Fig. 7). These results indicate that inositol plays a significant role in fungal phospholipid production, and could be part of the explanation for the inositol effect on Cryptococcus interaction with the HBMEC monolayer. To further understand how Cryptococcus cells respond to inositol, we analyzed transcriptional profiles of Cryptococcus that were treated or not treated with inositol to identify genes that are regulated by inositol. Cryptococcal cells were treated with 5 mM inositol for 24 hrs. Total RNA was prepared for hybridization with Cryptococcus 70-mer whole genome array chips. Our results showed that roughly 50 genes were significantly upregulated (>2 fold) (Table 1; Fig. S3), which is similar to the number of genes upregulated during mating under inositol induction conditions [23], while more than 300 genes were significantly downregulated (>2 fold) (Table 2). The complete list of genes with greater than two-fold change is shown in Table S1. Among upregulated genes, inositol oxygenase and beta-glucuronidase homologs were highly induced by inositol treatment (Table 1). Quantitative RT-PCR analyses were performed for six genes to confirm the microarray results (Fig. S3). The results demonstrated that the expression of two (myo)-inositol oxygenase genes (MIO1 and MIO2), as well as two beta-glucuronidase genes (CBG1 and CBG2) were indeed highly upregulated. Cryptococcus can use inositol as a carbon source. Conversion of inositol to glucuronic acid by inositol oxygenases (MIOs) is the first step of the only known pathway for inositol catabolism: the oxygenase controls the utilization of inositol as an energy source [50], [51]. C. neoformans has three MIO homologues (CNAG_06623, CNAG_03277, and CNAG_05316), which is another unusual feature that may be related to inositol function in this fungus. Multiple copies of the MIO gene is unique to Cryptococcus among the animal and fungal kingdoms [51]. Beta-glucuronidase is a member of the glycosidase family that catalyzes the breakdown of complex carbohydrates by releasing the glucuronic acid residues from polysaccharides [52]. In addition, a few inositol transporters are upregulated by inositol, while the gene encoding inositol 1-phosphate synthase (INO1) is significantly downregulated, suggesting, as we expected, that these two inositol acquisition pathways are regulated by inositol (Table 1 & 2). Cps1 is identified as the hyaluronic acid synthase in Cryptococcus [53], [54]. Hyaluronic acid has been reported to be a Cryptococcus ligand that can bind to the CD44 glycoprotein in HBMECs [55]. Using a quantitative RT-PCR analysis, we also detected high induction of CPS1 gene expression following treatment with inositol (Fig. 8A). The overproduction of hyaluronic acid likely increases the association and transmigration of Cryptococcus. Therefore, we measured hyaluronic acid production using a hyaluronic acid ELISA kit (Corgenix, Colorado, AZ). The results showed that the production of hyaluronic acid in wild type H99 was significantly increased (P<0.001) when the medium contained 1 mM inositol, confirming that inositol regulates hyaluronic acid production in C. neoformans, leading to an increased rate of fungal association and transmigration. In addition, the itr1aΔ itr3cΔ double mutant also produced a reduced level of hyaluronic acid compared to H99 in the presence or absence of inositol in the medium, but the reduction is less significant (P = 0.12) (Fig. 8B). This outcome suggests that inositol regulates the production of hyaluronic acid, but additional mechanisms may also be involved in fungal transmigration. Fungal inositol transporters are proton-dependent, which is different kinetically and pharmacologically from the sodium-dependent myo-inositol transporters (SMITs) in mammalian cells [56]. Thus, fungal inositol transporters have potential as attractive drug targets. Dinitrophenol (DNP) is a protonophore that dissipates transmembrane proton gradients and has been shown to effectively block inositol uptake in Candida albicans [56]. We treated Cryptococcus wild type strain H99 with DNP or with the human sodium-dependent inositol transport inhibitor phloretin before performing inositol uptake assays. The inositol uptake assays revealed that DNP produced a pronounced inhibitory effect on inositol import in Cryptococcus, while phloretin had no obvious effect (Fig. 9). Subsequently, DNP or phloretin pretreated H99 cells were used in fungal transmigration assays in the in vitro BBB model. The results demonstrated that the DNP pretreatment leads to a significant reduction in cryptococcal transmigration stimulated by the addition of inositol, compared to those treated with a vehicle (DMSO) or phloretin. These findings further confirm that cryptococcal uptake of host inositol is through proton-dependent inositol transporters, and is required for assisting Cryptococcus traversal across the HBMEC monolayer. Due to the nature of protonophore, DNP treatments may cause effects on cryptococcal cells in addition to inhibiting inositol uptake, such as effects on mitochondria function. To address the concern that DNP may alter the growth of cryptococcal cells, we compared the growth rates of H99 cells treated with or without either DNP or phloretin. The results showed that the different treatments did not affect growth curves (Fig. S4), indicating that the decreased cryptococcal transmigration was due to reduced inositol uptake in H99 rather than an effect on other fungal physiology. Taken together, these findings further confirm that host inositol plays a positive role in Cryptococcus traversal across the BBB. The mechanisms for the frequent occurrence of Cryptococcal CNS infection remain unclear. In this study, we explored our hypothesis that the high abundance of inositol in human brain contributes to virulence of Cryptococcus and the development of cryptococcal meningitis. In our in vitro model of BBB using the HBMEC monolayer, we observed that inositol promotes an increase in Cryptococcus association with and transmigration through the HBMEC monolayer. This increase is dependent upon fungal inositol transporters (ITRs), demonstrated by the fact that mutation of two major ITRs, ITR1A and ITR3C, partially abolishes the stimulation of Cryptococcus transmigration by inositol. Our results indicate that inositol plays a role in the traversal of Cryptococcus across the BBB, and showed that fungal cells can respond to inositol availability in the brain. Because we measured the transmigration rate by counting yeast cell numbers in the bottom chamber in our in vitro model, one concern in interpreting our results is whether addition of inositol to the bottom chamber has an effect on the proliferation of yeast cells. We measured the proliferation rate of Cryptococcus cells in HBMEC medium with or without addition of inositol, and observed a similar growth rate in all strains tested, confirming that the difference in cell numbers in the bottom compartment reflected a difference in yeast transmigration. Cryptococcus can use inositol as a carbon source although it is not a preferred source; Cryptococcus grows very slowly on medium containing inositol as the sole carbon source. Glucose is a much better carbon source for Cryptococcus. The HBMEC medium is enriched in nutrients and contains sufficient glucose for optimal fungal growth, which would explain why inclusion of 1 mM inositol did not affect cell growth. In addition, the association assays demonstrate that inositol promotes the association of Cryptococcus with the HBMEC monolayer. Because a better association often leads to increased transmigration, this effect could explain the increase in transmigration. In addition, because inositol is highly abundant in human and animal brains, for example, the astrocytes that directly interact with the BBB contain over 8 mM inositol that can be rapidly released [38], [39], we believe the inositol level in the brain side of the BBB is high and the inositol concentrations we used in in vitro assays are physiologically relevant. Our in vivo study using a murine model revealed a significant difference in numbers of CFUs recovered from mouse brains 24 hr post tail vein injection between mice infected by wild type versus the itr1aΔ itr3cΔ double mutant, confirming that the double mutant has a defect in brain infection. In the mutant, the observation of lower fungal burden in brains but not in lung during early infection suggests a defect in traversal across the BBB, consistent with our results from the in vitro system. We did observe a significant reduction in CFU in mutant-infected lungs 3 days post-inoculation, although the reduction was much less pronounced than in infected brains. Because the itr1aΔ itr3cΔ double mutant exhibits a moderate melanin defect [22], the reduced fungal burdens in the mutant infected lungs over time may be partially caused by reduced laccase activity, a known virulence factor [57]. However, these additional effects cannot account for the observed reduction in transmigration of cells carrying the double mutation, especially during early infection. Our confocal images of immunofluorescent staining further confirmed that Cryptococcus cells transmigrate into the brain by a mechanism that depend in part on inositol transporters, because strains lack of Itr1a and Itr3c have reduced fungal invasion in the brain. The outcome of our animal studies on transmigration could be influenced by the impact of the CNS environment on yeast growth and survival after transmigration. To address this biological artifact on transmigration analysis, we examined growth and/or survival of H99 and the double mutant within relevant biological fluids. Both in human CSF and in animal models of cryptococcal meningitis there was no apparent influence of the inositol transporters on yeast growth or survival in vivo, supporting their early impact on transmigration rather than direct effect on CNS compartmental survival and growth. However, we could not completely rule out that there may be a growth difference in certain microenvironment during brain infection, such as certain parts of the brain parenchyma, which is bypassed in the intracerebral injection model. In addition, we have shown that the presence of inositol transporters did reduce survival of the host in a murine intracerebral infection model [22]. Therefore, Itr1a and Itr3c are necessary for full virulence at this CNS site of infection. In fact, our preliminary results from a RNA-SEQ analysis of infected brains indicated that the double mutant caused more active host defense response than wild type, a potential explanation of the virulence attenuation of the double mutant during the establishment of the CNS cryptococcosis (our unpublished data). Therefore, inositol utilization likely influences on fungal transmigration across the BBB as well as subsequent disease development in the CNS. Furthermore, studies have shown that yeast cells cross the BBB early after infection, but that a dramatic increase occurs 24 hr post-injection [48]. Thus, a significant defect in the transmigration may not be detectable before 24 hr. We hypothesize that there is a threshold effect with respect to time or number of cryptococcal cells associated with brain microvascular endothelial cells before they can effectively penetrate the BBB in vivo. This is not without precedent; it has been reported that a bacteremia approaching 103 cells per milliliter in bloodstream is a prerequisite for meningitis-causing E. coli K1 to cross the BBB [58], [59]. In the in vitro BBB model used in this study, inositol added in the bottom compartment can diffuse to the top compartment to form an inositol concentration gradient. Our analysis showed that there was a time-dependent increase of inositol concentration in the top compartment with the presence of Cryptococcus, while very little increase of inositol level at the top without incubation with Cryptococcus. Furthermore, Cryptococcus incubation induces the modulation of the tight junction, as shown by the ZO-1 dislocation. These results suggest that the Cryptococcus-HBMEC interaction is required to modulate the permeability of the HBMEC monolayer to inositol. It has been reported that Cryptococcus invasion causes a modification of tight junctions [40]. In addition, inositol uptake by the host cells may contribute to the inositol level increase in the top compartment, since host cells also have inositol transporters. Our studies have shown that inositol alone is able to stimulate HBMEC and its uptake by HBMEC enhances the Cryptococcus-mediated phosphorylation of host signaling proteins and the permeability of dextran across the HBMEC monolayer in the presence of Cryptococci (Kim et al., unpublished). Alternatively, it is also possible that Cryptococcus and HBMECs may release some amount of inositol into the medium during their interactions. However, despite the increase in inositol permeability and ZO-1 dislocation, TEER measurement of the HBMEC monolayer was not significantly changed during incubation with Cryptococcus, suggesting that inositol diffuses through the monolayer without causing major damage to the BBB, which is consistent with previous reports [25], [47]. Further characterization on the inositol effects on the Cryptococcus-mediated modulation of HBMEC barrier is required to precisely understand the mechanisms for inositol diffusion. Because both Itr1a and Itr3c are major inositol transporters, Cryptococcus may be able to import inositol for its cellular function. It is possible that inositol uptake through inositol transporters modulates Cryptococcus cells to enhance their association with HBMEC and transmigration across the BBB. Our 2D-TLC analysis demonstrated that, in both the wild type and the itr1aΔ itr3cΔ double mutant, incubation with inositol significantly increases the amount of phospholipids produced (Fig. 7). Interestingly, among phospholipids, the amount of phosphoinositide (PI) detected after inositol treatment was much lower in the itr1aΔ itr3cΔ mutant than in the wild type. PI is a precursor in the production of inositol phospholipid and other downstream inositol metabolites that are essential for cellular function. Changes in PI production could profoundly affect Cryptococcus cellular signaling regulation and modification of cell surface dynamics. Phospholipids, including phosphatidylcholine (PC), have recently been identified to play a role in the capsule enlargement [60]. In our TLC results, PC production is highly induced by inositol treatment. The role of the capsule in Cryptococcus transmigration remains controversial in that some studies have demonstrated that capsule is involved [48], [61], while other reports have suggested a capsule-independent brain invasion process [26], [62]. Although we did not detect an obvious difference between the mutant and wild type in capsule size based on microscopy, it is possible that inositol may play a role in capsule structure or density, which may also influence cryptococcal transmigration and/or other functions during Cryptococcus-host interactions. Microarray analysis also revealed that inositol treatment induces upregulation of genes related to inositol catabolism and metabolism. Converting inositol to glucuronic acid by inositol oxygenases is the first step of the only known inositol catabolism pathway [50]. Cryptococcus has three genes encoding oxygenase enzymes and the expression of two of them is highly induced by addition of inositol, suggesting that the addition of inositol stimulates the production of glucuronic acid. The upregulation of beta-glucuronidase genes by inositol may indicate that yeast cells have increased carbohydrate turnover, resulting in overproduction of polysaccharides due to inositol supplementation. Whether inositol-converted glucuronic acid can be used to produce UDP-glucuronic acid, as one substrate of the polysaccharide capsule, remains uncertain. Previous studies have shown that UDP-glucuronic acid in C. neoformans is exclusively produced by the glycolytic pathway from glucose [63], [64]. However, it remains a possibility that under certain culture conditions, such as when inositol is used as a sole carbon source, glucuronic acid can also be a precursor of UDP-glucuronic acid. UDP-glucuronic acid is also one substrate for the synthesis of hyaluronic acid, a Cryptococcus ligand that interacts with CD44 on host cells during fungal invasion and transmigration of the BBB [36], [65]. Interestingly, we found that the gene encoding hyaluronic acid synthase, CPS1, is upregulated by inositol treatment and leads to the induction of hyaluronic acid production. The increased production of hyaluronic acid indicates that inositol may enhance the association between Cryptococcus and brain endothelial cells. Therefore, the changes induced in Cryptococcus through inositol uptake may play a positive role in fungal association as well as transmigration. The fact that the itr1aΔ itr3cΔ double mutant only showed a marginal reduction in hyaluronic acid production compared to wild type is consistent with results demonstrating that this mutant can invade the BBB and cause infection although less efficiently. Additional factors, such as the influence of inositol on the cellular lipid rafts, caveolin-1, cytoskeleton, etc., could also be in play to alter the transmigration. As well, additional redundant ITRs may contribute and/or partially compensate for losses in the double mutant background. A mutant strain lacking the whole ITR gene family would be valuable for understanding the complete role of ITRs. We are in the process of generating such a mutant strain. In addition, inositol itself is an important signaling molecule that affects many biological functions. Inositol stimulates the long distance sodium uptake from root to leaf in plants [66], and many species of caterpillar can sense inositol on plants for feeding [67], [68]. Cryptococcus may sense the concentration gradient of inositol and promote its association with endothelial cells and increase the transcytosis in response to higher inositol concentration, a phenomenon similar to chemotropism. The role of inositol as a potential chemoattractant to stimulate Cryptococcus to associate and penetrate the BBB needs to be further investigated.. Overall, we have demonstrated that brain inositol affects fungal cells to promote the traversal across the BBB as presented in our model (Fig. 10). Cryptococcal cells disrupt tight junctions of the BBB allowing leakage of inositol from the brain to the bloodstream to generate an inositol concentration gradient; Cryptococcus senses the presence of inositol gradient via inositol sensors that remain to be identified, and takes up inositol using these inositol transporters. Inositol import leads to modifications in fungal physiology such as hyaluronic acid production. The inositol-mediated changes on fungal cells lead to enhanced yeast binding to and transmigration of the BBB, resulting in cryptococcal brain infection and disease development. This model suggests that there is a complex interaction between fungal cells and host cells that involves host inositol and fungal inositol transporters. Because the transmigration and subsequent disease establishment in the brain is a continuous process that is difficult to separate definitely, we cannot rule out the possibility that in addition to its involvement in the fungal transmigration, inositol utilization could also play an important role in the establishment of cryptococcal meningitis in the CNS. In fact, this hypothesis is supported by our unpublished data on the difference in host immune response between brains infected by the wild type and the itrΔ mutant. Despite the importance of host inositol, we do aware that inositol may be only one of multiple factors that affect the development of the CNS cryptococcosis. Additional factors and mechanisms likely also exist to stimulate the fungal cell transmigration and the establishment of fungal infection in the brain. In fact, several factors, such as capsule [48], [61], urease [30], [31], phospholipase B1 [32], [33], and host copper utilization [69], have been reported to play a role in cryptococcal dissemination and/or brain infection. Nevertheless, our discovery of the involvement of host inositol and fungal inositol transporters in the development of cryptococcal CNS infection leads to a better understanding of this complex host-pathogen interaction during the development of cryptococcal meningitis. The animal studies conducted at Duke University and University of Medicine and Dentistry of New Jersey (UMDNJ) were in full compliance with all of the guidelines set forth by the Institutional Animal Care and Use Committee (IACUC) and in full compliance with the United States Animal Welfare Act (Public Law 98–198). The Duke and UMDNJ IACUCs approved all of the vertebrate studies. The studies were conducted in facilities accredited by the Association for Assessment and Accreditation of Laboratory Animal Care (AAALAC). C. neoformans var. grubii (serotype A) H99 and its isogenic mutant strains (itr1aΔ, itr3cΔ, itr1aΔ itr3cΔ) have been previously described [22], [23]. C. neoformans var. neoformans (serotype D) strain B-3501 was kindly provided by Dr. Kyung J. Kwon-Chung (NIAID), whereas Candida albicans strain ATCC90028 was kindly provided by Dr. David Perlin (UMDNJ). Yeast cells were grown on YPD (1% yeast extract, 2% peptone, 2% glucose) agar plates and synthetic (SD) medium at 30°C and stored at 4°C until use. The Anti-GXM antibody 18B7 was kindly provided by Dr. Arturo Casadevall (Albert Einstein College of Medicine). Horseradish peroxidase-conjuagated anti-mouse and anti-rabbit antibodies were obtained from Invitrogen (Grand Island, NY). Dinitrophenol and inositol isomers were purchased from Sigma-Aldrich (St. Louis, MO). Primary isolates of HBMEC were cultured as previously described [70]. HBMEC were routinely grown in RPMI 1640 supplemented with 10% heat-inactivated fetal bovine serum, 10% Nu-serum, 2 mM glutamine, 1 mM sodium pyruvate, penicillin (100 units/ml), streptomycin (100 µg/ml), essential amino acids, and vitamins. The cells were incubated at 37°C in a humidified incubator with 5% CO2. Before each experiment, the culture medium was replaced with experimental medium containing Hams-F12/M199 (1∶1, v/v), supplemented with 5% heat-inactivated fetal bovine serum. Total yeast cells associated with HBMEC were determined as previously described [41], [71]. Briefly, HBMEC were grown in 24-well tissue culture plates (or Transwell tissue-culture inserts with a pore diameter of 8.0 µm (Corning Costar)) until confluence. Inocula of 105 Cryptococcus cells in experimental medium were added to each well (or top compartments in Transwells), and then incubated for 3 hr at 37°C. Free unbound yeast cells were removed by washing 3 times with PBS. The HBMEC were lysed with sterile distilled water and the lysates were diluted and plated onto sheep blood agar plates. The colonies were counted and results were presented as the total number of yeast cells per monolayer. Each set was triplicated and repeated at least three times independently. The in vitro human blood-brain barrier (BBB) model was generated and used for fungal transmigration assays as previously described [70]. HBMEC were seeded on Transwell polycarbonate tissue-culture inserts with a pore diameter of 8.0 µm (Corning Costar) and cultured until their transendothelial electrical resistance (TEER) reached over 350 Ω/cm2, as measured by an Endohm volt/ohm meter in conjunction with an Endohm chamber (World Precision Instruments). The medium was replaced with experiment medium before each experiment. Yeast cells were washed with phosphate-buffered saline (PBS) and resuspended in HBMEC culture medium. 105 Cryptococcus cells were added to the top compartment and then incubated at 37°C. At 3, 6, and 9 hr, the medium in bottom compartments was collected and immediately replaced with fresh medium. Fungal cell numbers in the collected medium were addressed by CFU counts to determine the number of transmigrated viable yeast cells. To determine the specificity of myo-inositol, the transmigration assay was performed in the presence of myo-inositol, scyllo-inositol, D(+) galactose or D(+) mannose (1 mM each) in the bottom compartments prior to addition of Cryptococcus cells (105) in the top compartment of Transwells. Results are presented as the total number in the bottom chamber. Each set was triplicated and repeated three times independently. The statistical analysis of the data from our in vitro studies was done with a two-tailed Student t test. Statistical significance was determined at P<0.001. The inositol concentration in medium was determined as previously described with minor modifications [46]. Briefly, the medium containing inositol collected from the top compartments of the BBB model was incubated with hexokinase to reduce interference from glucose by phosphorylation. The mixture was then incubated with 4.1 U/ml myo-inositol dehydrogenase for 15 min. Subsequently, 100 µl medium was mixed with an equal volume of detection reagent. The inositol concentration was determined by measuring optical density at 492 nm with a microplate reader (BioTek, Winooski, VT). Each assay was triplicated and repeated three times independently. Virulence of the C. neoformans strains was assessed using both a murine intravenous infection model and a rabbit CSF model of cryptococcosis as previously described [22], [72], [73]. For virulence study in a murine intravenous injection model, Cryptococcus strains were grown at 30°C overnight and cultures were washed twice with 1× PBS buffer by centrifugation, and resuspended at a final concentration of 5×105 cells/ml. Groups of 15 female A/JCr mice (NCI-Frederick, MD) were used for each infection. Mice were infected with 5×104 yeast cells of each strain in 100 µl PBS through tail vein injection [22], [74]. At each time point, 3 mice infected with either H99 or the mutant were sacrificed after 1, 6, 24, 48, and 72 hr post-infection. Fungal burden in infected brains was analyzed by CFU counts. Data from the murine experiments were statistically analyzed between paired groups using the long-rank test and the PRISM program 4.0 (GraphPad Software) (P values of <0.01 were considered significant). For the murine intracerebral injection model, mice were sedated with a Ketamine-Xylazine combination and the top of the head was sterilized using antiseptic. A total of 500 yeast cells in 50 µl were directly injected into the cerebrum as previously described [22]. For Rabbit infection, male New Zealand White (NZW) rabbits were treated with cortisone via daily injection and intrathecal inoculation into the subarachnoid space with 108 cells of each C. neoformans strain (3 rabbits per group). Fungal burden in brain was analyzed by CFU counts. H99 overnight culture was washed with dH2O twice. Equal amount cells were inoculated on SD medium with or without 5 mM inositol and incubated for 24 hr before cells were collected for total RNA purification. Total RNAs were extracted using Trizol Reagents (Invitrogen) and purified using with Nucleospin RNA cleanup kit (Clontech, Mountain View, CA). Cy3 and Cy5-labeled cDNA were generated by incorporating amino-allyl-dUTP during reverse transcription of 5 µg of total RNA as described previously [75] and competitively hybridized to a JEC21 whole-genome array generated previously at Washington University in Saint Louis. After hybridization, arrays were scanned with a GenePix 4000B scanner (Axon Instruments) and analyzed by using GenePix Pro version 4.0 and BRB array tools (the National Cancer Institute, http://linus.nci.nih.gov/BRB-ArrayTools.html) as described previously [76]. The original microarray data was provided as supplementary file (Table S1) and also was submitted to GEO database (GSE41211). To confirm the microarray results, we measured the mRNA levels of 6 genes under different conditions via quantitative real-time PCR (qPCR). First strand cDNAs of the purified RNAs were synthesized using a Superscript III cDNA synthesis kit (Invitrogen, Grand Island, NY) following the instructions provided by the manufacturer. Expression of candidate genes and GAPDH were analyzed with the comparative CT method using SYBR green QPCR reagents (Clontech) as described previously [23]. Overnight cultures of Cryptococcus were washed with dH2O twice and re-inoculated to 5 ml SD with or without 5 mM inositol and shaken for 2 hrs; 25 µCi/ml of 32P-phosphorus was added and the cultures were incubated with shaking overnight. Yeast cells were collected and concentrations were determined by hemocytometer counts. Steady-state labeling of phospholipid with 32Pi and lipid extraction were performed as described previously [77]. Lipids were dried in a SpeedVac apparatus, and resuspended into 500 µl chloroform. Five microliter aliquots were removed to measure the radioactivity in a scintillation counter. The remaining lipids were dried and frozen at −80°C. The individual phospholipids were resolved using two-dimensional silica gel TLC plates (EMD, Rockland, MS) using chloroform/methanol/ammonium hydroxide/water (90∶50∶4∶6, v/v) as the solvent system for dimension one and chloroform/methanol/glacial acetic acid/water (32∶4∶5∶1, v/v) as the solvent system for dimension two. The identity of the labeled lipids on thin-layer chromatography plates was confirmed by comparison with standards after exposure to iodine vapor. Radiolabeled lipids were visualized by phosphorimaging analysis with a Storm PhosphorImager (GE, Pittsburgh, PA). The relative quantities of labeled lipids were analyzed using ImageQuant software. The hyaluronic acid enzyme-linked immunosorbent assay (ELISA) kit (Corgenix, Denver, CO) was used to assay hyaluronic acid. The ELISA for hyaluronic acid production was followed the method as previously described with a few modifications [53]. Yeast cells (107 cells) in the exponential-growth phase were incubated in individual wells at room temperature to trap the surface polysaccharide. After 60 min, the wells were washed with washing buffer carefully according to the manufacturer's instructions. A second solution containing a hyaluronic acid-binding protein-HRP conjugate was added to the wells and incubated for 30 min before adding substrates. The intensity of the resulting color was measured in optical density units with a spectrophotometer at 450 nm. The concentrations of hyaluronic acid were calculated by comparing the absorbance of the sample against a reference curve prepared from the reagent blank and hyaluronic acid reference solutions. The statistical significance was assessed by a 2-pair student t-test. HBMECs were grown on coverslips coated with type-I collagen from rat tail (Millipore, Billerica, MA) until confluence. After incubation with C. neoformans (H99) for 1 hr, HBMECs were washed three times with PBS and processed for immunofluorescent staining as previously described [78]. Briefly, the cells were fixed with 4% paraformaldehyde for 30 min, permeabilized with 0.5% Triton X-100 for 5 min and then incubated with ZO-1 antibody, followed by AlexaFluor 488-conjugated secondary antibody (Invitrogen) to visualize. The coverslips were mounted with Vectashield mounting solution with DAPI (Vector Laboratory, Burlingame, CA) and observed with a Nikon fluorescence microscope. Images were taken with a MetaMorph Microscopy Automation & Image Analysis Software. Analysis of fungal infection in mouse brain using confocal fluorescent microscope was performed as reported previously with modifications [79]. Thirty-fifty microns paraffin brain tissue sections were de-paraffined, subjected to antigen retrieval and permeabilized with 0.01% Triton X-100. Tissue sections were washed three times in PBS and incubated in blocking solution (5 mM EDTA, 1% fish gelatin, 1% essentially Ig-free BSA, 2% human serum and 2% horse serum) for 60 min at room temperature. Tissue sections were incubated in the proper diluted primary antibody (anti-GFAP, 1∶500, or anti-GXM, 1∶1000, provided by Dr. Arturo Casadevall, Albert Einstein College of Medicine) overnight at 4°C. Samples were washed several times with PBS at room temperature and incubated with the appropriate secondary antibodies conjugated to FITC or Cy3 for 2 hr at room temperature, followed by another wash in PBS for 1 hr. Tissue sections were then mounted on slides and stained with DAPI, and the cells were examined by a SP2 confocal microscopy (Leica). To identify and observe the tissue lesions induced by Cryptococcus, optical sections were acquired and reconstituted to focus in the lesion in these thick tissue sections. Images were analyzed with the NIS Elements Advance Research Program (Nikon). Antibody specificity was confirmed by replacing the primary antibody with a non-specific myeloma protein of the same isotype or non-immune serum.

10.1371/journal.pcbi.1000620

Design Principles for Ligand-Sensing, Conformation-Switching Ribozymes

Nucleic acid sensor elements are proving increasingly useful in biotechnology and biomedical applications. A number of ligand-sensing, conformational-switching ribozymes (also known as allosteric ribozymes or aptazymes) have been generated by some combination of directed evolution or rational design. Such sensor elements typically fuse a molecular recognition domain (aptamer) with a catalytic signal generator (ribozyme). Although the rational design of aptazymes has begun to be explored, the relationships between the thermodynamics of aptazyme conformational changes and aptazyme performance in vitro and in vivo have not been examined in a quantitative framework. We have therefore developed a quantitative and predictive model for aptazymes as biosensors in vitro and as riboswitches in vivo. In the process, we have identified key relationships (or dimensionless parameters) that dictate aptazyme performance, and in consequence, established equations for precisely engineering aptazyme function. In particular, our analysis quantifies the intrinsic trade-off between ligand sensitivity and the dynamic range of activity. We were also able to determine how in vivo parameters, such as mRNA degradation rates, impact the design and function of aptazymes when used as riboswitches. Using this theoretical framework we were able to achieve quantitative agreement between our models and published data. In consequence, we are able to suggest experimental guidelines for quantitatively predicting the performance of aptazyme-based riboswitches. By identifying factors that limit the performance of previously published systems we were able to generate immediately testable hypotheses for their improvement. The robust theoretical framework and identified optimization parameters should now enable the precision design of aptazymes for biotechnological and clinical applications.

Aptamers are nucleic acids that bind their cognate ligands (ranging from metal ions to small molecules to proteins) specifically and tightly. Through rational design and/or directed evolution, aptamers can be engineered into allosteric nucleic acids whose conformations can be regulated by their ligands. Aptamer beacons, aptazymes, and riboswitches all undergo ligand-dependent conformational changes, and have been adapted to signal the concentration of their ligands. However, there is currently no model that can be used to predict how the energetics of conformational change affects signaling, either in vitro or in vivo. We have developed a model that identifies what parameters can be optimized to best yield signals. By focusing on these parameters, it should be possible to more readily design or select more effective conformation-switching nucleic acid biosensors.

Nucleic acid binding species (aptamers) have emerged as a powerful tool for molecular recognition, and have begun to be widely adapted as biosensors, in drug-delivery systems, and as regulatory elements that control gene expression [1]–[4]. Naturally occurring nucleic acid regulatory elements, riboswitches, have been discovered in a variety of organisms and control the expression of a wide range of genes [5]. One of the major advantages of aptamers over their protein counterparts is that they can be easily coupled to other functional RNAs based largely on secondary structural considerations in order to generate allosteric constructs. To a large extent aptamer-based biosensors (both in vitro and in vivo) can be classified into two major categories: (i) those in which the aptamer binding influences the hybridization state of other nucleic acids (for in vitro examples see [6],[7]; for in vivo examples, see [8]), and (ii) those in which aptamer binding influences the catalysis of a ribozyme (for in vitro examples, see [9]–[11]; for in vivo examples, see [12]–[15]. These allosteric ribozymes derived from aptamers are also known as aptazymes. While there are numerous empirical examples of aptazymes operating as biosensors and regulatory elements, quantitative analyses of aptazyme performance and the development of design principles for aptazymes have seldom been attempted and are largely incomplete [10],[16]. Recently, Beisel and Smolke developed a similar model for riboswitch function [16]. However, only qualitative trends were reported. For example, while it was concluded that “a design that is biased toward forming the disrupted-aptamer conformation will generally increase the dynamic range …(but) require higher ligand concentrations to modulate protein level,” the more useful quantitative relationship between dynamic range of activity and ligand sensitivity that should enable rational design was not described. Similarly, the impact of fundamental kinetic parameters such as the ribozyme cleavage rate constant and mRNA degradation rate constant on the behavior of riboswitches was not analyzed. Additionally, those numerical solutions that were given were based on arbitrary parameters. For all of these reasons it is unclear what parameters need to be measured for the quantitative prediction of riboswitch function. It is also unclear how and to what extent the parameters can be optimized for improved function. To establish a better quantitative understanding of aptazyme-based biosensors and riboswitches, we analyze a two-state model for aptazyme function and illustrate: (i) the quantitative relationship between the dynamic range of activity and ligand sensitivity; (ii) the variables that limit aptazyme function; (iii) the minimal set of readily measurable parameters that are necessary and sufficient to quantitatively predict aptazyme function; and (iv) strategies to design optimal aptazyme-based biosensors for both in vitro and in vivo applications. In addition, we apply this model to published data for a previously engineered riboswitch system [14] and show that this system is severely limited both by slow ribozyme cleavage relative to mRNA degradation and likely by the intracellular concentration of theophylline. The ability to predict the secondary structure of functional RNA molecules has made it possible to rationally design allosteric ribozymes. Aptamer secondary structures are superimposed upon or swapped with portions of ribozyme secondary structures (Figure 1A), and interactions between the two domains are often controlled by junction sequences (so-called communication modules). One commonly used strategy to design ligand-activated aptazymes can be described as ‘binding assisted stem-formation’ (Figure 1B) in which a weak but functionally important stem that is shared by the aptamer and the ribozyme is stabilized by ligand-binding [12],[13]. Other design strategies include ‘slip structures’ (Figure 1C; [9]) and ‘strand replacement’ (Figure 1D; [14],[15]). In these latter strategies the ligand-induced stabilization of the aptamer helix causes a conformational change in the secondary structure of the ribozyme that either promotes or inhibits catalysis. Taken together, all of these strategies assume a two-state model for the aptazyme in which one of the states is stabilized by ligand-binding. To garner better insights into how to design aptazymes, we will attempt to model the interrelationships between aptazyme conformational change, ligand-binding, and catalysis. In this way we can separate intrinsic variables (including the aptamer∶ligand affinity and the ribozyme catalytic rate constant) from extrinsic or ‘engineerable’ variables (including the equilibrium constant between the two conformers). While the catalytic rates of the less active conformer and the more active conformer are also extrinsic variables, they should almost always be minimized (to zero if possible) and maximized (to the rate of the ribozyme sans aptamer if possible), respectively. For simplicity, we develop our analyses with self-cleaving aptazymes, but the model should be generalizable to aptazymes with other catalytic activities. The general model for ligand-modulated ribozymes is similar to that for allosteric protein enzymes (Figure 2A). In this model, the aptazyme can assume two interchangeable conformations A and B with internal equilibrium constant Kint (see Text S3 for a summary of terms), each of which has particular (but different) ligand-binding affinities defined by association constants Ka(A) and Ka(B), respectively, and particular (but different) cleavage activities defined by cleavage rate constants kCle(A) and kCle(B), respectively. Since in most cases it is the local structure of the catalytic core (as opposed to the ligand-binding site) that determines the catalytic activity of the aptazyme, it is assumed that the aptazyme-ligand complexes AL and BL have the same cleavage rate constants as the unbound aptazymes A and B, respectively. Furthermore, we only consider the situation where all four species (A, B, AL and BL) are in equilibrium at the start of the reaction. When the conformer that possesses higher ligand-binding affinity also has higher catalytic activity the aptazyme is called ligand-activated; when the conformer that possesses higher binding affinity has lower catalytic activity the aptazyme is called ligand-inhibited. In general we assign conformation B to have the higher ligand-binding affinity (Ka(B)>Ka(A)). Thus, ligand can be thought to thermodynamically shift the A conformer towards B. Since most two-state models for allosterism assume that the ligand primarily influences the population of catalytically inactive and active conformations, we assume that the less catalytically active conformer has zero activity and the more catalytically active conformer has the same cleavage rate constant as the ribozyme sans aptamer (denoted as kCle). Formally,and Another simplifying assumption is that the complex BL is much more thermodynamically stable than AL, and thus we can ignore the existence of AL and reduce the model to the path outlined in green in Figure 2A. This reduced model assumes that the A conformer must spontaneously refold into the B conformer in order to bind the ligand and thus excludes ligand-induced refolding of the aptazyme. This reduction is valid when two conditions are met: (i) the energy barrier between A and B is not much higher than that between AL and BL, so that aptazyme refolding does not rely on the ligand as a catalyst; and (ii) when the aptazyme is bound to the ligand the aptazyme almost exclusively assumes the BL conformation. We will use this reduced model in the following analyses. The in vitro performance of a self-cleaving aptazyme is usually evaluated by plotting the first-order apparent cleavage rate constant (kapp; the initial cleavage rate divided by the total concentration of aptazyme) against the total ligand concentration ([Ltot]). As a starting point of our model, we show how kapp, which is in fact contributed to by all three aptazyme conformations, is determined by the variables shown in Figure 2A. Assuming that ligand-binding is much faster than aptazyme cleavage ([L]kon(B) + koff(B)≫kCle(B)) the initial cleavage rate constant should directly reflect the initial fraction of each of the three conformers A, B, and BL:(1)If ligand-binding is slow relative to cleavage, the apparent rate constant would reflect the rate of binding (the rate limiting step) instead of cleavage. Based on the assigned definitions for parameters (see Text S1 for derivation) the fraction of A can be calculated to be:(2)and the total fractions of B and BL are:(3)where is relative ligand concentration, defined as the ligand concentration divided by the dissociation constant (Kd or ) of the aptamer domain. The introduction of relative ligand concentration means that Kd is only a scaling factor for ligand concentration. In other words, two aptazymes with the same Kint but different Kd values would be indistinguishable in terms of their performance with respect to relative ligand concentrations. In the absence of ligand, fB+BL and fA equal to and , respectively. Thus the ratios and are the fraction of cleavage-competent conformers in the absence of ligand for ligand-activated aptazymes and ligand-inhibited aptazymes, respectively. We term these ratios ‘cleavage tendency’ and denote them as ω. Formally: It should be noted that the relationship between ω and Kint is dependent on the type of the aptazyme (ligand-activated or ligand-inhibited). When the aptazyme type is specified, ω can be used interchangeably with Kint. Since in many cases the equations are in simpler form when ω is used instead of Kint, we will primarily use ω in the following derivations and analyses. From equations (1∼3) and earlier assumptions, the relationship between kapp and the relative ligand concentration are:or(4)for ligand-activated aptazymes, and:or(5)for ligand-inhibited aptazymes. The kapp-vs- curve is an increasing hyperbola for ligand-activated aptazymes. The relationship between the parameters that describe the hyperbola (highest value, lowest value, and half-value concentration) and the model parameters (ω and kCle) can be determined by rewriting equation (4) as:where kapp(min) and kapp(max) are the minimal and maximum apparent cleavage rate constants. These rate constants are reached in the absence of ligand and at a saturating concentration of ligand, respectively. is the relative ligand concentration at which the kapp is half-way between kapp(min) and kapp(max). As a result:(6)(7)(8) According to the definition of relative concentration, is dimensionless and scales relative to the Kd of the aptamer domain, the absolute (with unit of a concentration) can be calculated with the equation:It is noteworthy that EC50 is often regarded as the ‘apparent Kd of the aptazyme’ and can be confused with Kd. In fact, the Kd is an intrinsic variable reflecting the affinity between the aptamer and the ligand, while EC50 is design-dependent. From equation (8) it can be seen that is always greater than Kd and is inversely correlated with ω, since the ligand binding-competent conformation B is only a fraction of the total aptazyme population and a smaller ω means this conformation is proportionately disfavored. In addition to , another important parameter for describing the performance of a ligand-activated aptazyme is the fold-activation of the cleavage rate constant when ligand concentration increases from 0 to infinite. We denote this fold-activation as which is defined as:Comparing equations (6) and (7) it is obvious that for ligand-activated aptazymes,(9)which means that the maximum fold-activation is solely determined by the cleavage tendency of the aptazymes. In order to engineer aptazyme that have a higher , one must minimize ω, i.e. the cleavage-competent conformation should be disfavored in the absence of ligand. For example to achieve a >102-fold activation in the presence of ligan ω should also be no greater than 10−2, which in turn means that the free energy of conformation A should be disfavored by at least 2.8 kcal/mole (at 37°C) relative to conformation B. However, a low value of ω would also increase the concentration of ligand that was required to fully activate the aptazyme. This can be seen by comparing equations (8) and (9), yielding:or(10) In other words, high sensitivity (low EC50) and a large dynamic range of kapp (high ) cannot be obtained simultaneously (Figure 2B). Conversely, if an aptazyme displays a mediocre and also has a large fold-activation it can be inferred that the aptamer domain may actually have a very high affinity for its ligand. For example, a lysozyme-dependent L1-ligase previously selected by Robertson and Ellington [11] exhibits an EC50 of 1.5 µM but has a 3100-fold activation in the presence of saturating concentration of ligand (which means ≥3100). According to equation (10), the aptamer domain of this aptazyme may have a Kd as low as 500 pM. To reach the full theoretical dynamic range of kapp, the ligand concentration should vary from 0 to infinite, which is of course impossible. The upper limit of the realistic dynamic range of kapp for a ligand-activated aptazyme is determined by the kapp at the highest possible concentration of ligand. Therefore, when designing aptazymes it is important to consider the fold-activation of the cleavage rate constant when ligand concentration increases from 0 to its highest possible concentration. We denote this fold-activation as and formally define it as:where the is the highest possible relative ligand concentration. Since decreasing cleavage tendency is a double-edged sword in that it increases but at the same time requires higher ligand concentration to achieve half activation, it is important to find the cleavage tendency that gives optimal aptazyme performance (the highest ). To find the optimal cleavage tendency, it is useful to determine the explicit expression of as a function of ω, which is:(11)Interestingly, from this equation it is clear that for any >0, increases monotonically as the cleavage tendency ω decreases, as shown in (Figure 2C, left panel). In other words, it is always beneficial to have a lower cleavage tendency when the goal is to design the aptazyme to maximize . Practically, the only negative effect of engineering small cleavage tendencies in aptazymes is that the absolute value of is small, and thus the rate of cleavage and signal generated by the aptazyme may be small. Therefore, as a practical guideline for designing ligand-activated aptazymes as in vitro biosensors the cleavage tendency should be minimized as long as the value still falls within a range that is readily detected by a given assay. The kapp-vs- curve for a ligand-inhibited aptazyme is a decreasing hyperbola, whose descriptor can be solved by rearranging equation (5) to the form:yielding:(12)(13)(14) Here the definition of the fold-inhibition over the theoretical dynamic range of kapp () is problematic since the theoretical lower limit of kapp is 0 and therefore for a ligand-inhibited aptazyme would be infinite. The value (now defined as ) will be dependent on the design of the aptazyme (i.e., the choice of the cleavage tendency ω) and on the highest available concentration of ligand (). Because the inhibited aptazyme is hyperbolically controlled by the ligand (see Figure 2D), the lower realistic limit of kapp will be very hard to reach, and the range of kapp values for ligand-inhibited aptazymes will be heavily dependent on the ratio of to . A low will be crucial if the highest possible concentration of ligand is limited or if the intrinsic affinity of the aptamer domain is low. According to equation (14), a lower should be engineered by decreasing ω. However, by comparing equations (13) and (14) we find:(15)which in turn implies that lowering will decrease the upper bound on possible kapp values (Figure 2D). Once again there is a compromise between ligand sensitivity and the dynamic range of activity. Again, to find the cleavage tendency that yields the highest for a given , the expression of as a function of ω should be considered. This expression is:(16)Interestingly, as cleavage tendency ω increases from 0 to 1, decreases linearly from 1+ to 1 (Figure 2C, right panel). Consequently, when designing ligand-inhibited aptazymes as in vitro biosensors, it is also always beneficial to choose a low cleavage tendency as long as is still readily detectable. In summary, for both ligand-activated and ligand-inhibited aptazymes there are trade-offs between ligand sensitivity and the dynamic range of activity, reflected by equations (10) and (15), respectively. However, when attempting to maximize it is always a good strategy to choose a low cleavage tendency, as shown by equations (11) and (16) and Figure 2C. Aptazymes can be inserted into mRNAs in order to regulate their stabilities and translation efficiencies, thereby functioning similar to natural riboswitches in vivo. In such applications, aptazyme regulation will of necessity be further modulated by the dynamic processes surrounding RNA metabolism, including transcription, processing, transportation, translation and degradation. In addition, the most readily observed signals will be steady state mRNA or protein concentrations, instead of cleavage rate constants. The most straightforward strategy for adapting aptazymes to gene regulation is to engineer a drug-responsive cleavase (such as a hammerhead aptazyme) to target a particular mRNA. However, despite decades of effort, gene regulation based on trans-cleaving ribozymes has proven largely unsuccessful. Gene regulation via ligand-responsive ribozyme was paradoxically first demonstrated in a natural system, where a novel ribozyme located at the 5′ UTR of the glmS gene of B. subtilis was found to self-cleave primarily in the presence of GlcN6P [17]. This cleavage has been shown to destabilize glmS mRNA and thus to down-regulate glmS expression [18]. Interestingly, biochemical study revealed that glmS ribozyme is not an allosteric ribozyme per se, since GlcN6P does not allosterically regulate glmS ribozyme but rather serves as a cofactor which directly contributes to catalysis [19]. More recently, the engineering of artificial riboswitches based on cis-cleaving aptazymes has achieved some success. By connecting the anti-theophylline or anti-tetracycline aptamers to the tobacco ringspot virus (TRSV) HHRz via rationally designed or selected communication modules, Win and Smolke engineered aptazymes that, when inserted to the 3′ UTR of the GFP gene, could regulate GFP expression in yeast in response to theophylline or tetracycline concentration [14]. The reported dynamic range of GFP expression level was 20∼25-fold (Figure 2 of [14]). However, closer inspection of the raw data provided in the supplementary material (Figure S13 of [14]) showed that the dynamic range of GFP expression level was actually much lower. Among all the aptazyme constructs that were designed and tested, most displayed only ∼1.5-fold regulation and the best ones displayed ∼2.5-fold regulation. The discrepancy between the interpretation and the data was due to redefinition of the word ‘fold’ by the authors. Although the word ‘fold’ is generally used to express the ratio of two quantities, Win and Smolke used ‘fold’ as a unit of absolute quantity of GFP expression [14]. For example, the GFP expression level from an unengineered plasmid was defined as ‘50 fold.’ Therefore, when the GFP expression level from an engineered plasmid changed from ‘20 fold’ in the absence of theophylline to ‘43 fold’ in the presence of theophylline, a dynamic range of ‘(43−20 = ) 23 fold’ could be claimed. Most researchers would instead estimate the dynamic range to be (43/20 = ) 2.2-fold. Win and Smolke have also reported that multiple aptazymes inserted into the 3′ UTR could act as logic gates for gene expression, but the raw data necessary to evaluate these claims were not immediately available [15]. These designs were of necessity eukaryote-specific, since the 3′ polyA:5′ cap interaction is crucial for efficient protein translation. A prokaryote-specific system has been developed by Wieland et al. in which the ribosome-binding site (RBS) of a reporter gene was embedded in stem I of the Schistosomal HHRz, such that the self-cleavage of the HHRz liberated the RBS for translation initiation [12],[13]. Through rational design and genetic screening, a theophylline-responsive aptazyme that exhibited 10-fold regulation of the expression of the reporter gene was generated. The fold-regulation achieved by these authors (1.2- to 10- fold) are far smaller than those that have been routinely demonstrated in vitro (102-∼104- fold ). To explain this discrepancy, we will explore a simple kinetic model. In this model, the eukaryotic-specific system, where an aptazyme is placed within the 3′ UTR of a mRNA, will be used. That said, it should be noted that self-cleaving HHRzs placed within the 5′ UTR can abet even stronger inhibition of gene expression [20], but such a model would be inherently more challenging because it would have to take into account the continuous scanning by the pre-initiation complex. We first model how gene expression can be inhibited by a constitutively active, self-cleaving ribozyme (Figures 3A and 3B). In these models, we assume that the steady-state concentration of a protein is proportional to the steady state concentration of its intact mRNA. In contrast, mRNA with a cleaved 3′ UTR is assumed to have a negligible translation efficiency or is rapidly degraded [21]. In the absence of ribozyme cleavage (Figure 3A) the steady state concentration of mRNA ([R]ss) is . When a constitutively active self-cleaving ribozyme is inserted to the 3′ UTR of the mRNA (Figure 3B), the steady state concentration of intact mRNA should depend on its cleavage rate, as well as on the transcription and degradation rates, specifically:(17) If we define the relative steady-state intact mRNA concentration without ribozyme as 1, then the relative steady-state intact mRNA concentration of an mRNA that harbors a ribozyme is:(18)where is the ratio of cleavage rate constant to the spontaneous degradation rate constant. The extent to which gene expression can be inhibited by an inserted ribozyme is directly determined by this ratio D, which implies that the rate of spontaneous degradation of mRNA also directly influences how much inhibition a given ribozyme can potentially achieve [22]. As before, we assume that the inactive conformer in a two-state model is completely inactive, and that the active conformer has the same cleavage rate constant as the ribozyme sans aptamer. The kinetic model for gene regulation via ligand-activated self-cleavage is shown in Figure 3C. For simplicity only the 3′ UTR is shown. In this model, mRNA is transcribed from the ‘gene’ (G) with a zero-order rate constant of vTxn. The nascent transcript (I) can fold into either aptazyme conformer [cleavage-incompetent conformer (A) or cleavage-competent conformer (B)] with folding and unfolding rate constants kFoldA, kFoldB and kUnA, kUnB, respectively. The B (but not A) conformer can also bind the ligand L to form aptazyme∶ligand complex BL which has the same catalytic activity as B (kCle). The second-order association rate constant and first-order dissociation rate constant are denoted as kOn and kOff, respectively. Under this model (see Text S2 for derivation), the relationship between steady-state relative concentration of intact mRNA (including I, A, B and BL) and the concentration of total ligand L ([Ltot]) is expressed in the following equation:(19)where(20)(21)and(22) This definition of relative concentration is similar to our earlier definition of relative ligand concentration, except that in this case Kd is replaced by:which we term the apparent dissociation constant and denote as . is similar in form to Kd () and will have a similar value to Kd when the dissociation rate constant of the ligand∶aptamer complex (kOff values typically 10−3 to 101 s−1) is much higher than the cleavage rate constant of the ribozyme (kCle values typically 10−2 to 1 s−1). However, it may also have a larger value than Kd when kOff is comparable to or lower than kCle. Again, is the scaling factor for ligand concentration. Since the degradation rate constant of mRNA in eukaryotic cells is much slower (by up to 10 orders of magnitude; [23]) than structural transition, ligand dissociation, and ribozyme cleavage rates, α and β should have values similar to the equilibrium constants for the reactions I↔A () and I↔B (). Notably, β can be treated as a constant although it is actually a function of ligand concentration. When β is treated as a constant, is similar to Kint in Figure 2 and consequently is equivalent to the cleavage tendency ω. Moreover, since the folded state is typically of lower energy (and thus more occupied) than the intermediate (I) or unfolded state, α is usually much greater than 1. Given these two conditions, the equation (19) can be written as:(23) It is interesting that equation (23) can be simply obtained by replacing kCle in (17) with kapp in (4). This suggests that the equation for the function of aptazymes in vitro (4) can be used for aptazymes in vivo, with the only significant error coming when kCle is on the same order as or larger than kOff, which would in turn lead to a significant difference between and Kd. The model for a ligand-inhibited self-cleaving ribozyme is diagramed in Figure 3D. The primary difference from the model for a ligand-activated aptazyme (Figure 3C) is that now only the conformer A, instead of both B and BL, can undergo self-cleavage. Given the parameters in Figure 3C, the relationship between relative steady-state concentration of intact mRNA ([R]Rel) and ligand concentration ([Ltot]) is (see Text S2 for derivation):(30)where:(31)(32)(33) In this case the apparent dissociation constant () is closer in value to Kd since kCle does not appear in the definition of . As before, α and β are similar to the equilibrium constants for the reactions I↔A () and I↔B (), respectively, and β can be treated as a constant. Given that , when α is much greater than 1 then equation (30) can be re-written as:(34) Since the inhibition of aptazyme cleavage would result in a increase of gene expression, the [R]Rel-vs- curve is an increasing hyperbola, whose descriptor can be obtained by re-writing (34) to:where:(35)(36)and(37) From these results it can be seen that the ‘Roof Gap’ (Figure 4B) for a ligand-inhibited aptazyme is always 0, since the mRNA can theoretically be completely protected when the concentration of the ligand approaches infinite. In contrast, the width of the ‘Floor Gap’ is dependent on D and the cleavage tendency ω. As before, the theoretical and realistic dynamic ranges of gene expression are graphically represented as a regulatory landscape (Figures 5E–5F). Analytically, by defining:andit can be shown that:(38)and(39) Once again, for each given D and there is an optimal ω to maximize (Figure 6B). Interestingly, though, for ligand-inhibited aptazymes a much wider range of cleavage tendencies give satisfactory values (Figures 6B∼6D). A major advance in our modeling compared to previous work ([16]) is that we provide practical guidelines for what experiments should be carried out to develop a quantitative understanding and prediction of riboswitch function. Based on our analysis, the performance of an aptazyme-based riboswitch can be quantitatively predicted when four parameters are known: (i) the gene expression level of an unengineered mRNA; (ii) the ratio of the ribozyme cleavage rate constant to the mRNA degradation rate constant (D); (iii) cleavage tendency of the aptazyme (ω); and (iv) the maximum available relative concentration of ligand (). Among these four parameters, the gene expression level of an unengineered mRNA can be trivially measured. Using equation (18), D can be obtained by measuring the gene expression level of an mRNA harboring a ribozyme sans aptamer at its 3′ UTR (or elsewhere). Once D is determined, the cleavage tendency can be predicted based on RNA folding energetics or by measuring the gene expression level of an aptazyme-harboring mRNA in the absence of ligand, according to equations (25) and (35). The only parameter that cannot be directly measured is . However, is ligand-specific, aptamer-specific, and organism-specific, but not design-specific. Therefore if the for one aptazyme is measured, can be calculated and used to predict the performance of other aptazymes which contain the same aptamer and are used in the same organism. To calculate from one need only solve equations (29) and (39), yielding:(40)for ligand-activated aptazymes and(41)for ligand-inhibited aptazymes. With such a theoretical framework we can attempt not only to promulgate engineering principles, but also to analyze previously designed aptazyme-based riboswitches. As we discussed above, Win and Smolke generated a series of theophylline-responsive hammerhead ribozymes by grafting the anti-theophylline aptamer onto loop I or loop II of the TRSV ribozyme via various communication domains [14]. When these different constructs were placed in the 3′ UTR of a reporter gene (GFP) modest ∼2-fold effects on gene regulation were observed. One rationale for the disappointing results was that introduction of aptamer domains into loop I and loop II disrupted a known, critical tertiary interaction [24]. Although the original TRSV ribozyme inserted into the 3′ UTR can inhibit the expression of GFP expression to 2% of the unengineered mRNA level, when loop II was extended the inhibition was only to ∼10%. If the steady-state GFP signal reflects the steady-state concentration of intact mRNA, the D value for the engineered aptazymes was thus likely to be ∼10. Therefore, the maximum activation and inhibition could never exceed 10-fold, as shown by Figures 6A and 6B (top panels). The constructs were inherently restricted by their very design. Beyond limitations on catalysis, we also suspect that there were limitations on either the allosteric binding sites or the available intracellular ligand concentration. Using the data from Figure S13 of Win and Smolke [14] and equations (25) and (35), the cleavage tendencies of each aptazyme were calculated (Table 1). was also calculated from each aptazyme construct using equations (40) and (41) (Table 1). Although many values fall into a narrow range, they were not consistent. Possible explanations for this inconsistency include: (i) the existence of ‘non-productive’ aptazyme conformations not considered in the model (e.g., a non-binding and non-cleaving conformation of the ligand-inhibited aptazyme); and (ii) the possibility that the basic functionality of either the aptamer or the ribozyme were significantly altered in the aptazyme designs. To further our analysis, we assume that the aptazymes showing the largest (∼12) did not operate under the caveats stated above. If so, the maximum available cellular theophylline concentration was only about 12 times the of the anti-theophylline aptamer. The anti-theophylline aptamer has a reported Kd<1µM [25]. Assuming the aptamer retains its affinity for theophylline in the cellular environment, the calculated indicates that the intracellular concentration would be on the order of 12 µM, even though the extracellular concentration of theophylline was 5mM. This discrepancy is consistent with an early finding that the intracellular concentration of theophylline in E.coli is 103-fold lower than the concentration in media [26], and with the previous performance of an engineered antiswitch in yeast [8]. The comparison between the model and the experimental data from these studies can be visualized in the regulatory landscapes shown in Figure 7A and 7B, where the calculated cleavage tendencies and the relative gene expression values are shown both in the absence of theophylline (circles) and in the presence of 5 mM theophylline (triangles). For most constructs, there was quantitative agreement between the model and experimental data with acceptable variance. It should be noted that if we had used the original, published estimates for the fold-change due to the aptazyme there would have been virtually no agreement between model and experiment. With aptazymes that had an intrinsically limited D (∼10) and a small upper limit of L* (∼12), it was ultimately to be expected that the maximum fold-change that might be available through optimization of the communication module was only ∼3.5-fold (Figures 6A and 6B, upper panels) for both ligand-activated and ligand-inhibited aptazymes. In order to actually obtain better aptazyme and riboswitch functionality both a larger D and a higher upper limit of L* must be engineered. Our model predicts that by using a 10-fold more stable mRNA the maximum fold-change can be increased to ∼7-fold (Figure 6A and 6B, lower panels; keeping the upper limit of L* constant). For this more stable mRNA when L* is also increased to 50 (by using a tighter binding aptamer∶ligand pair and/or a ligand that is better able to penetrate the cell), ∼17-fold regulation can be achieved (Figure 6A and 6B, lower panels). In summary, the dynamic range of gene expression in the current aptazyme-based riboswitch system is severely limited by the cleavage rate of the ribozyme relative to spontaneous mRNA degradation rate and the achievable intracellular ligand concentration relative to the in vivo Kd of the aptamer. Reasonable improvements of these factors should lead to a wider dynamic range of gene expression. Although throughout the above analyses we assume that the cleavage tendency can be freely tuned, this is based on the assumption that for a given sequence design the aptazyme conformations and their relative energetics can be reliably predicted. This assumption is questionable. For example, we have recently designed a series of biosensors based on the anti-thrombin aptamer, and demonstrated that biosensor properties did not align with the stabilities based on secondary structural features alone, but were fit much better by measured stabilities [27]. Similarly, attempts to computationally design hammerhead aptazymes based only on secondary structural hypotheses (the ‘slip structure’ model; [9]) yielded aptazymes that were much less activated [28]. Such discrepancies are likely to be even greater when intracellular energetics need to be predicted. For example, for the aptazymes designed by Win and Smolke [14], the cleavage tendencies calculated from experimental data (Table 1) largely disagree with the predicted cleavage tendencies calculated from the thermodynamics data (taken from Table S1 of [14]), as shown in Figure 7C. In principle, designed aptazymes should be characterized in vitro to better understand whether and how they fit either in silico data or the in vivo data. Similarly, a recent attempt at model-driven design of allosteric shRNAs also yielded only qualitative agreement with modeling based on secondary structures [29]. To better ensure coherence between model and reality, many assumptions and predictions made in our model of aptazyme-based biosensors and riboswitches need to be tested experimentally. First of all, it is critical to test to what extent the two-state structural and energetic model is acceptable. In a recent elegant study on the kinetics of a previously engineered theophylline-activated hammerhead ribozyme [9], de Silva and Walter observed four conformations relevant to activation using single-molecule fluorescent resonance energy transfer (FRET) [30]. Moreover, upon the addition of theophylline the conformational change of the aptamer domain was observed to be much faster than that of the ribozyme core. Based on these results the authors suggested a model for ligand-induced conformational change in which the aptamer domain is capable of binding ligand even in the cleavage-incompetent conformation of the aptazyme. Consequently, the ligand binding of the aptamer domain primes the conformational change of the communication domain and the ribozyme domain (induced fit). Whether this mechanism proves to be general will strongly impact how the kinetics of effector modulation are modeled, and may alter the equilibrium arguments we make herein, depending on how the different energy states are populated. In addition, parameters relevant to the in vivo environment need to be characterized in greater detail in order to understand aptazyme function. For example, translation efficiency and the half-life of cleaved mRNAs should be carefully determined since these factors, although ignored in the current model, would contribute to the background expression level when a ribozyme or aptazyme is cleaving at full speed [21]. A more fundamental and largely unknown issue is how the energetics and kinetics of RNA folding are influenced by the biochemical properties (ionic strength, viscosity, the presence of RNA chaperons and helicases) in cellular environments. While predictive models are incomplete in the absence of such information, it is nonetheless worthwhile to formulate them so that the functionality of aptazymes can be more routinely evaluated as these additional variables are acquired. The derivations of the fundamental equations (equations (2), (3), (19) and (30)) that describe how energetic parameters dictate the performance of aptazymes in vitro and in vivo are provided in the Text S1 and S2. All figures were produced with MatLab using the equations described in the text. The Supplemental Information and Figure 13 from Win and Smolke [14] were used to derive data for our analyses. The ‘designed cleavage tendency’ presented in our Figure 7C was calculated using the equations:andwhere the values of and were taken from the Supplemental Information Table 2 of reference [14].

10.1371/journal.ppat.1004403

Ontogeny of Recognition Specificity and Functionality for the Broadly Neutralizing Anti-HIV Antibody 4E10

The process of antibody ontogeny typically improves affinity, on-rate, and thermostability, narrows polyspecificity, and rigidifies the combining site to the conformer optimal for binding from the broader ensemble accessible to the precursor. However, many broadly-neutralizing anti-HIV antibodies incorporate unusual structural elements and recognition specificities or properties that often lead to autoreactivity. The ontogeny of 4E10, an autoreactive antibody with unexpected combining site flexibility, was delineated through structural and biophysical comparisons of the mature antibody with multiple potential precursors. 4E10 gained affinity primarily by off-rate enhancement through a small number of mutations to a highly conserved recognition surface. Controverting the conventional paradigm, the combining site gained flexibility and autoreactivity during ontogeny, while losing thermostability, though polyspecificity was unaffected. Details of the recognition mechanism, including inferred global effects due to 4E10 binding, suggest that neutralization by 4E10 may involve mechanisms beyond simply binding, also requiring the ability of the antibody to induce conformational changes distant from its binding site. 4E10 is, therefore, unlikely to be re-elicited by conventional vaccination strategies.

4E10 is an antibody that neutralizes a broad variety of HIV strains. However, 4E10 is uncommon in infected patients and has not been successfully elicited by any vaccine approach attempted. Hurdles to re-eliciting 4E10 include the accumulation of many mutations during development, demonstrated reactivity against host proteins and significant structural flexibility. Lacking a confirmed sequence for precursors of 4E10, we studied the recognition and biophysical properties of an ensemble of eight of the likeliest candidates. Surprisingly, 4E10 gained host reactivity and structural flexibility, but lost stability during development when compared to candidate precursors. However, recognition of HIV was remarkably conserved, despite a considerable improvement in binding. Since these results run counter to those expected from conventional vaccination protocols, 4E10 is unlikely to serve as the basis of a useful HIV vaccine.

An effective HIV vaccine will likely need to elicit broadly-neutralizing antibodies (bnAbs) that target the viral envelope protein (Env) as part of a protective immune response [1]–[6]. Env-derived and reverse-engineered immunogen-based vaccines, however, have consistently failed to elicit bnAbs. Possible explanations include that: (1) immunogens may be unable to bind germline-encoded precursors (GEPs) of bnAbs with sufficient affinity to initiate B cell activation and affinity maturation, which has a ∼micromolar threshold [7]–[9]; (2) rearranged VH and VL genes compatible with the development of bnAbs may not be common in human or animal model vaccinee GEP repertoires; (3) some bnAbs are autoreactive, which hinders their elicitation through self-tolerance mechanisms; (4) the unusual characteristics inherent to bnAbs, such as long complementarity determining regions (CDRs), functionally-required polyspecificity, and a high degree of somatic mutation (typically observed in Abs elicited in response to chronic infections, including bnAbs), may not be easily achieved through conventional vaccination strategies; (5) imperfect immunogens may elicit off-target (non-neutralizing or non-Env) or humoral responses with limited breadth; and, finally, (6) neutralization mechanisms may involve complexities beyond simply binding a particular epitope on Env (e.g., inducing specific conformational changes), which may be difficult to recapitulate, since selective expansion of particular B cell clones is based solely on BCR binding properties, not higher-order functionalities [10], [11]. The bnAb 4E10, the focus of our studies, has a conserved, linear epitope (core epitope: 671NWFD/NIT676) immediately adjacent to the viral membrane in the Env gp41 subunit membrane proximal external region (MPER) [12], [13]. While 4E10 displays admirable breadth [14], has been the target of a successful design effort to reverse-engineer tight-binding immunogens [15], has recognizable GEPs present at finite frequencies in human naïve repertoires [16], and arguably does not require a high degree of polyspecificity to neutralize HIV [17], [18], its viability as a vaccine target is hampered by limited potency, demonstrated autoreactivity and exceptional combining site flexibility [17], [19], [20]. The neutralization mechanism of 4E10 also has not been clearly defined and may involve higher order effects [17], [21], [22]. The ontogeny of 4E10, therefore, must be elucidated in order to understand how these properties were acquired and to what degree they impose constraints that might hinder re-elicitation by vaccination. Mutations acquired during Ab maturation occur preferentially in the CDRs, which make up the six loops (CDRs 1, 2, and 3 on the heavy (HCDRs) and light (LCDRs) chains) comprising the combining site [23], [24]. CDRs are responsible for the majority of direct contacts to an antigen, as opposed to the intervening framework regions (FWRs), which form the immunoglobulin β-sheet structure stabilizing the combining site, helping define CDR loop conformations. While CDR mutations are typically thought to more directly affect antigen binding and neutralization, bnAbs consistently depend on FWR substitutions to a surprising degree, though 4E10 is an exception to this exception [4]. bnAbs are notorious for their high degree of somatic hypermutation, the product of a long process of affinity maturation against a rapidly mutating virus during a persistent, chronic infection [25]. While typical affinity-matured Abs have acquired 15 to 20 VH mutations, bnAbs accumulate up to 100 VH mutations [4]. These mutations are crucial because reversion to germline sequences drastically reduces epitope affinity and neutralization potency and breadth. In many cases, bnAb GEPs are unable to bind Env, though the actual eliciting isolate may not be known [26]–[31]. In addition, bnAbs can contain extraordinarily long HCDR3s, up to 33 residues long versus an average of 13 for non-HIV bnAbs [32], [33]. Using phylogenetically-inferred GEP sequences [16], [34] (Fig. 1), 4E10 has acquired between 33 and 35 mutations during maturation, 20 to 22 in VH, depending on gene segment selection, and 13 in VL, and has an HCDR3 22 residues long, values at the less exceptional end of the bnAb spectrum and not unheard of for conventional Abs. Affinity-matured Abs display univalent equilibrium binding constants (KD) for their cognate antigens, typically ranging from 10−6 to 10−10 M, that are orders-of-magnitude stronger than their GEPs [35]. Multiple approaches, including computational analyses and biophysical comparisons of affinity-matured Abs and their associated GEPs, have generated a consensus model for the molecular mechanism of affinity maturation [36]–[47], perhaps better understood as binding optimization, that traces its roots back to Pauling [48]. In the consensus model, the antigen specificities of the naïve, germline-encoded repertoire, while diverse and extensive, are further extended by encoding a high degree of polyspecificity into GEPs. This is accomplished partly through increased combining site plasticity in GEPs, more formally stated as the ability of GEP CDRs to dynamically sample a broader ensemble of structural conformers. In response to immunogen stimulation, Ab binding properties are iteratively optimized through cycles of somatic hypermutation and selection, resulting in improved binding affinities, kinetics and thermodynamics. While mutations have been observed to improve or add direct contacts to antigen, typically improving enthalpies of interaction and off-rates (kd), a majority of measurably favorable mutations do not directly contact antigen. These mutations indirectly optimize binding by: (1) increasing shape complementarity between paratopes and epitopes through more global effects on structure; (2) increasing antibody stability, typically measured as solution thermostabilities (Tm), thus compensating for deleterious effects of other mutations that improve affinity; and (3) structurally rigidifying the combining site conformer optimal for binding antigen from the accessible ensemble. Rigidifying the combining site can affect measured interaction parameters in different ways depending on the mechanism of binding. The two mechanistic extremes are known as “conformer-selection” and “instructive-encounter”, or “induced-fit”, binding. In conformer-selection mode, binding does not occur until the compatible conformer is adopted. Rigidification of the binding site through mutation then typically improves entropies of interaction and on-rates (ka). In instructive-encounter mode, initial binding occurs to sub-optimal conformers which affects the rate of interchange with the optimal conformer. Rigidification of the binding site through mutation then typically improves affinity through changes distributed over kd and ka. However, the consensus of studies of binding proteins and enzymes suggests that conformer selection is the preeminent recognition mechanism [49]. Surprisingly, comparisons of the bound and unbound structures of 4E10 revealed that this affinity-matured bnAb incorporates considerable HCDR conformational flexibility [17], in excess of what has been observed in most other antibodies, mature or GEP, suggesting that the ontogeny of 4E10 may be an exception to the consensus model and may pose unique challenges as a vaccine target. In order to fully understand the ontogeny of this unique bnAb and consequences for vaccine development, we characterized the unbound and complex crystal structures, and the functional and binding properties, of 4E10 and an ensemble of its most likely GEPs. GEPs showed detectable, but extremely weak, binding to soluble Env gp140s and extremely limited neutralization potency, though some reverse engineered epitope-scaffolds (ESs) showed robust GEP affinities, well above the B cell activation threshold. 4E10 and GEP paratopes displayed a remarkable degree of structural conservation in the antigen-bound state, with little improvement in overall shape complementarity. Multi-log improvements in affinity for ESs were the result of improved off-rates or combined improvements in on- and off-rates, with a small number of enhanced contacts to antigen observed in the crystal structures. However, minimal mutations of GEP sequences to include these enhanced direct contacts only marginally increases affinity. FWR mutations had little discernable effect on global or local structure. Controverting the consensus model of ontogeny, 4E10 thermostability was appreciably worse than its GEPs; while 4E10 and GEPs displayed similarly constrained VH/VL interdomain movements upon binding, 4E10 maturation involved negligible combining site rigidification, with both 4E10 and GEP HCDRs sampling extensive conformer ensembles. The narrowing of polyspecificity assumed to concur with maturation was not observed with 4E10, as both 4E10 and its GEPs showed similar patterns of limited polyspecificity to a phage-displayed human peptidome (Phage Immunoprecipitation Sequencing; PhIP-Seq) [50]. While 4E10 is demonstrably autoreactive, GEPs exhibited a distinct profile of autoantigen recognition by PhIP-Seq. When combined, these results inform efforts to re-elicit 4E10 by vaccination and its mechanism of neutralization. Lacking access to the original donor, identification of a single GEP sequence with high confidence for many bnAbs, including 4E10, is complicated by extensive editing and TdT N-nucleotide insertion during rearrangement [26], leading to our decision to study an ensemble of the 12 likeliest candidates (Fig. 1). Due to the prediction that the sequence differences introduced by alternate heavy chain J gene segments may not affect any discernable GEP property, the initial ensemble was limited to eight GEPs (IGHJ4*02 paired with all six D segments plus IGHJ1*01 paired with IGHD1-1*01 and IGHD6-19*01), with the intention of generating additional GEPs if the IGHJ4*02/IGHJ1*01 substitution exhibited any differences in structure or binding properties on the IGHD1-1*01 or IGHD6-19*01 backgrounds. These eight GEPs also recapitulated some of the variability seen in potential 4E10 GEPs identified by deep sequencing of uninfected individuals (Fig. 1) [16], providing an additional justification for studying the ensemble. However, GEPs in our ensemble differed from one another by no more than four mutations, though the mutations were quite non-conservative. GEPs were engineered as cleavable, single-chain, variable domain cassette (Fv; VH+VL) constructs to ease expression, analysis, and crystallization and to prevent monobody-diabody interchange, following protocols developed for 4E10 [18]. The prior study confirmed that these Fv constructs retained the structural and binding properties of Fab fragments of intact 4E10. All eight GEP Fvs expressed at high levels as bacterial inclusion bodies and, in all but one case (GEP 5), were refolded in vitro with what in our experience were exceptionally high efficiencies of 20 to 40%. GEP 5 was not included in subsequent experiments because its extremely poor in vitro refolding efficiency suggested that this was not a viable pairing. GEP constructs were stable and monodisperse in solution, running exclusively as monomers by size exclusion chromatography (SEC). Reduced/non-reduced PAGE analysis of GEPs confirmed purity and proper disulfide bond formation. GEP Tm values (Fig. 2A), determined by circular dichroism (CD) spectroscopy as previously described [18], narrowly ranged from 64.2°C to 67.0°C, showing that GEP Fvs were well folded, but had even higher Tm values than 4E10 Fv (52.8°C). The neutralization potency of GEPs was tested against clade A (Q461.d1, Q461.e2 [51]) and B (SF162 [52]) HIV-1 isolates using standard TZM-bl assays (Fig. 2B) [53]. Overall, GEP Fv potencies were markedly reduced in comparison with 4E10 Fv. GEPs displayed only very weak and likely insignificant neutralization potencies, though with a trend of greater effect against the clade A isolates, particularly the neutralization-sensitive strain Q461.d1, and with GEP 1 and GEP 7 showing marginally better potencies across tested isolates. In order to characterize the change in binding properties during 4E10 ontogeny, the binding of 4E10 and GEP Fvs to three (SF162, SF162K160N and Q461.e2) soluble Env gp140 trimers (gp1403) [54] and four engineered 4E10 ES immunogens (T72, T93, T117 and T344) [15], [55] was evaluated by surface plasmon resonance (SPR) interaction analysis (Figs. 2C, 2D, Figs. 3–7, Table 1). Isolates and ESs were selected to span a range of binding properties, where previous studies had shown 4E10 bound a free peptide spanning its linear epitope with a KD of 12 nM, SF162 gp1403 with a KD of 98 nM, and ESs with KDs of either ≤10 picomolar (T117) or ∼100 picomolar (T72, T93 and T344) [15],[18],[55]. All seven GEPs showed unquantifiably weak, but detectable binding to chip-coupled clade A (Q461.e2) and clade B (SF162, SF162K160N) gp1403 and T72, T93, and T344 in qualitative SPR analyses, with KDs all estimated to be well above the ∼micromolar B cell activation threshold. Quantitative SPR analyses of GEPs binding to T117 showed KDs ranging from the low nanomolar to low micromolar range. 4E10 interactions with ESs ranged from 100- to 10,000-fold stronger than GEPs, which was qualitatively consistent with the observed difference in 4E10 versus GEP interactions with gp1403. Since the GEP/gp1403 interactions were too weak to quantitate, peptide binding studies were not performed on the expectation that they would also be too weak to measure accurately. Five GEPs (1, 2, 3, 7 and 8) bind T117 with nearly identical behavior, including GEP 1 and GEP 7, which differ only by alternate J segment utilization, showing that the two incorporated mutations did not affect binding, so no further IGHJ1*01 GEPs were produced for study. GEP 4 and GEP 6, which both incorporate differences from G96H in 4E10 (A or V), showed approximately 10-fold (GEP 4) or 100-fold (GEP 6) reductions in affinity relative to the cluster of other GEPs. Kinetically, the five clustering GEPs (1, 2, 3, 7 and 8) showed affinity reductions for T117 relative to 4E10 overwhelmingly through faster off-rates (kds). GEPs 4 and 6, in addition to comparable increases in kd, also showed progressive reductions in on-rates (kas). In order to shed light on potential structural differences accounting for reduced GEP binding affinities, crystal structures of GEP 1, 2 and 7 in complex with T117 were determined at resolution values of 2.9 Å, 1.8 Å, and 3.1 Å respectively, rebuilt and refined with good statistics (Table 2), and compared to two reference 4E10/antigen complex structures: 4E10 bound to an epitope peptide (2FX7.pdb [12]) or a related ES, T88 (3LH2.pdb [15]). Superpositions showed that almost all direct contacts to the core epitope (NWFDIT) and epitope conformation are conserved between 4E10 and GEP complexes to a remarkable degree (Fig. 8, Tables 3 and 4). Only six of the 35 predicted somatic mutations affect sequence positions making direct contacts to the core epitope: Y/K32L, S/Q93L, T/S31H, I/V51H, F/L54H and T/I56H. However, residues at two of these positions (93L, 51H) contacted the epitope solely through main-chain interactions, which were not affected by the mutated side-chains. Replacements at three other positions (31H, 54H, 56H) nearly perfectly recapitulated contacts, also conserving the interface. The 4E10 and GEP paratope/epitope interface is largely hydrophobic, thus binding is mainly mediated by Van der Waals contacts and desolvation entropy, which is conserved through equivalent positioning of nonpolar groups. This was reflected in the close concordance between the surface areas (SA) buried in the complexes by core epitope and the corresponding shape complementarity (Sc) [56] between Ab and core epitope. Only two somatic mutations were predicted to contribute to a stronger binding interaction. The Y/K32L mutation replaces a hydrogen bond involving the tyrosine hydroxyl with a salt bridge. The P/L95H mutation, involving non-contacting residues, restructured LCDR3 to reposition the side-chain of the conserved serine at 94L from a non-contacting position in GEPs to one contributing a hydrogen bond in 4E10 complex structures. However, the Y/K32L-P/L95H double mutation made on a GEP 1 background (GEP 1m) did not affect Tm, only marginally increased affinity by about three-fold, and did not improve neutralization potency (Figs. 2A, 2D, 6 and Table 1), arguing that affinity and neutralization are largely influenced by somatic mutations through indirect effects beyond the ability to adopt the optimal binding conformer. In order to determine whether the 4E10 combining site rigidified during affinity maturation, undergoing binding site preconfiguration, the crystal structures of GEP 7 and GEP 1 were determined in the unbound state at resolution values of 1.9 Å and 1.7 Å respectively, rebuilt and refined with good statistics (Table 2), and compared to the unbound structure of 4E10 (4LLV.pdb [17]). Superpositions of 4E10 and GEP bound and unbound structures showed that interdomain movements, while present, were limited compared to CDR rearrangements (Fig. 9), and that 4E10 retained at least as much CDR flexibility as was present in GEPs, particularly in HCDR2 and HCDR3, while LCDRs were relatively constrained (Figs. 10A and 10B). Calculated root mean square deviations (RMSDs; Fig. 10C) of VH and VL superpositions, with and without CDRs, confirmed that while the FWR regions of GEPs and 4E10 were nearly identical, the bulk of rearrangements observed during binding occurred in HCDRs, with 4E10 movements as large, or larger, than observed in GEPs. LCDR movements, while much smaller overall, contributed less to rearrangements in 4E10 than GEPs, suggesting some minimal degree of rigidification during maturation. Comparison of the surfaces directly contacting epitopes (Movies S1 and S2) showed that GEPs retained higher degrees of structural conservation than 4E10 between the bound and unbound states. These analyses need to be interpreted cognizant of the constraints imposed by crystallization, which involved varying non-physiological solution conditions and variable crystal contacts between structures. However, these caveats were minimized by comparing multiple structures with multiple copies per asymmetric unit. The goal of reverse engineering an Ab is to scaffold the desired epitope to re-elicit Abs that solely recognize the epitope [6], [57]. However, the 4E10 linear epitope, as currently defined, is smaller than typical Ab/antigen interfaces, which poses the design challenge of isolating humoral responses to the epitope and not contiguous scaffold surfaces. Previous crystal structures of 4E10/ES complexes showed that many ESs achieved this goal well, including T93 [15], [55]. However, GEP/T117 complex structures reported here showed extensive GEP/scaffold contacts (SA for GEP2 contacts to scaffold minus epitope = 308 Å2) (Fig. 11). A dominant feature of these interactions was the binding of the side-chain of the GEP-specific residue F54H in a deep hydrophobic cleft of the T117 scaffold protein, a putative phosphotransferase from S. typhimurium. These additional contacts raise the concern that Abs elicited by T117 immunizations may have off-target (non-Env) specificities. Nevertheless, the T117 scaffold is highly complementary to 4E10 (Sc for GEP2/T117 = 0.69), which may help explain the increased affinity of T117 for 4E10 and GEPs, and may be ideal for preferentially targeting GEPs through the F54H interaction (F54H is present in the heavy chain gene used by all GEPs). The additional T117 contacts appear to have had the effect of increasing the affinity for T117 over other ESs by two orders of magnitude for 4E10 and by one to three orders of magnitude for the GEPs. However, structural superpositions show that these extra Ab/scaffold interactions do not affect Ab/epitope interactions, or the structure of the epitope in the complexes, which are highly conserved (Fig. 8B). A validated phage-displayed library consisting of 413,611 overlapping 36-mer peptides spanning the entire human proteome combined with phage immunoprecipitation sequencing (PhIP-Seq) was used to assess the polyspecificity and autoantigen recognition profiles of GEPs in comparison to 4E10 [17], [50] (Fig. 12A, Tables 5, 6, and 7). GEP 2 and GEP 4 were selected to represent both the clustering (GEP 2) and 96H mutant (GEP 4) GEPs, and an affinity-matured, murine anti-canine CD28 Ab, 1C6 [58], was included for comparison (Fig. 12A, Table 8). The top hit in the 4E10 Fv PhIP-Seq analysis reported here, a peptide derived from the type 2 inositol 1,4,5-trisphosphate receptor (IP3R), matches the top hit from the previous PhIP-Seq analysis of IgG 4E10 [17]. However, overall scores were considerably reduced in the PhIP-Seq analysis of 4E10 as an Fv construct versus intact IgG (replicate average −Log10 P-values for the top ten scoring peptides of 35.3 to 255 for IgG versus 4.35 to 12.3 for Fv), likely due to decreased accessibility of coupled Fv relative to IgG, an increased chance of inactivating Fv versus IgG during chemical coupling, and the inherent increase in local avidity of bivalent IgG versus univalent Fv on potentially sparsely-coupled beads. Given these caveats, overall scoring behavior was very similar across the Fvs tested, with 4E10 and 1C6, the affinity-matured Abs, showing the highest average scores. GEP 4 and 1C6 showed the largest number of high-scoring hits, with 61 and 194 peptides with replicate-averaged −Log10 P-values of ≥4.0 respectively; 4E10 and GEP2 had 12 and 20 peptides, respectively, scoring ≥4.0. Qualitatively, the results were not dramatically different, but with GEP 4 and 1C6 showing nominally greater spreads of top-scoring peptides rising above the bulk responses. None of the top-scoring three-dozen 4E10 peptides appeared in the top three-dozen hits from either GEP; however two of the top-scoring dozen GEP peptides (derived from zinc finger Ran-binding domain-containing protein 3 or hyaluronidase-3 isoform 1 precursor) were in common between GEP 2 and GEP 4, scoring 1 and 12 (GEP 2) and 4 and 2 (GEP 4), respectively. However, no common peptide motifs could be identified within or between Fv results. None of the IP3R peptides scored in the top 65,000 GEP hits. While 1C6 showed the highest scoring spread of top hits (replicate average −Log10 P-values for the top ten scoring peptides of 12.14 to 30.9), the top ten scoring peptides displayed a considerably higher average hydrophobic character (Φ), with average Φ values of: 4E10 = 0.37; GEP 2 = 0.42; GEP 4 = 0.45; 1C6 = 0.71 (higher values are more hydrophobic). Using relative hydrophobicity of the top-scoring PhIP-Seq peptides as a surrogate measure of the overall hydrophobicity of the combining site was consistent with the structures of the Abs (Fig. 12B), where the 1C6 combining site is structured as a large, broad, very hydrophobic concavity and 4E10 and its GEPs sporting smaller, convex hydrophobic surfaces. In order to test the hypothesis that 4E10 may induce global conformational changes in gp1403 as part of a higher-order neutralization mechanism [17], [21], consistent with downstream effects such as gp120 release, binding of Abs to epitopes distant from the MPER (the V3 loop, by 447-52D [59], and the CD4 binding site, by b12 [60]) was assayed by SPR in the presence or absence of saturating 4E10 (Fig. 13). The qualitative results show that 4E10 pre-binding does not affect 447-52D binding to the flexible, extended V3 loop, but does alter the association kinetics of b12 at the CD4 binding site, suggesting that 4E10 binding induced global conformational changes in gp1403 registering at distant sites. Since b12 dissociation kinetics were unaffected, 4E10 did not preclude achieving a similar b12 bound-state conformation. Reverse engineered ESs are ideal reagents for studying the binding properties of GEPs. Where peptides and Env do not display sufficient affinities, the best ESs display strong affinities across the GEP ensemble, allowing for the biophysical characterization of 4E10 ontogeny by using ESs to compare 4E10 and GEP binding kinetics and bound-state crystal structures. Consistent with current theory and previous results comparing GEPs with matured bnAbs or other Abs, 4E10 displays orders-of-magnitude better affinities than candidate GEPs for both Env gp1403 proteins and engineered ESs. This gain in affinity is potentially sufficient to account for the concurrent gain in neutralization potency over GEPs. Previous results showing minimal contributions from FWR mutations [61] and comparisons of complex structures of 4E10 with GEP/T117 complex structures, which showed near-complete conservation of the recognition interface (Fig. 8), argue that affinity maturation was likely due, at least in part, to the small number of somatic mutations in the CDR regions, e.g. 32L and 95H, that add or improve direct contacts to epitope. This limited number of mutations is presumably readily achievable during conventional vaccination. Based on PhIP-Seq peptidome binding results (Fig. 12A), 4E10 and GEPs have distinct autoantigen profiles, suggesting that 4E10 autoreactivity was acquired during ontogeny and not inherent in GEPs, consistent with recognizable GEPs populating naïve repertoires of mature B cells at relatively high frequencies [16]. Among ESs, T117 in particular interacts with GEPs sufficiently strongly to drive B cell activation and maturation (Fig. 2D) and selectively interacts with GEP-specific structural features, e.g. F54H (Fig. 11). These aspects of 4E10 combine to seemingly argue that 4E10 ontogeny follows a relatively short path and that directed re-elicitation of 4E10, and perhaps other bnAbs, may be an achievable goal. In many aspects, however, 4E10 ontogeny appears more convoluted. Contrasting with the consensus model of Ab ontogeny and results from other systems [38], 4E10 is considerably less thermostable than GEPs, which was difficult to account for only from static views of structures. This lowers the headroom available to tolerate mutations that increase affinity or potency that otherwise might degrade stability, potentially adding a significant hurdle to re-elicitation of 4E10, but clearly does not limit the ability of 4E10 to bind to or neutralize HIV. The conformer ensemble sampled by GEP combining sites echoes the startling plasticity of 4E10, with structural comparisons of bound versus unbound states (Figs. 9 and 10) showing perhaps a small degree of 4E10 LCDR rigidification, but an increase in the conformer ensemble sampled by the HCDRs. This directly contradicts the consensus model, and specific examples [40], of Ab maturation, showing alternatively that preconfiguration of 4E10 does not occur during, or contribute to, affinity maturation. HCDR flexibility is highlighted by F29H, which is able to flip out into solvent in the unbound 4E10 and GEP 1 structures, allowing HCDR1 to dynamically sample multiple conformers. A high number of conserved glycine residues may contribute to 4E10 and GEP HCDR mobility. The reduced affinities due to decreased kas (Fig. 2D), relative to the clustering GEPs, of GEP 4 and 6 with non-glycine residues at position 96H, predicted to restrict HCDR3 conformer sampling, suggests that HCDR3 mobility was needed to achieve the bound-state conformation with optimal kinetics by destabilizing non-optimal conformers. Retention of significant combining site plasticity also strongly argues for a functional role other than polyspecificity, which is unremarkable in comparison to other Abs, based on PhIP-Seq peptidome binding results (Fig. 12A). Minimal focusing of apparent polyspecificity also contradicts the consensus model, and specific examples [62], of maturation. Also unexpectedly, the ∼100-fold increase in affinity of 4E10 over GEPs to T117 is through a decreased kd (Fig. 2D), suggesting that improvements in binding affinity arose from a more favorable bound state and not a decrease in entropic barriers imposed by more plastic GEPs. Thermodynamic studies of 4E10 and GEPs were not possible because no single antigen, needed for valid comparisons, bound to both 4E10 and GEPs with parameters accessible to measurement. However, while crystal structures showed a highly-conserved bound state in both 4E10 and GEPs, the small number of improved contacts could not account fully for the observed improvement in affinity. Complicating this analysis, and its use as a vaccine immunogen, are the extensive contacts between GEPs, and presumably also 4E10, and the scaffold of T117 outside of the stabilized MPER epitope (Fig. 11). The interaction with T117 also highlights another possibility. The 4E10 combining site extends beyond the minimal, linear epitope in T117, making contacts to the scaffold. Unless the MPER epitope forms an isolated structure extended away from the rest of Env, 4E10 may make contacts to elements of Env outside of the linear 4E10 epitope, as it does with T117. HCDR conformer dynamics may therefore be understood as enabling such interactions, which may foster conformational changes in Env leading to neutralization by any of several mechanisms. Prebinding of 4E10 to gp1403 affected b12 binding at the CD4 site distant from the MPER (Fig. 13), indirectly demonstrating a global conformational change consistent with this supposition. Additional interactions between GEPs and the scaffold moiety of T117 would also be predicted to drive rigidification of an even greater portion of the combining site during conventional immunization, increasing the distinction between 4E10 and a matured Ab derived from a 4E10 GEP stimulated with T117. T117 was the product of a cutting-edge design effort to generate a structurally-stabilized MPER epitope with optimized binding properties for use as a vaccine immunogen to drive 4E10 ontogeny based on the current paradigm of affinity maturation. However, 4E10 appears to take the affinity maturation pathway less traveled, one that contrasts, in many fundamental ways, the current paradigm. This raises the concern that conventional immunization protocols, based on the current paradigm of structural stabilization of optimal binding conformers, will not successfully re-elicit 4E10 or other unconventional Abs. Understanding unconventional maturation pathways then becomes paramount for the future of molecular vaccinology, allowing efforts to focus on re-eliciting Abs identified as the products of conventional ontogeny in the near term while developing unconventional vaccination strategies to target exceptional Abs in the long term. The 4E10 heavy and light chain nucleotide sequences were analyzed using JoinSolver [63], IMGT/V-QUEST [64], [65], Ab-Origin [66], SoDA [67] and iHMMune [68] to compositely identify segments with the fewest nucleotide mismatches, generating a combinatorial ensemble of 12 GEPs (Fig. 1). Each derived GEP candidate consisted of a single VL domain sequence assigned with high confidence (IGKV3-20*01 plus IGKJ1*01), one VH gene segment also confidently assigned (IGHV1-69*06), plus one of six likely D segments (IGHD1-1*01, IGHD6-19*01, IGHD7-27*01, IGHD6-25*01, IGHD1-26*01, IGHD4-17*01) and either of two likely J segments (IGHJ4*02, IGHJ1*01). The CDR3 junctions and presumed N-nucleotide insertions in mature 4E10 were retained in the candidate GEPs. Sequences are numbered following Kabat [24]. 4E10, eight of the 12 GEPs, and 1C6 Fvs were engineered, expressed as inclusion bodies in E. coli BL21(DE3) RIL cells (Invitrogen), refolded, purified, and validated by SEC and PAGE as previously described [18]. Circular dichroism (CD) spectra of 4E10 Fv and GEPs were measured on a J-815 spectrometer (Jasco) at a concentration of 10 µM in 10 mM Na2HPO4/KH2PO4 (pH = 7.4). Temperature melts were performed at 210 nM with a temperature ramp from 25°C to 95°C at a slope of 2°C/minute, data pitch of 2°C, and delay time of 10 s (Fig. 2). Tms were determined by nonlinear least-squares analysis using a linear extrapolation model with Spectra Analysis software (Jasco). Fabs of 447-52D and 4E10 were prepared by digestion of IgG with papain (Pierce), affinity chromatography with protein A (Pierce) and preparative SEC (Superdex 200 16/10 column; GE Healthcare); IgG 4E10 was purchased from Polymun Scientific, and IgG 447-52D and Fab b12 were kindly provided by Pamela Bjorkman (Caltech). Neutralization assays (Fig. 2B) were performed using single-round entry-competent viruses and TZM-bl cells as previously described [53]. Percent neutralization at concentrations of 0.96 µM (4E10 and GEP Fvs) or 0.17 µM (4E10 IgG) was calculated as previously described [18], [53]. All SPR experiments were performed at 25°C on a Biacore T100 instrument with the T200 sensitivity enhancement (GE Healthcare). For analyses of the binding of 4E10 and GEP Fv analytes to chip-captured ES ligands (Fig. 2D, Table 1), ESs at ∼1 µg/mL were captured on a SA sensor chip (GE Healthcare) following either carboxy (T117) or amine (T72, T93, T344) biotinylation, following the manufacturer's recommended protocol (EZ-Link, Thermo Scientific), in a running buffer of HBS-EP+ (10 mM HEPES, pH 7.4, 150 mM NaCl, 3 mM EDTA, 0.05% P-20; GE Healthcare) plus 0.1 mg/mL bovine serum albumin. A reference flow cell was left blank. Duplicate 4E10 and GEP Fv analyte injections were randomized and run at a flow rate of 50 µL/minute. Regeneration, if needed, was achieved by injection of 10 mM glycine at 50 µL/minute followed by buffer stabilization. Sensorgrams obtained from SPR measurements were double-reference subtracted [69] with BIAevaluation 2.0.3 software (GE Healthcare) employing previously described methodology [18]; data were fit with either 1∶1 or steady-state binding models. For analyses of the binding of 4E10 and GEP Fv analytes to gp1403 (Fig. 2C), gp1403 at 30 µg/mL in 10 mM sodium acetate (pH = 5.0) were direct amine coupled at a density of ∼2200 RUs on a CM5 sensor chip (GE Healthcare); a reference surface was prepared by activating and deactivating a flow cell without the addition of protein; gp1403 were prepared as previously described [18]. Duplicate 5 minute, 300 nM injections of 4E10 and GEP Fv were made in HBS-EP+ at 50 µL/minute followed by 5 minutes of dissociation, regeneration with 10 mM glycine (pH = 1.5) for 5 seconds and then 6 minutes of stabilization. Assays of Fab b12 or Fab 447-52D binding to SF162 gp1403 in the presence or absence of 4E10 Fv (Fig. 13) were conducted in HBS-EP+ buffer. Using standard amine coupling chemistry, 1273 RUs of SF162 gp1403 were coupled to a CM5 sensor chip. A reference surface was prepared by activating and deactivating a flow cell without the addition of protein. Two samples were injected at a flow rate of 50 µL/minute using the dual injection command in the T100 control software (v2.0.3, GE Healthcare) with injection 1 at 5 minutes, injection 2 at 2 minutes and a final dissociation time of 5 minutes. Fab-alone curves were generated by injecting HBS-EP+ followed by an injection of 10 nM Fab b12 or 447-52D and double referenced by subtracting a dual injection of HBS-EP+ followed by HBS-EP+. Fab with 4E10 Fv curves were generated by injecting 3 µM 4E10 Fv (∼90 RUs bound) followed by an injection of 10 nM Fab b12 or 447-52D in the presence of 3 µM 4E10 Fv and double referenced by subtracting a dual injection of 3 µM 4E10 Fv followed by 3 µM 4E10 Fv. Optimal regeneration was achieved by injection of 10 mM glycine (pH = 1.5) at a flow rate of 50 µL/minute for 5 seconds followed by a 6 minute buffer stabilization phase. Sensorgrams were corrected by the double-subtraction method [69] in Scrubber 2.0b software (BioLogic Software). Fig. 13 shows the second sample of each set of injections with baselines zeroed just prior to the second injection. Crystals of GEPs and 4E10 Fv were grown by the hanging-drop vapor diffusion method at 25°C with the following well solutions: Crystals were cryopreserved in mother liquor containing 15% v/v glycerol (GEP 1, GEP 1/T117, GEP 7/T117), mother liquor containing 10% (1C6) or 20% (GEP 7) v/v glycerol, or 20% w/w sucrose (GEP 2/T117). Diffraction data for GEP 1, 4E10, and 1C6 were collected at the Advanced Light Source beamline 5.0.2 (Lawrence Berkeley National Laboratory, Berkeley, CA) and reduced using HKL-2000 [70] or d*TREK (Rigaku) [71]. Diffraction data for GEP 7, GEP 1/T117, GEP 7/T117 and GEP 2/T117 were collected in house with CuKα radiation on a R-AXIS IV++ image plate detector with HR optics (Rigaku) at −170°C. Initial phase information for all data sets was determined by molecular replacement, as implemented in the CCP4i program suite [72]–[74], using 3LF6.pdb (T117), 3LH2.pdb (GEPs) and 1JP5.pbd (1C6) as initial search models. Phases were improved by subsequent rounds of model building and refinement using COOT [75] and REFMAC [76]. Structure validation was carried out with PROCHECK [77], the MolProbity server [78], and the RCSB ADIT validation server. Data collection and structure refinement statistics are shown in Table 2. Crystals of GEP 2 alone could not be grown, despite considerable effort. 4E10, GEP and 1C6 Fvs were coupled to beads and analyzed in duplicate. For each Fv analyzed, 3 mg of magnetic beads (Invitrogen M-270 Epoxy Dynabeads) were resuspended in 60 µL 0.1 M NaPO4 (pH = 7.4). Beads were rocked at ambient temperature for 24 hrs with 60 µg of each Fv in 1 M (NH4)2SO4 and then washed with 10 mM glycine in PBS to cap unreacted epoxy groups. Activity of Fv coupled beads was confirmed by epitope-scaffold binding prior to PhIP-Seq analyses. PhIP-Seq analyses were performed as previously described [17], [50]. Results were plotted as replicate #1 versus replicate #2 −Log10 P-values; highly discordant, and therefore spurious, hits falling near the axes were excluded from analysis. 241 peptides were also discarded because they displayed nonspecific binding to all Abs tested [17]. Peptide hydrophobicity was determined with the Sigma-Aldrich PEPscreen Library Design Tool and overall charge at neutral pH was determined with the Innovagen Peptide Property Calculator. Coordinates and structure factor amplitudes have been deposited in the Protein Data Bank [79] under accession codes: 4M8Q.pdb (ligand-bound GEP 1), 4LRN.pdb (unbound GEP 1), 4M62.pdb (ligand-bound GEP 2), 4ODX.pdb (ligand-bound GEP 7), 4OB5.pdb (unbound GEP 7), and 4LCI.pdb (1C6).

10.1371/journal.ppat.1002379

A Novel Role for the NLRC4 Inflammasome in Mucosal Defenses against the Fungal Pathogen Candida albicans

Candida sp. are opportunistic fungal pathogens that colonize the skin and oral cavity and, when overgrown under permissive conditions, cause inflammation and disease. Previously, we identified a central role for the NLRP3 inflammasome in regulating IL-1β production and resistance to dissemination from oral infection with Candida albicans. Here we show that mucosal expression of NLRP3 and NLRC4 is induced by Candida infection, and up-regulation of these molecules is impaired in NLRP3 and NLRC4 deficient mice. Additionally, we reveal a role for the NLRC4 inflammasome in anti-fungal defenses. NLRC4 is important for control of mucosal Candida infection and impacts inflammatory cell recruitment to infected tissues, as well as protects against systemic dissemination of infection. Deficiency in either NLRC4 or NLRP3 results in severely attenuated pro-inflammatory and antimicrobial peptide responses in the oral cavity. Using bone marrow chimeric mouse models, we show that, in contrast to NLRP3 which limits the severity of infection when present in either the hematopoietic or stromal compartments, NLRC4 plays an important role in limiting mucosal candidiasis when functioning at the level of the mucosal stroma. Collectively, these studies reveal the tissue specific roles of the NLRP3 and NLRC4 inflammasome in innate immune responses against mucosal Candida infection.

In this manuscript we describe a new role for a group of molecules termed the “inflammasome” that process key immune response proteins including interleukin-1-β. In previous work, we and others have shown that the NLRP3 inflammasome is important in protecting from severe fungal infections. We now show that, in addition to the NLRP3 inflammasome, a different inflammasome containing NLRC4 is also important in protecting against infection with Candida albicans, and appears to be functioning in the mucosal lining of the mouth and intestines, rather than in immune cells. Our research explains a new mechanism of mucosal immunity to fungal infections and has broad implications for developing new treatments against fungal infections, which are a serious cause of illness and death, particularly in immunocompromised persons. Additionally, this research may also lead to new ways to identify those individuals who are at the highest risk for serious fungal infections.

Candida sp. are dimorphic fungi that commonly colonize the oral cavity of adult humans, with overgrowth prevented by competing commensal bacteria as well as local host immune responses. Perturbations of the normal oral flora through antibiotic treatment, for example, or immunocompromised states can lead to mucosal Candida overgrowth resulting in the development of oropharyngeal candidiasis (OPC, also known as thrush). Candida albicans has now been identified as the leading cause of fatal fungal infections, with mortality rates as high as 50%, and ranks 4th among all pathogens isolated from bloodstream and nosocomial infections [1]–[3]. Host recognition of Candida requires engagement of surface receptors on innate immune cells, including TLR2 and Dectin-1 [4]–[7]. A major consequence of receptor activation is the induction of pro-inflammatory gene expression including interleukin 1 beta (IL-1β), a zymogen which requires proteolytic processing by caspase-1 to become biologically active [8]–[11]. Activation of caspase-1 requires signaling through recently described protein complexes termed inflammasomes, consisting of either NOD-like receptor (NLR) molecules or the PYHIN protein, Absent in melanoma-2 (AIM2) [12]–[16]. NLRs are characterized by the presence of a Leucine Rich Repeat domain, a central NACHT domain involved in oligomerization and protein-protein interactions, and a CARD or PYRIN domain [17]. Conformational changes in NLR proteins, resulting from the introduction of activating stimuli, cause oligomerization of NLR proteins together with ASC adapters, permitting autocatalytic cleavage of pro-caspase-1 to an active state capable of cleaving pro-IL-1β. Although intracellular danger signals and crystalline compounds such as uric acid crystals, cholesterol crystals, amyloid and asbestos have been shown to activate the NLRP3 inflammasome [18]–[22], the precise mechanism(s) underlying inflammasome activation are not defined. Currently, several theories have been proposed for the molecular mechanisms underlying activation of the NLRP3 inflammasome including mitochondrial ROS production [23], phagosomal or endosomal rupture and cell membrane disturbances [24]–[27]. The NLRP3 inflammasome has been linked to IL-1β responses to pathogen-derived molecules including bacterial muramyl dipeptide [28] and toxins [20], [28], as well as in response to a range of bacterial, viral and fungal pathogens, including Candida albicans [6], [29]. Another NLR molecule, NLRC4, also forms an inflammasome capable of activating caspase-1 and IL-1β cleavage. During some bacterial infections, such as with Shigella, Salmonella, Pseudomonas or Legionella, NLRC4 detects inadvertently translocated flagellin or PrgJ rod protein, a component of the type III secretion system [30]–[35]. Although limited in vitro studies using NLRC4 deficient macrophages or dendritic cells challenged with Candida albicans revealed no defects in caspase-1-dependent IL-1β responses [29], [36], [37], the role of NLRC4 in live fungal infection models has not been thoroughly defined. In this study, we sought to examine the role of other inflammasome in anti-fungal defenses in vivo. We show that infection with Candida albicans leads to up-regulation of NLRP3 and NLRC4 expression in the oral mucosa and this induction is impaired in both NLRP3 and NLRC4 deficient mice. Additionally, we reveal a role for the NLRC4 inflammasome in regulating resistance to mucosal infection with Candida as well as preventing systemic dissemination. We show that inflammasome driven IL-1β responses via both the NLRC4 and NLRP3 inflammasome are essential for epithelial antimicrobial peptide production, and other inflammatory responses including IL-18 and IL-17 in response to Candida infection. Inflammatory cell recruitment to Candida infected oral mucosa is significantly impaired in NLRC4 deficient mice compared to wild-type mice. Using bone marrow chimera mice, we reveal that the activity of NLRC4 is mediated at the level of the mucosal stroma, in contrast to that observed with NLRP3 which is active in both hematopoietic and stromal compartments. Collectively our studies show that, in addition to the NLRP3 inflammasome, there is a tissue specific role for the NLRC4 inflammasome in host sensing and immune defense to non-bacterial pathogens such as Candida albicans. During Candida infection, the oral mucosa acts as a physical barrier to infection as well as the initial tissue to respond to fungal growth and invasion. To assess the impact of Candida treatment on the oral mucosa, we monitored gene expression levels in oral mucosa by quantitative real-time PCR. We first examined the level of NLR expression in buccal tissues of Candida infected mice and observed a strong induction of NLRP3 in wild-type mice following oral challenge with Candida albicans. Induction of NLRP3 was significantly reduced in both Nlrc4−/− and Asc−/− mice (Figure 1A). Similarly, NLRC4 was induced in WT mice and negligible in Nlrp3−/− and Asc−/− mice (Figure 1B). As expected, ASC was not induced in any of the strains after Candida infection. These data indicate that genetic knockdown of a single NLR may have profound effects on the expression profile of other NLR proteins and is, to our knowledge, the first evidence of cross-regulation of NLRP3 and NLRC4. We next assessed the impact of the NLRP3 and NLRC4 inflammasomes on expression levels of members of the IL-1 family. There was a significant difference in the induction of IL-1β between the WT and Nlrc4−/−, Nlrp3−/−, and Asc−/− mice (Figure 1C). This defect in IL-1β production was confirmed in the serum of infected mice at 3 days of infection (Figure 1D). Levels of IL-1R1 expression were similar between WT and Nlrc4−/− or Asc−/− mice with reduced induction observed in Nlrp3−/− mice, although this was not significant (Figure 1E). Induction of IL-1R antagonist (IL-1Rn) was not significantly different between any of the inflammasome knockout mice and WT mice (Figure 1F). Overall, the induction of IL-1R1 and IL-1Rn was minimal compared to IL-1β in all the infected mice. Our previous studies demonstrated that NLRP3 signaling is critical for the prevention of fungal growth as well as dissemination in a murine model of oropharyngeal candidiasis [6]. A role for the NLRC4 inflammasome in response to oral fungal challenge has yet to be characterized. In order to ascertain the impact of loss of NLRC4 function on disease progression, we infected wild-type (WT) and Nlrc4−/− mice with Candida albicans as previously described [6]. Oral fungal burdens were elevated in Nlrc4−/− mice compared to WT mice by day 7, and persistently higher fungal burdens were observed to day 21 (Figure 2A). In our model of persistent, low virulence oral candidiasis, WT mice rarely show blood borne dissemination of infection, as measured by quantitative fungal burdens in the kidneys (Figure 2B). In contrast, Nlrc4−/− mice show a significantly increased susceptibility to dissemination of infection, peaking at day 7 but returning to WT levels by day 21. In agreement with these findings, Nlrc4−/− mice also had elevated gross clinical scores, a qualitative measure of oral infection severity, at all time points (Figure 2C). Survival in the Nlrc4−/− mice was reduced when compared to WT mice when infected with a virulent strain of Candida albicans (Figure 2D). Elevated quantitative fungal colonization was observed in tissues of the gastrointestinal tract including esophagus, stomach, and small intestine in Nlrc4−/− mice compared to WT (Figure S1). These data contrast with studies in Nlrp3−/− mice, in which it was determined that oral fungal colonization was similar at day 3, becoming slightly elevated at day 7 and 14, and returning to WT levels by day 21 (Figure 2A). These mice exhibited elevated levels of systemic dissemination throughout the 21 day timecourse (Figure 2B). By day 21, the gross clinical score of both Nlrp3−/− and WT mice were between 0 and 1, indicating minimal signs of infection, which contrasts the sustained elevated clinical score seen in Nlrc4−/− mice (Figure 2C). Taken together, our studies imply that NLRC4 and NLRP3 are differentially functioning in the innate response to Candida infection, with NLRC4 playing a more prominent role in the clearance of oral infection. One of the earliest inflammatory cells that migrate to the site of microbial infection are neutrophils, and this chemotaxis is necessary for proper inflammatory responses and anti-microbial defenses. Given the known capacity for IL-1β to mediate leukocyte infiltration into infected tissues, we used histology to examine the impact of inflammasome deficiency on cellular infiltration to the mucosa of the tongue. By day 2, a robust cellular infiltration was observed in the dorsal epithelium of a WT tongue, particularly in areas showing the presence of fungal hyphae and epithelial erosion (Figure 3A). These cells morphologically appear to have multi-lobulated nuclei, consistent with neutrophils. In contrast, minimal cellular infiltration was observed in a tongue from Nlrc4−/− mouse, despite the presence of erosive lesions and fungal hyphae (Figure 3C). Nlrp3−/− and Asc−/− tongues exhibited cellular infiltration, although not to the extent of WT; and these areas of concentrated cellular infiltrates also correlated with the presence of fungal hyphae and tissue erosion (Figure 3E, G). Neutrophils have been implicated in the control of a range of microbial infections, including Candida [38]–[40]. Given the presence of significant cellular infiltration at 2d post infection, we sought to specifically characterize the extent of neutrophil infiltration in these tissues. Using a monoclonal antibody shown to specifically stain neutrophils, we observed significant neutrophil staining in the outer epithelium of the WT tongue. This immunofluorescent staining localized to the regions of increased cellularity observed in the epithelium with PAS/H staining (Figure 3B). Neutrophils were also observed throughout the sub-mucosal tissue. In agreement with our finding with PAS/H staining, neutrophil influx into the Nlrc4−/− was drastically reduced (Figure 3D). As expected, a significant influx of neutrophils was observed in both Nlrp3−/− and Asc−/− tongues (Figure 3F, H). Intriguingly, it was observed that not only was there a reduction in neutrophil infiltration in the Nlrc4−/− tongue but the neutrophils present failed to infiltrate the epithelium where the presence of hyphae was detected. This is evidenced by a ∼25 fold reduction in the percent of dorsal epithelium that stains positive for neutrophils in Nlrc4−/− mice when compared to WT (Figure 4). Nlrp3−/− and Asc−/− mice showed a ∼3 fold decrease in positive staining (Figure 4). These findings indicate that the activation of Nlrc4−/− is required for neutrophil recruitment into infected tissues and proper trafficking to the site of active fungal infection. Recent reports have implicated the IL-17 family as a critical mediator of protective host responses to a range of extracellular pathogens, including Candida [41]–[45]. To assess the impact of inflammasome activation on IL-17 in our model of oral candidiasis, we measured expression levels of IL-17 family members in oral mucosal tissues after infection. A robust increase in IL-17A and IL-17F gene expression was detected in the oral mucosal tissue of WT animals, which was significantly reduced in Nlrc4−/−, Nlrp3−/−, and Asc−/− mice (Figure 5A, B). In contrast, the induction of IL-17F was dependent on NLRP3, but not NLRC4 or ASC (Figure 5B). A robust induction of interleukin 17A receptor (IL-17RA) expression was detected in WT, Nlrp3−/−, and Asc−/− mice while this response was abrogated in Nlrc4−/− mice (Figure 5C). As many downstream inflammatory responses are dependent on IL-17, the failure to upregulate the IL-17A receptor may have implications for local anti-Candida inflammation and chemotaxis of inflammatory cells. We next examined expression of other inflammatory cytokines in the oral mucosa of mice infected with Candida albicans. We observed significantly lower induction of IL-18, another cytokine requiring inflammasome mediated cleavage, in Nlrp3−/− and Asc−/− mice, while Nlrc4−/− mice showed a slight reduction compared to WT which was not statistically different (Figure 5D). As shown in Figure 5E, murine CXCL1, a homolog of human IL-8, was dramatically induced in WT mice following Candida infection and levels were significantly reduced in all strains of inflammasome deficient mice. Induction of the pro-inflammatory cytokine IL-6 was also significantly reduced in Nlrc4−/−, Nlrp3−/−, and Asc−/− mice compared to WT (Figure 5F). In addition to inflammatory cytokines responses following pathogenic encounter, immune cells in the oral mucosa, including epithelial cells, release small antimicrobial peptides designed to disrupt microbial function as well as act as strong chemoattractant signals for the migration of inflammatory cells such as neutrophils and macrophages [46]–[48]. To better define the inflammasome dependence of antimicrobial peptide responses in our murine model of OPC, we quantified several murine beta-defensins as well as cathelicidin responses in the oral mucosa following fungal infection. In contrast to murine beta-defensin 1 (mBD1), which was not induced following infection in any of the mice (Figure 6A), mBD2, -3, -and 4 all showed elevated expression in oral tissues following Candida infection in WT mice (Figure 6B, C, D). Reduced or negligible up-regulation in mBD2 and mBD4 gene expression was observed in all of the inflammasome deficient mice (Figure 6B, D). In contrast, mBD3 induction was similar between WT and Nlrc4−/− mice but significantly reduced in Nlrp3−/− and Asc−/− mice (Figure 6C). No appreciable induction of mBD14, a murine homolog of human beta-defensin 3, was observed in any of the mice (Figure 6E). Another antimicrobial peptide, cathelicidin or CAMP, has also been implicated as an activator of the P2X7 receptor and a potential inducer of IL-1β release from cells [49]. Expression levels of CAMP were dramatically elevated in WT buccal tissue following Candida infection, and this induction was dependent on NLRC4, NLRP3 and ASC (Figure 6F). NLRC4 does not appear to be involved in inflammasome activation in innate immune cells exposed to Candida in vitro [29], [36], [37]. Yet in our in vivo model on mucosal candidiasis, NLRC4 is required for protection from mucosal colonization, prevention of early dissemination of infection, and neutrophil infiltration. Therefore, we hypothesized that the impact of NLRC4 activation during fungal infection was manifested in mucosal and/or stromal tissues versus hematopoietic cells. The innate immune system is comprised of cells of embryonic origin, including epithelial cells, as well as infiltrating leukocytes derived from the bone marrow. To assess the contribution of different inflammasome molecules in these compartments, we generated bone marrow chimera mice and infected them orally with C. albicans. Lethally irradiated recipient mice were reconstituted with bone marrow progenitor cells and allowed to fully reconstitute prior to infection. WT mice that were reconstituted with WT bone marrow exhibited no increase in oral infection or systemic dissemination of infection at 7 days when compared to non-chimeric WT mice, indicating that the chimera procedure does not predispose the mice to higher levels of fungal infection (Figure S2). WT mice reconstituted with Nlrc4−/− bone- marrow showed no significant difference in oral fungal colonization compared to WT/WT chimera mice (Figure 7A). In contrast, Nlrc4−/− mice reconstituted with WT bone marrow showed enhanced oral colonization with C. albicans (Figure 7A), to a degree that is similar to native Nlrc4−/− mice (Figure 2A). The Nlrc4−/− mice reconstituted with WT bone marrow also exhibited increased disease severity when compared to WT/WT chimera mice (Figure 7C). These results demonstrate that intact NLRC4 function in the stromal or epithelial compartment is associated with protection from mucosal infection with Candida albicans. To evaluate the role of the NLRP3 inflammasome in the stromal versus hematopoietic compartments, bone marrow chimeras were generated using Nlrp3−/− as well as Asc−/− mice. Nlrp3−/− mice receiving WT bone marrow showed difference in oral fungal burdens relative to WT/WT chimera controls (Figure 7A). However, WT mice reconstituted with Nlrp3−/− bone marrow exhibited significantly elevated oral infection when compared to WT/WT chimera mice (Figure 7A), although both sets of NLRP3 chimera mice showed elevated clinical scores compared to WT/WT chimera mice (Figure 7C). A similar pattern was seen with ASC chimeric mice (Figure 7A), where higher oral fungal colonization was observed in the WT mice receiving ASC deficient bone marrow compared to Asc−/− mice receiving WT cells or WT/WT chimera mice. These observations indicate that the function of the NLRP3/ASC inflammasome complex in infiltrating inflammatory cells is critical for control of oral mucosal infection. We next assessed the role of NLRC4 and NLRP3 in protection from systemic dissemination of infection. As a marker of dissemination, we quantified the fungal burdens of the kidneys in bone marrow chimera mice infected with C. albicans for 7 days. As shown in Figure 7B, neither Nlrc4−/− donor nor recipient chimera mice showed dissemination to the degree seen in native Nlrc4−/− mice (Figure 2B), although a trend towards increased dissemination was observed compared to WT/WT chimera mice. In contrast, both Nlrp3−/− donor and recipient chimera mice showed higher systemic dissemination when compared to WT/WT chimera mice (Figure 7B). WT mice receiving Asc−/− bone marrow showed enhanced dissemination of infection, whereas Asc−/− mice receiving WT bone marrow showed similar kidney fungal burdens to WT/WT chimera mice (Figure 7B). These results demonstrate that the NLRP3/ASC inflammasome plays a dominant role in protection against disseminated fungal infection compared to the NLRC4 inflammasome which plays a role in protection of the host from mucosal infection. Disseminated fungal infections present a significant health risk to both immune-competent and -compromised individuals, making studies into early host immune responses involved in the prevention of dissemination critical for the development of new therapeutic approaches. The release of inflammatory mediators from resident cells at the site of infection is critical for antimicrobial responses including the recruitment of inflammatory cells such as neutrophils and macrophages. IL-1β has been implicated in protective host immune responses to a range of infectious pathogens including viruses, bacteria, and fungi [11]. We and others have previously shown that the NLRP3 inflammasome is important for control of Candida infection in both mucosal and disseminated models [6], [29], [36], [37]. Here we present the first experimental evidence implicating the NLRC4 inflammasome in the induction of protective host responses to challenge by a fungal pathogen. Our studies reveal that the NLRC4 inflammasome is important for the control of Candida albicans infection in vivo, particularly in the oral cavity, with increased oral fungal colonization and disease severity observed in Nlrc4−/− mice. Lack of NLRC4 also increased susceptibility to disseminated fungal infection, particularly early in infection. Our model of OPC using a clinical isolate recapitulates human oral infections in which mortality is rarely observed. We carried out OPC infection studies using a highly virulent strain of C. albicans (ATCC 90234) and observed similar increases in oral colonization and dissemination at 7d to that observed using the oral isolate GDH2346 (Figure S1). However, the virulent strain resulted in significant mortality in Nlrc4−/− mice when compared to WT (Figure 2E). When we examined cellular infiltrations in infected tongues, we observed a substantial impact in both cellular infiltration, and specifically neutrophil-influx in the absence of NLRC4, compared to a robust neutrophil infiltration into the infected epithelium of WT mice. Interestingly, the neutrophils that were present in the Nlcr4−/− tongue remained in the sub-mucosa. This defect in neutrophil infiltration into the epithelium of the tongue may explain the extended defect in oral clearance observed in Nlrc4−/− mice throughout the course of our OPC infection studies. The importance of neutrophils in anti-fungal defenses is well known. Recent studies have shown that the innate inflammatory mileu is critical for effective neutrophil activity against fungal pathogens including IL-6, GM-CSF, and IL-17 responses [38], [41], [42], [50], [51]. In order to better define the molecular basis for protection from mucosal infection, we evaluated a panel of innate inflammatory responses in the oral mucosa from Candida infected mice. Previous studies have demonstrated that NLRP3 expression is inducible following infection and implicate this as the rate limiting step in inflammasome activation [52], [53] but little is known about the regulation of NLRC4 expression. IL-1β has been shown to increase the expression level of other pro-inflammatory cytokines following IL-1 receptor engagement [54]-[56], and IL-1 receptor activation may result in the increased expression of inflammasome proteins in the responding cells, priming them for quick activation. We show that Candida infection up-regulates NLRP3 and NLRC4 expression in mucosal tissues compared to mock infected mice. Additionally, we demonstrate that this induction is impaired in Nlrp3−/−, Asc−/− and Nlrp4−/− mice. As expected, ASC expression was found to be constitutive and not induced by Candida infection (data not shown). Our data provides novel insight into transcriptional changes induced by activation of inflammasome complex by an infectious pathogen. The impairment of mucosal NLRP3 induction in NLRC4 deficient mice may enhance the impact of the lack of this receptor on susceptibility to mucosal infection. As expected, we found that IL-1β was up-regulated in wild-type mice in response to fungal infection. However, these responses were partially abrogated in NLRC4 deficient mice and absent in NLRP3 and ASC deficient mice. Similarly, the IL-1 receptor 1 (IL1R1) and IL-1 receptor antagonist (IL1Rn) were poorly induced by Candida infection in WT and inflammasome deficient mice. Consistent with published studies, we did not observe a role for the NLRC4 inflammasome in IL-1β induction or processing in inflammatory cells stimulated in vitro with Candida albicans [29], [36], [37]. IL-18 is another member of the IL-1 family which requires proteolytic enzyme cleavage for activation [57], [58]. Although caspase-1 mediated cleavage of pro-IL-18 into the bioactive form is the accepted paradigm, alternative mechanisms for cleavage have been proposed including PR3, granzyme B and mast cell chymase [59], [60]. In our studies, in vivo IL-18 induction by Candida infection showed a similar pattern in mucosal tissues as IL-1β. Next, we investigated the regulation of the cytokine IL-17, which has been shown to play a role in anti-fungal immunity in humans and in animal models [41], [42], [44], [45]. In response to infection with Candida albicans, IL-17A, IL-17F and IL-17RA were up-regulated in mucosal tissues. Interesting patterns of induction of the IL-17 family by Candida infection were observed; IL-17A was dependent on NLRC4, NLRP3 and ASC whereas IL-17F was only dependent on NLRP3 with comparable induction seen in WT, NLRC4 and ASC deficient mice. The induction of IL-17RA was only dependent on NLRC4. The inflammasome dependence of IL-17 family responses during microbial infection is not well understood, and the regulation of these genes is likely multi-factorial, including dependence on IL-1β as well as other inflammatory mediators. The cytokine IL-6 was induced in wild-type mice in response to Candida infection but was significantly reduced in NLRC4 deficient mice and completely abrogated in NLRP3 and ASC deficient mice. The chemokine KC (CXCL1), a murine homolog of human IL-8, was also found to be highly dependent on IL-1β as NLRC4, NLRP3, and ASC deficient mice exhibited significant decreases in gene expression following fungal challenge. Although IL-6 and KC/IL-8 are not known to be major mediators of anti-fungal immunity, they play an important role in the cytokine network of innate cellular communication. IL-6 has been shown to be directly up-regulated in macrophages by Candida cell wall components [61] but can also be induced by secondary effect of other inflammatory responses. IL-6 signaling has also been shown to be important in the recruitment of neutrophils in response to Candida [38], [62]. IL-6 production is regulated by a complex network of signaling pathways that include NF-κB as well as a newly described pathway mediated by the protein tyrosine phosphatase, Src homology domain 2-containing tyrosine phosphatase-1 (SHP-1) leading to activation of Erk1/2–C/EBPβ [63]. This is particularly relevant to mucosal anti-fungal immunity as SHP-1, also known as PTPN6 and PTP1C, has been shown to negatively regulate the effects of epidermal growth factor [64], an important regulator of epithelial homeostasis, as well as affect tight junction formation in epithelium [65]. We propose, based on our histological analysis, that the defect in the induction of inflammatory mediators observed in mice deficient in the NLRP3/ASC inflammasome is most likely due to defective functioning of inflammatory leukocytes in the absence of these proteins. In contrast, the partial defect in inflammatory responses observed in mucosal tissues from infected NLRC4−/− mice is likely due to impaired infiltration of immune cells into infected tissue. Another key innate immune response during infection is the release of antimicrobial peptides designed to limit pathogen growth and survival. Antimicrobial peptides (AMPs) consist of a diverse group of small cationic peptides including the defensins, cationic and amphipathic peptides which have broad antimicrobial and chemotactic properties. Beta-defensins are primarily secreted by epithelial cells and play an important role in the microbial homeostasis of the skin, oral cavity, lung and gut. Human β-defensin (hBD)-1 is primarily expressed in the urinary and respiratory tracts [66], [67] and although constitutively expressed, may be up-regulated by infection or inflammation. A defect in hBD-1 activity in the lung has been associated with cystic fibrosis [68], [69]. Polymorphisms in the defensin-1 gene, defB1, have been associated with low oral colonization with Candida albicans (Jurevic 2003), protection from HIV [70]–[72], chronic obstructive pulmonary disease [73] and Crohn's disease [74]. The murine homolog of hBD-1, murine β-defensin (mBD)-1, is also expressed by epithelial surfaces, lung and kidney and has salt sensitive antimicrobial activity [75], although its role in antifungal defense is unclear. Both hBD-2 and hBD-3 have known anti-Candida [46], [76]–[78] as well as anti-HIV activity [79], [80]. The role of mBD-2, the murine homolog of hBD-2, in oral mucosal health is unclear although its expression in the lung is highly inducible by LPS [81]. The murine ortholog of hBD-3, mBD-14, has inducible expression in the respiratory and intestinal tracts as well as in dendritic cells and shows anti-bacterial and chemotactic activity [82]. To better define the role of AMPs in anti-fungal defense, we examined AMP responses in oral mucosa after infection with Candida albicans. We show that mBD-1 appeared to be constitutively expressed in WT mice; however, gene expression was inhibited in NLRP3 and ASC deficient mice. We discovered that mBD-2, -4 and -14 were highly dependent on inflammasome activation as both NLRC4 and NLRP3 as well as ASC deficient mice exhibited dramatically reduced expression levels compared to WT. Interestingly, mBD-3 responses were found to have little dependence on NLRC4 but were dependent on NLRP3 and ASC. Another class of AMPs, the cathelicidins, consisting of human LL-37 and murine CAMP (also known as CRAMP), are known to have anti-Candida as well as chemotactic activity [83]–[87]. We observed that CAMP was highly up-regulated in WT mucosa in infected mice, but not in NLRC4, NLRP3 or ASC deficient mice. A recent report found that IL-17A augmented vitamin D3-mediated CAMP production in keratinocytes during psoriasis [88]. In concurrence with these findings, we observed abrogated CAMP expression in NLRC4 and NLRP3 deficient mice, which also lacked IL-17A gene expression, following Candida challenge. In addition to its direct antimicrobial effects, CAMP has been identified as a modulator of the P2X7R which has a known role in ATP-induced IL-1β release [49], [89], [90]. From our studies, it can be inferred that initial production of IL-1β may induce IL-17A and CAMP production which can in turn positively regulate further production of IL-1β to create an inflammatory environment which limits fungal infection. This mechanism may serve to explain the strong in vivo phenotype observed in our model for NLRC4 and NLRP3 deficient mice, perhaps via a failure to engage this positive feedback loop resulting in an immune state that is prone to persistent infection. In addition to driving IL-1β and IL-18 responses, the NLRC4 inflammasome has been shown to induce a specialized form of programmed cell death, termed pyroptosis or pyronecrosis, characterized by the release of cytoplasmic contents, which include inflammatory mediators such as ATP and arachidonic acid metabolites, to the extracellular matrix. A defect in pyroptosis may partially account for the critical role for NLRC4 activation in our model of candidiasis and provides an opportunity for future research. Interestingly, our data shows that activation of the NLRC4 inflammasome is important in the stromal compartment, where its role is critical for in vivo anti-fungal host defense, but not in the hematopoietic compartment. Using bone marrow chimera mice we differentiated between inflammasome activity in hematopoietic derived cells such as infiltrating macrophages and neutrophils, and embryonic derived mucosal tissues in our model of oropharyngeal candidiasis. This approach demonstrated that NLRP3 and ASC activity in both hematopoietic and stromal compartments are important for protection against oral infection and dissemination. This is in agreement with our previously published findings that the NLRP3 inflammasome was the primary mediator of IL-1β cleavage in murine macrophages stimulated with Candida in vitro [6]. In contrast, in vivo infection of bone marrow chimera mice showed higher mucosal colonization in NLRC4 deficient recipient mice reconstituted with WT inflammatory cells than WT recipients reconstituted with Nlrc4-/- cells, which had similar levels of oral mucosal infection as WT controls. Overall, these studies utilizing chimera mice in our murine model of mucosal fungal infection implicate a novel, tissue-specific role for the NLRC4 inflammasome. Many key questions are raised by the findings in this paper. Known microbial activators of the NLRC4 inflammasome include Salmonella typhimurium [91], Shigella flexneri [92], [93], Legionella pneumophila [94] and Pseudomonas aeruginosa [34]. Previous reports implicated the activation of NLRC4 by the release of flagellin through the Type-III secretion apparatus and by components of the basal rod proteins of the Type III secretion system itself [93], [95], [96]. Despite these findings, it still remains unclear the mechanism by which NLRC4 senses these activators. Given the homology between the known bacterial activators of NLRC4, it is possible that these proteins may function as a direct receptor recognizing a conserved sequence or structural feature. Current models of NLRP3 activation indicate it does not act as a traditional receptor but rather as a nexus for different pathways invoked following cellular injury and/or infection, which may also be true for the NLRC4 inflammasome. Future studies will seek to elucidate the mechanism of NLRC4 recognition of its activators and also identify the molecule(s) in Candida that induce NLRC4 activation. We hypothesize that mucosal NLRC4 activation may occur as an early event in fungal infection, perhaps as a result of cellular damage or direct effect of infection, leading to the induction of innate responses such as anti-microbial peptides and cytokines that recruit inflammatory cells including neutrophils and macrophages that infiltrate the sites of infection. Candida induced activation of the NLRP3/ASC inflammasome then provides a critical amplification of the innate response leading to protection of the host from overwhelming mucosal and disseminated candidiasis. The animals described in this study were housed in the AAALAC accredited facilities of the Case Western Reserve University School of Medicine. All animal use protocols have been approved by the Institutional Animal Care and Use Committee of Case Western Reserve University and adhere to national guidelines published in Guide for the Care and Use of Laboratory Animals, 8th Ed., National Academies Press, 2001. Candida albicans strains GDH2346 (NCYC 1467), a clinical strain originally isolated from a patient with denture stomatitis, or ATCC 90234 were utilized for in vitro and in vivo studies. Master plates were maintained on Sabouraud Dextrose (SD) agar. For OPC infection, yeast were grown for 12–16 h in SD broth, pelleted at 3000 rpm for 5 min and washed 2x with sterile 1X PBS. Yeast cells were manually counted using a hemocytometer and diluted to 5×107 cells/mL for live infection. Wild-type C57BL/6 mice were purchased from Jackson Laboratories. Nlrp3−/−, Nlrc4−/− and Asc−/− mice were generated by Millenium Pharmaceuticals. Animals were housed in filter-covered micro-isolator cages in ventilated racks. Infection and organ harvesting was performed as described previously [6]. Briefly, after pre-treatment with antibiotic containing water, the mice were anesthetized and light scratches made on the dorsum of the tongue following by the introduction of 5×106 C. albicans yeast. The scratches are superficial, limited to the outermost stratum corneum, and do not cause trauma or bleeding. After infection of 3 to 21 d, the mice were euthanized, organs harvested and homogenized and fungal burdens assessed by growth on SD agar. For gross clinical score assessment, visual inspection of fungal burdens on the tongues was performed under a dissection microscope. A score of 0 indicates the appearance of a normal tongue, with intact light reflection and no visible Candida, a score of 1 denotes isolated patches of fungus, a score of 2 when confluent patches of fungus are observed throughout the oral cavity, and a score of 3 indicates the presence of wide-spread fungal plaques and erosive mucosal lesions. For assessment of inflammatory gene induction, buccal tissue was isolated from infected mice and immediately placed into a RNA stabilization reagent (RNAlater, Qiagen). After homogenization in lysis buffer for 1.5 min using a bead-beater homogenizer (Retsch), total RNA was isolated using PrepEase RNA Spin Kit (USB/Affymetrix) followed by conversion to cDNA using SuperScript III Reverse Transcriptase (Invitrogen). Whole blood was collected via retro-orbital bleeding into EDTA pre-coated tubes, followed by centrifugation and removal of serum. Serum was stored at −80°C until used. Quantitative real-time PCR was done as described [6]. Specific primer sequences are listed in Table S1. Cytokines were measured in serum by ELISA (R&D). NCBI gene accession numbers are as follows: Nlrc4:NM_001033367.3; Nlrp3: NM_145827.3; Asc: NM_023258.4; Il1b: NM_008361.3; Il1r1: NM_008362.2; Il1rn: NM_031167.5; Il17a: NM_010552.3; Il17f: NM_145856.2; Il17ra: NM_008359.2; Il18: NM_008360.1; Cxcl1: NM_008176.3; Il6: NM_031168.1; Defb1: NM_007843.3; Defb2:NM_010030.1; Defb3: NM_013756.2; Defb4: NM_019728.4; Defb14: NM_183026.2; Camp: NM_009921.2. Intact tongues are removed at necropsy, immediately immersed in Tissue Freezing Medium (EMS) and flash frozen in liquid nitrogen. After cryo-sectioning, 5 µm sections were fixed with10% formalin for 2 min then stained with Periodic Acid Schiff and Hematoxylin (PAS/H). For immunofluorescent staining, the sections were blocked with 5% normal goat serum/PBS, stained with rat monoclonal anti-neutrophil primary antibody (NIMP-R14; specific for Ly-6G and Ly-6C) and Alexa Fluor 488-conjugated goat anti-rat secondary antibody (Invitrogen) and mounted in Vectashield containing DAPI (Vector Laboratories). For quantitative analysis, images were taken using a Leica DMI 6000B inverted microscope, and the number of neutrophils in each section was digitally quantified using the imaging program MetaMorph (Molecular Devices). Briefly, the number of pixels in a region containing the dorsal epithelial portion of the tongue was counted. A threshold value was then assigned which corresponds to a minimum fluorescent value of a neutrophil. The number of pixels at or above this threshold was determined and a percentage of fluorescent pixels determined by dividing by total overall number of pixels. Lethally irradiated mice (exposed to a Cesium-139 γ-radiation source for a total full body dose of 900 rads) received 5×106 bone marrow cells from pooled donor mice via tail vein injection and allowed to recover for 4 weeks. Data were analyzed using commercial software (GraphPad) and Student's two-sample independent t tests or Mann Whitney U tests were used for comparative statistical analysis of qPCR, ELISA, and quantitative fungal load data. Comparison of survival curves was done using a mean Logrank test. P values are presented when statistical significance was observed (significance set at P≤0.05 at a confidence interval of 95%).

10.1371/journal.pntd.0000721

Structural Optimization and De Novo Design of Dengue Virus Entry Inhibitory Peptides

Viral fusogenic envelope proteins are important targets for the development of inhibitors of viral entry. We report an approach for the computational design of peptide inhibitors of the dengue 2 virus (DENV-2) envelope (E) protein using high-resolution structural data from a pre-entry dimeric form of the protein. By using predictive strategies together with computational optimization of binding “pseudoenergies”, we were able to design multiple peptide sequences that showed low micromolar viral entry inhibitory activity. The two most active peptides, DN57opt and 1OAN1, were designed to displace regions in the domain II hinge, and the first domain I/domain II beta sheet connection, respectively, and show fifty percent inhibitory concentrations of 8 and 7 µM respectively in a focus forming unit assay. The antiviral peptides were shown to interfere with virus:cell binding, interact directly with the E proteins and also cause changes to the viral surface using biolayer interferometry and cryo-electron microscopy, respectively. These peptides may be useful for characterization of intermediate states in the membrane fusion process, investigation of DENV receptor molecules, and as lead compounds for drug discovery.

Virus surface proteins mediate interactions with target cells during the initial events in the process of infection. Inhibiting these proteins is therefore a major target for the development of antiviral drugs. However, there are a very large number of different viruses, each with their own distinct surface proteins and, with just a few exceptions, it is not clear how to build novel molecules to inhibit them. Here we applied a computational binding optimization strategy to an atomic resolution structure of dengue virus serotype 2 envelope protein to generate peptide sequences that should interact strongly with this protein. We picked dengue virus as a target because it is the causative agent for the most important mosquito transmitted viral disease. Out of a small number of candidates designed and tested, we identified two different highly inhibitory peptides. To verify our results, we showed that these peptides block virus:cell binding, interfere with a step during viral entry, alter the surface structure of dengue viral particles, and that they interact directly with dengue virus envelope protein. We expect that our approach may be generally applicable to other viral surface proteins where a high resolution structure is available.

Fusogenic viral envelope glycoproteins are multimeric proteins that facilitate the fusion of viral and target cell lipid membranes during the initiation of infection. The membrane fusion process is energetically favorable and essentially irreversible, but has a considerable kinetic energy barrier [1]. These proteins allow rapid membrane fusion by drawing the opposing membranes together and either stabilizing or providing the activation energy to surmount the transition state [1], [2]. In this way, they behave in many aspects like a fusion catalyst. Because they effect a macromolecular process that involves large scale conformational changes in the substrate membranes and the proteins themselves, these proteins possess multiple interacting surfaces that could be targeted by inhibitors [3]. There are several distinct types of viral fusion proteins, including the class I, primarily alpha helical proteins (such as HIV TM and influenza HA), the class II, primarily beta sheet proteins (such as the flavivirus E and alphavirus E1), and mixed helix/sheet proteins (including herpes virus gB and rhabdovirus G) [3], [4]. To date, most progress with viral fusion protein inhibitors has focused on class I alpha helical proteins. The HIV TM protein provides an excellent example of targeting distinct, interacting surfaces for inhibition. The HIV TM functions as a homotrimer with each monomer contributing two alpha helical regions that interact to form a post-fusion six-helix bundle. Inhibition of the formation of this six-helix bundle can be accomplished by exogenous peptides mimicking either of the two reciprocally interacting helices [5]–[7]. Only a few examples of viral entry inhibitors with activity against the primarily beta sheet envelope proteins (E) from flaviviruses have been described [8]–[10]. However, few of these have taken advantage of the available crystal structures of flavivirus E proteins, including both pre-fusion and post-fusion forms [11]–[22]. The authors of some of these structures have predicted that several regions of these proteins might be targets for inhibition [11], [14], [15]. Here we report the use of structural data from the pre-fusion dengue virus-2 (DENV-2) E protein as a model for a computational approach to the design of new peptide inhibitors of DENV-2 entry. This approach makes use of a residue-specific all-atom probability discriminatory function (RAPDF) score to identify in situ amino acid sequences that are likely to have high structural and binding stability [23], [24]. Out of seven computationally designed peptides that were synthesized and tested, two were identified as possessing fifty percent in vitro inhibitory activity (IC50) below 10 µM and another with IC50 activity below 40 µM. Two of the inhibitors (DN57opt and DN81opt) are binding optimized variants of peptides originally designed from DENV inhibitory peptide sequences located in domain II near the domain I/domain II hinge region [9]. The other (1OAN1) is an entirely novel peptide designed from an extended beta sheet region comprising the first connection between domains I and II. We show that the two peptides with the highest inhibitory activity interfere with virus:cell binding, cause structural changes to the surface of DENV-2 virions, and bind specifically to purified DENV-2 E protein. The causative agent of dengue fever, dengue hemorrhagic fever and dengue shock syndrome, DENV has emerged in the past several decades as the most important mosquito borne viral disease with an estimated 2.5 billion people living in areas at risk for epidemic transmission and 50–100 million people infected annually [25], [26]. Complicating this situation, the four distinct serotypes of DENV generate only low level immunological cross protection, allowing for repeated epidemic outbreaks in the same populations [27], [28]. The phenomenon of antibody dependent enhancement has been shown to result in more severe disease in individuals who have been previously infected with a different serotype [29]–[33]. With no specific treatment or prevention available other than vector control, DENV is an important target for the development of antivirals and vaccines. The results presented here indicate that the DENV E glycoprotein has multiple accessible surfaces that can be targeted by distinct inhibitors and is an amenable target for rational inhibitor design. Peptide inhibitors were designed to have improved in situ binding compared to naturally occurring sequences using the residue-specific all-atom probability discriminatory function (RAPDF) [24]. The x-ray diffraction structure of DENV-2 envelope protein (Protein Data Bank identifier 1OAN) was used as a template for creating mutant structures from which the peptides were derived [14]. For each peptide, we randomly selected a residue side chain and substituted it with a new side chain. The substitution was performed using a backbone-dependent side chain rotamer library and a linear repulsive steric energy term provided by SCWRL version 3.0 [34]. The resulting all-atom models were energy minimized for 200 steps using the Energy Calculation and Dynamics (ENCAD) program [35]–[37]. RAPDF scores were then calculated to estimate the structural stability of a given E protein structure derivative. For a selected residue, side chain substitution was carried out ten times. The amino acid that produced the best RAPDF score was selected and used as a template for further mutation. The entire mutation process was repeated 100,000 times to enable a rigorous search for peptides that produced the best RAPDF score (i.e., highest predicted stability). A 20 residue acid sliding window that moved from the N to the C terminus of the E protein in 10 residue acid increments was evaluated by a structural stability (pseudoenergy) optimization protocol using the RAPDF. A Metropolis Monte Carlo search algorithm [38] was used to change each amino acid in the 20 residue window to one of the other 19 naturally occurring amino acids, and the stability of corresponding peptide in the context of the entire E protein structure was evaluated. This process was iterated 100,000 times using RAPDF as the target scoring function. The Metropolis criterion was used to select a particular change in the simulation: if a particular change resulted in a better RAPDF score (lower pseudoenergy), then it was retained. If a particular change resulted in a worse RAPDF score (higher pseudoenergy), then a random choice, based on the score difference between the previous change and the current one, was made to retain the corresponding change. This procedure enables not only enables design of peptides that will result in high structural and binding stability (i.e., the best RAPDF scores/pseudoenergies), but also enables surmounting local minima encountered during the search. Computational optimization was performed on the four regions corresponding to the best RAPDF score, and therefore the highest binding potential, within the E protein as described above to generate variant peptides sequences. DENV-2 strain NG-C was obtained from R. Tesh at the University of Texas at Galveston. Virus was propagated in the Macaca mulatta kidney epithelial cell line, LLC-MK2 (ATCC catalog number CCL-7). Cells were grown in Dulbecco's modified eagle medium (DMEM) with 10% (v/v) fetal bovine serum (FBS), 2 mM Glutamax, 100 U/ml penicillin G, 100 µg/ml streptomycin and 0.25 µg/ml amphotericin B, at 37°C with 5% (v/v) CO2. Peptides were synthesized by solid-phase N-α-9-flurenylmethyloxycarbonyl chemistry, purified by reverse-phase high performance liquid chromatography and confirmed by amino acid analysis and electrospray mass spectrometry (Genemed Synthesis, San Antonio, TX). Peptide stock solutions were prepared in 20% (v/v) dimethyl sulfoxide (DMSO): 80% (v/v) H2O, and concentrations determined by absorbance of aromatic side chains at 280 nm. LLC-MK2 target cells were seeded at a density of 1×105 cells in each well of a 6-well plate 24 h prior to infection. Approximately 200 focus forming units (FFU) of virus were incubated with or without peptide in serum-free DMEM for 1 h at rt. Virus/peptide or virus/control mixtures were allowed to infect confluent target cell monolayers for 1 h at 37°C, with rocking every 15 m, after which time the medium was aspirated and overlaid with fresh DMEM/10% (v/v) FBS containing 0.85% (w/v) Sea-Plaque Agarose (Cambrex Bio Science, Rockland, ME) without rinsing. Cells with agar overlays were incubated at 4°C for 20 m to set the agar. Infected cells were then incubated at 37°C with 5% CO2 for 5 days. Infected cultures were fixed with 10% formalin overnight at 4°C, permeablized with 70% (v/v) ethanol for 20 m, and rinsed with phosphate buffered saline, pH 7.4 (PBS) prior to immunostaining. Virus foci were detected using a specific mouse mAb from hybridoma E60 (obtained from M. Diamond at Washington University), followed by horseradish peroxidase-conjugated goat anti-mouse immunoglobulin (Pierce, Rockford, IL), and developed using AEC chromogen substrate (Dako, Carpinteria, CA). Results are expressed as the average of at least two independent trials with three replicates each. IC50 values were determined using variable slope sigmoidal dose-response curve fits with GraphPad Prism 4.0 software (LaJolla, CA), except for DN81opt, which was determined graphically due to a lack of data points to produce a reasonable curve fit. Cytotoxicity of peptides was measured by monitoring mitochondrial reductase activity using the TACS™ MTT cell proliferation assay (R&D Systems, Inc., Minneapolis, MN) according to the manufacturer's instructions. Dilutions of peptides in serum-free DMEM were added to confluent monolayers of LLC-MK2 cells in 96-well plates for 1 h at 37°C, similar to the focus forming inhibition assays, and incubated at 37°C with 5% (v/v) CO2 for 24 h. Absorbance at 560 nm was measured using a Tecan GeniosPro plate reader (Tecan US, Durham, NC). DENV-2 NGC strain used for the cryoEM reconstructions was propagated in mosquito C6/36 cells. Virus was purified by precipitation with 40% PEG 8000 and then ultracentrifugation onto a 25% sucrose cushion. Virus was further purified by banding on a 10%–30% potassium tartrate gradient. The virus band was removed and dialyzed against 12 mM Tris pH 8.0, 120 mM NaCl, 1 mM EDTA, and concentrated using a Millipore Centricon filter. Purified virus was mixed with 1OAN or DN57opt at a concentration of 1 molecule of peptide for every E protein on the surface of the virus. The complex was incubated for half an hour at 37°C followed by half an hour at 4°C and then flash frozen on holey carbon grids in liquid ethane. Images of the frozen complex were taken with a Philips CM200 FEG transmission electron microscope (Philips, Eindhoven, The Netherlands) at a magnification 51,040 using an electron dose of approximately, 25e-/Å 2 using a Charge-Couple device. Real time binding assays between peptides and purified DENV-2 S1 E protein were performed using biolayer interferometry on an Octet QK system (Fortebio, Menlo Park, CA). This system monitors interference of light reflected from the surface of a fiber optic sensor to measure the thickness of molecules bound to the sensor surface. Purified, recombinant, 80% truncated DENV-2 S1 E protein was obtained from Hawaii Biotechnology (Honolulu, HI). Peptides were N-terminally biotinylated with a 5∶1 molar ratio of NHS-LC-LC-Biotin (Pierce/ThermoFisher, Rockford, IL) in PBS pH 6.5 at 4°C. Excess biotinylation reagent was removed using Pepclean C-18 spin columns (Pierce/ThermoFisher, Rockford, IL). Biotinylated peptides were coupled to kinetics grade streptavidin high binding biosensors (Fortebio, Menlo Park, CA) at several different concentrations. Sensors coated with peptides were allowed to bind to E protein in PBS with 0.02% (v/v) Tween-20 and 1 mg/ml BSA at several different E protein concentrations. Binding kinetics were calculated using the Octet QK software package, which fit the observed binding curves to a 1∶1 binding model to calculate the association rate constants. E protein was allowed to dissociate by incubation of the sensors in PBS. Dissociation curves were fit to a 1∶1 model to calculate the dissociation rate constants. Binding affinities were calculated as the kinetic dissociation rate constant divided by the kinetic association rate constant. Approximately 200 FFU of DENV-2 without peptide was allowed to bind and enter target cells for 1 h at 37°C as described for the focus forming unit assay. Unbound virus was then removed by rinsing with PBS and peptide was added to the cells for 1 h at 37°C. Cultures were washed again in PBS and agarose overlays, incubation, and immunological detection was conducted as described for the focus forming unit assay. Approximately 200 FFU of DENV-2 were allowed to attach to cells for 45 min at 4°C, and then rinsed with cold PBS before peptide was incubated with the target cells for 45 min at 4°C. The cells were rinsed again with cold PBS, and agarose overlays, incubation, and immunological detection were conducted as described for the focus forming unit assay. Hemagglutination inhibition (HI) was performed according to [39] adapted to microtiter plates. Binding inhibition assays were modified from Thaisomboonsuk, et al [40].}. Briefly, LLC-MK2 monolayers were rinsed in 4°C DMEM containing 0.8% BSA and 25 mM HEPES, pH 7.5. Virus was incubated at 4°C with peptides, control anti-dengue serum, or heparan sulfate in DMEM/BSA/HEPES for one hour before adding to the monolayers for 2 hours at 4°C. Monolayers were rinsed 3 times with cold DMEM/BSA/HEPES media prior to RNA extraction using the Qiagen RNeasy mini kit (Valencia, CA) per manufacturers instructions. Quantitative, real time, reverse transcriptase polymerase chain reaction (qRT-PCR) was conducted utilizing the Roche Lightcycler RNA Master SYBR Green 1 qRT-PCR kit (Basel, Switzerland), using primers Den_F (TTAGAGGAGACCCCTCCC) and Den_R (TCTCCTCTAACCTCTAGTCC) from Chutinimitkul et al [41].}. and the following cycling conditions: 1 h at 61°C, 30 s at 95°C, followed by 45 cycles of: 5 s at 95°C, 20 s at 61°C, and 30 s at 72°C. Cp values were used to estimate infectious units according to a standard curve. Independent assays were repeated three times, in duplicate or triplicate. Graphs were generated using KaleidaGraph v.3.6 graphing software (Synergy Software, Reading, PA). Statistical analyses were performed using the GraphPad Prism 4.0 software package (GraphPad Software, San Diego, CA). P values less than 0.05 were considered significant. We had previously identified several E protein regions where peptides mimicking the E protein sequence might function as inhibitors. Several of these mimic peptides did not show substantial DENV inhibitory activity [9]. These included a peptide derived from the domain II fusion sequence (DN80, corresponding to amino acids 96–114 in the DENV-2 E protein) and two overlapping peptides derived from the domain II hinge region (DN57 and DN81, corresponding to amino acids 205–223 and 205–232, respectively). Predictions from crystal structures [11], [14], [15], as well as the previously confirmed inhibitory activity of an analogous WNV domain II hinge region peptide [9] lent support to the idea that the domain II hinge region was an attractive target for inhibition. Energy minimized peptides with sequences computationally optimized for structural stability and binding to the target regions, as evaluated by our residue-specific all-atom probability discriminatory function (RAPDF), were selected for further characterization and evaluation. These sequences generally had the best RAPDF scores (or “pseudoenergies”) for structural stability and binding, much better (lower) than the original wild type sequences (see Table 1 for original and optimized sequences.). These sequences, DN57opt, DN80opt and DN81opt, were selected for synthesis and evaluated for antiviral activity. To identify additional novel peptide inhibitors and their corresponding targets, a 20 residue sliding window that moved from the N to the C terminus of the DENV-2 strain S1 E protein (PDB ID 1OAN) in 10 residue acid increments was evaluated by a structural stability (pseudoenergy) optimization protocol using the RAPDF. A Metropolis Monte Carlo search algorithm [38] was used to change each amino acid in the 20 residue window to one of the 19 other naturally occurring amino acids, and the stability of each corresponding peptide in the context of the entire E protein structure was evaluated. Our approach identified four E protein regions with the potential for the highest in situ binding affinities. These correspond to DENV-2 strain S1 E protein amino acids 41–60, 131–150, 251–270, and 351–370 (see Figure 1) that were selected for synthesis and antiviral testing (1OAN1, 1OAN2, 1OAN3, and 1OAN4). In order to verify the effectiveness of the binding optimization process and peptide design, synthesized peptides were tested for antiviral activity against DENV-2 strain NG-C in a focus forming unit (FFU) reduction assay. DENV-2 strains S1 (GenBank accession number M19197.1) and NG-C (GenBank accession number AF038403.1) share 98% amino acid sequence identity in the E protein and the majority of differences are conservative. Dose response curves generated for the optimized peptides DN57opt, DN80opt, and DN81opt are shown in Figure 2A. The domain II region peptides, DN57opt and DN81opt displayed IC50 values of 8±1 µM and 36±6 µM (mean ± sem) respectively, while no inhibition of infection was observed with the fusion region peptide, DN80opt. Correspondingly, maximum inhibition of 97% and 57% was achieved at 20 µM and 50 µM for DN57opt and DN81opt. Both DN57opt and DN81opt showed improved inhibition of DENV-2 compared to their non-optimized counterparts, with DN57opt and DN81opt showing a nearly 14 fold and a 2 fold increase, respectively, in inhibition of DENV-2 at equivalent concentrations [9]. The most active inhibitor, DN57opt was chosen for further study. A scrambled version of DN57opt (DN57optscr) did not display inhibition at any concentration tested (Figure 2B). Four de novo designed peptides, 1OAN1, 1OAN2, 1OAN3, and 1OAN4 were also tested for inhibitory activity using the same FFU reduction assay (Figure 2C). 1OAN1 was found to be an effective inhibitor of DENV-2 infection with an IC50 of 7±4 µM and a maximum inhibition of 99% at 50 µM. A scrambled version of 1OAN1 (1OAN1scr) did not inhibit infection by DENV-2 at any concentration tested (Figure 2D). In addition to these full dose response inhibition experiments using approximately 100 infectious units of virus, both the DN57opt and 1OAN1 peptides were also capable of inhibiting 4,000 infectious units of virus (data not shown). Because toxicity could result in a decrease in focus formation and be interpreted as evidence of antiviral activity, the inhibitory peptides and their scrambled versions were assessed for cellular toxicity. Confluent monolayers of LLC-MK2 cells used in FFU reduction assays were exposed to increasing concentrations of peptide before measuring mitochondrial reductase activity using an MTT mitochondrial reductase activity assay (Figure 3). When we initially performed these assays to exactly mimic the focus forming unit assay by waiting five days after peptide exposure, we saw no evidence of toxicity at any concentration of any peptide (data not shown). However, we found that a shorter post-exposure incubation time revealed a subtle toxicity on the part of one of the peptides. Apparently, waiting more than 24 h post-exposure gives the cells a chance to recover and conceals this effect. At 24 h post-exposure, DN57opt was found to be mildly toxic to cells at 40 µM (one-way ANOVA with Dunnet's post hoc test, P = 0.0004, N = 18), so only inhibitory data using lower, nontoxic concentrations was considered. Peptides DN57optscr, 1OAN1, and 1OAN1scr were not toxic at any concentration tested (one-way ANOVA, P>0.05). Cryoelectron microscopy (cryoEM) was used to visualize the effect of the DN57opt and 1OAN1 peptides on DENV-2 viral particles. Control dengue virions exhibited the normal, nearly smooth outer surface typical of mature flaviviruses [42]. The surfaces of the virus particles werebecome followingrough after treatment with peptides, implying a possible rearrangement of the envelope glycoproteins (Figure 4). The treated virions no longer showed icosohedral symmetry, Attempts to reconstruct the structure of virus complexed with DN57opt and 1OAN1 structures by imposing icosahedral symmetry failed, indicating the viruses are no longer icosahedral. Control treatments with equivalent DMSO alone did not produce this morphological alteration. Biolayer interferometry was performed to examine binding of the peptides to purified, truncated DENV-2 E protein. Amino terminally biotinylated peptides were immobilized onto streptavidin biosensors and then the association and dissociation of truncated E protein with the immobilized peptides was monitored. The interactions of three different concentrations of truncated E protein to peptides DN57opt and 1OAN1 are shown (Figure 5). A buffer blank containing no E protein was run for each peptide. The affinities of the peptides for the truncated E protein were calculated with a 1∶1 binding model: DN57opt KD = 1.2×10−6±0.6×10−6 M (mean±sd), 1OAN1 KD = 4.5×10−7±2.0×10−7 M. While the data for the 1OAN1 peptide show a lower KD, these numbers are not statistically different (unpaired student's T-test, P = 0.16, N = 3). The association rate constants were: DN57opt ka = 8.0×102±5.0×102 M−1s−1, 1OAN1 ka = 3.9×103±1.5×103 M−1s−1. The dissociation rate constants were: DN57opt kd = 7.7×10−4±1.7×10−4 s−1, 1OAN1 kd = 1.6×10−3±0.2×10−3 s−1. We have previously used this system to characterize the binding affinities of several human monoclonal antibodies for DENV E proteins [43]. In order to determine if the peptides were exerting their effects on post-entry steps in the virus replication cycle, DENV-2 was allowed to infect LLC-MK2 cells before peptide was added to the cells (Figure 6). No inhibition of viral replication was observed at any concentration of DN57opt (Figure 6A) or 1OAN1 in these assays (Figure 6B), indicating that the peptides are not acting at a post-infection step. Since we had determined that inhibition with both peptides occurs at a viral entry step, we asked if infection could still be inhibited after virus had bound to the surface of target cells. We bound virus to cells at 4°C, then treated with increasing concentrations of DN57opt or 1OAN1 before warming the cells back to 37°C and allowing the infections to progress (Figure 6C and D). Inhibition of viral entry was observed for both peptides when added to the virus after it was bound to target cells. To determine if the peptides interefere with virus:cell interactions, we conducted two different experiments. We first performed hemagglutination inhibition assays, but were unable to detect any inhibition of the ability of viral antigen to agglutinate red blood cells (data not shown). To further investigate virus:cell binding in a more relevant system, we treated virus with DN57opt or 1OAN1, bound the virus to cells, and washed the cells repeatedly at 4°C before measuring the amount of virus remaining on the cells by quantitative rt-PCR. Both peptides showed evidence of ability to block virus:cell binding compared to control virus without peptide (Figure 7). Treatment of virus with pooled human anti-dengue serum or heparan sulfate similarly showed reduced cell binding. We have used computational methods to design multiple peptide inhibitors of the DENV E glycoprotein. Importantly, out of seven peptides synthesized and tested, two peptides with high activity and one peptide with intermediate activity were identified. A high resolution crystal structure of the pre-fusion conformation of the DENV-2 E [14] was used as the starting point to generate in situ energy minimized peptides. Two distinct approaches were used for the design of these peptides. First, we built upon previous work targeting DENV fusion peptide and domain II hinge regions with naturally occurring E protein sequences from these regions [9]. No inhibitory activity was found for the optimized fusion peptide region sequence (DN80opt), indicating that this region may not be a promising target mechanistically for DENV peptide inhibitors. Since an analogous, naturally occurring WNV domain II hinge region peptide was shown to be inhibitory against WNV [9], we reasoned that a more tightly binding analog of this region in the DENV E protein could be designed and might have improved inhibitory activity. This turned out to be correct, and we identified two distinct binding-optimized peptide sequences to this region with antiviral activity, DN57opt and DN81opt. This supports previous predictions of hinge region inhibitors and the proposed mechanism of fusion based on hinge region movements [11], [14], [15]. The second approach to designing peptide inhibitors was to identify peptides with non-native sequences derived from E protein regions that are highly stable in terms of structure and binding as evaluated by an all-atom scoring function (RAPDF). This identified four regions that were used to derive additional optimized peptides (Figure 1). Of the four resulting peptides tested, one, 1OAN1, was identified as having antiviral activity. This confirms the use of the sliding window RAPDF minimization approach for finding tightly binding protein ligands [23], [24]. It is perhaps not surprising that computational binding optimization increased the activity of previously inactive peptides that were based on naturally occurring E protein sequences. Naturally occurring sequences have multiple balancing selection pressures that may limit their binding stability in vivo. The combined use of primary sequence prediction tools [9] and structural optimization tools [23], [24] should be a valuable approach for finding binding partners and inhibitors for other protein targets. Neither peptide showed inhibitory activity when added directly to cells after infection had already occurred, indicating that the peptides were acting during an entry step in the virus life cycle, and sequence scrambled versions of the two most active peptides were inactive, confirming sequence specific activity. Both peptides also block virus:cell binding, but are still capable of inhibiting infection even when added after virions have already bound to the surface of target cells. CryoEM was used to visualize the effect of the peptides on DENV-2 virions. The surface of virions appeared to change from smooth to rough after incubation with the antiviral peptides. This suggests that there may be an alteration of the arrangement of the surface envelope protein (Figure 4). Biolayer interferometry was used to measure the kinetics of binding between the peptides and soluble, truncated E protein (Figure 5). These binding studies showed a direct interaction between the peptides and DENV-2 E protein with affinities in the 1 µM range and relatively fast on/off rates. The cryoEM images demonstrate that these inhibitory peptides probably cause structural deformations in intact viral particles, but do not provide information about the kinetics of these changes. It is possible that the peptides trap the viral E proteins in some conformation that is part of the normal breathing of the viral particles, and that this interferes with cell binding and entry. The DN57opt and 1OAN1 peptides were designed for optimized binding to the pre-fusion E structure and we show direct evidence for this interaction, both with the purified, monomeric E protein, and with virion particles. These peptides likely function by displacing portions of the E protein and interfering with normal cell binding or the structural changes during entry. Although separate in the primary protein sequence, the regions targeted in the design the DN57 and 1OAN1 peptides are partially adjacent to each other in the crystal structure, with the C terminus of the 1OAN1 region occupying a pocket surrounded by the DN57 region (See Figure 1). We stress that we do not know the structures of the bound and inhibited peptide/E protein complexes, but these structures may shed light on the mechanistic details of cell binding and fusion. Taken together, our results support the hypothesis that both of these peptides interact directly with DENV-2 E proteins and are entry inhibitors. Despite difficulties with oral administration and degradation in the digestive tract, peptides may make useful antiviral agents when targeted against viral envelope proteins. Directing inhibitors to viral surface proteins avoids the major difficulty of crossing cellular membranes in order to reach the target. For example, peptide inhibitors of intercellular viral targets, such as proteases or polymerases, would need to cross the cell plasma membrane, and in the case of flaviviruses, possibly internal membrane bound replication and assembly compartments. The HIV entry inhibitor T-20 (Fuzeon) is a peptide, and in the context of a chronic infection, repeated life-long injections are problematic. DENV is an acute infection and most severe DENV infections require intravenous fluid support, facilitating delivery of anti-DENV peptides by this route. We have established the existence of multiple, distinct inhibitory peptides targeting the DENV E glycoprotein and confirmed the utility of rational design using structural data for developing DENV E protein inhibitors. Applications of this strategy should also be possible for the generation and refinement of lead compounds for other viral envelope fusion proteins. It would be optimistic to propose that any single antiviral would provide an effective treatment for DENV given the enormous genetic variability of the four serotypes and multiple substrains. Different classes of inhibitors targeting the E protein and other DENV targets [44], [45], could form the basis for the development of a combination treatment plan to combat this disease.

10.1371/journal.pntd.0003034

Patterns and Risks of Trichinella Infection in Humans and Pigs in Northern Laos

Several outbreaks of trichinellosis associated with the consumption of raw pork have occurred in Laos since 2004. This cross-sectional study was conducted in four provinces of northern Laos to investigate the seroepidemiology of trichinellosis in the human population and determine the prevalence and species of Trichinella infection in the domestic pig population. Serum samples and questionnaire data were obtained from 1419 individuals. Serum samples were tested for Trichinella antibodies by ELISA using larval excretory–secretory (ES) antigens and a subset of 68 positive samples were tested by western blot. The seroprevalence of Trichinella antibodies was 19.1% (95% confidence interval (CI) = 17.1–21.1%). The risk of having antibodies detected by ELISA using ES antigens increased with age, being of Lao-Tai ethnicity, living in Oudomxay province and being male. Tongue and diaphragm muscle samples were collected from 728 pigs and tested for Trichinella larvae by the artificial digestion method. Trichinella larvae were isolated from 15 pigs (2.1%) of which 13 were identified as T. spiralis by molecular typing; the species of the two remaining isolates could not be determined due to DNA degradation. Trichinella spp. are endemic in the domestic environment of northern Laos and targeted preventative health measures should be initiated to reduce the risk of further outbreaks occurring.

Trichinellosis is one of the most widely distributed parasitic zoonoses worldwide and is caused by infection with nematodes of the genus Trichinella. Infection occurs after consuming larvae in the muscle of infected animals. Several outbreaks of trichinellosis have occurred in Laos since 2004, resulting in a substantial public health problem. The principal risk factor for trichinellosis is consumption of uncooked or partially cooked meat from domestic pigs and game. We visited communities in four ethnically diverse provinces of northern Laos to determine the seroprevalence of trichinellosis in the human population and explore the population and individual level risk factors of exposure. In addition, we also examined muscle samples collected from pigs post-slaughter to determine the prevalence of Trichinella infection and identify the species of Trichinella circulating in the domestic pig population. Our findings indicate that Trichinella spp. are endemic in the domestic environment of northern Laos and consumption of uncooked pork is common for all ethnic groups. Targeted preventative health measures, taking a one health approach by bringing together medical, veterinary and health sociology professionals, should be initiated to prevent the transmission of Trichinella and other foodborne pathogens in Laos.

Trichinellosis is one of the most widely distributed zoonoses worldwide and is caused by infection with nematodes of the genus Trichinella [1]. Infection occurs after consuming larvae in the muscle of infected animals with domestic and wild pigs the most common vehicles of human infections [2]. The severity of human disease is dependent on multiple factors including the number of viable larvae consumed, the frequency of consuming infected meat, meat being consumed raw or rare, the Trichinella species involved and individual susceptibility [3]. Trichinella spp. are endemic throughout Southeast Asia (SE Asia), from southern China to the Indonesian archipelago [4], [5] in domestic pigs and wildlife, causing frequent outbreaks of human disease. Three species of Trichinella have been detected in the SE Asian region, the encapsulated T. spiralis and the non-encapsulated T. pseudospiralis and T. papuae [6]. Trichinella spiralis has a regional distribution [7] with many of the recognised outbreaks occurring in the ethnically diverse regions of central and northern Laos, northern Thailand and northwest Vietnam where consumption of uncooked pork is common [8], [9], [10], [11]. Outbreaks of human trichinellosis involving T. pseudospiralis and T. papuae have occurred in Thailand after consuming wild pig meat [12], and cases of trichinellosis involving T. papuae have been detected in Papua New Guinea [15], [16] and a Thai patient returning from Malaysia [17]. Several outbreaks and sporadic cases of trichinellosis have occurred in Laos over the past five years [8], [18], [19] with the majority of the reported cases being associated with consumption of raw pork. Notwithstanding the propensity for Lao people to consume uncooked meat, including pork [20], little is known of the population and individual level risk factors of exposure and the meat consumption habits across an ethnically diverse country. Furthermore, relatively little is known about the prevalence of Trichinella infection in pigs and the species circulating in the domestic pig population. We report here the results of a cross-sectional serological survey of the human population and a concurrent survey in domestic pigs using muscle digestion in four provinces of northern Laos. Written informed consent was obtained from all participants 15 years and older and from the parents or legal guardians of children <15 years of age. The age of consent for this study was 15 years old. The study protocol was reviewed and approved by the Murdoch University Human Ethics Committee (Project no. 2008/266) and the Lao Ministry of Health National Ethics Committee for Health Research (no. 239/NECHR) before commencing this study. For the pig study, the protocol was reviewed and approved by the Murdoch University Animal Ethics Committee (Project no. R2108/07), which adheres to the Australian Code of Practice for the Care and Use of Animals for Scientific Purposes. Laos is an ethnically diverse country with 49 distinct ethnic groups classified into four ethno-linguistic families (Lao-Tai, Mon-Khmer, Hmong-Mien, and Sino- Tibetan), comprising 67%, 24%, 8%, and 1% of the population, respectively [21]. The study was conducted in four provinces in northern Laos (Oudomxay, Luangprabang, Huaphan, and Xiengkhuang), where all four ethnolinguistic families are represented. One district in each province (Xay, Xiengngeun, Viengxay, and Pek Districts, respectively) was randomly selected for inclusion in this study. The human survey was conducted in the dry season during January–March 2009 to maximize study participation and minimize negative impacts on seasonal labour demands. The survey design, sample size calculations and methodology have been described in detail elsewhere [20], [22]. The sample size calculations were based on estimates of taeniasis prevalence in the target populations. In brief, 14 households were randomly selected in each village and all household members ≥6 years of age were asked to participate. A venous blood sample of 2–3 mL was collected and the serum fraction was stored at −20°C. A household questionnaire was administered to the head of each household with his/her family present to assess the house characteristics, assets owned, ownership of animals, age of each household member, ethnicity and education levels, literacy of the male and female heads of household. Individual questionnaires were administered to collect data on meat consumption. For those family members who consumed raw meat, either pork, beef or fermented pork sausage, we asked them to estimate the frequency of raw meat consumption: weekly, monthly, every few months, and infrequently (once or twice per year or less often). Trichinella excretory secretory antigens (ESA) prepared from T. spiralis larvae [23] and four positive control human serum samples were provided by the European Union Reference Laboratory for Parasites (EURLP; Department of Infectious, Parasitic and Immunomediated Diseases, Istituto Superiore di Sanita, Rome, Italy) and supplied lyophilised and stored at 4°C on receipt in Laos. In brief, T. spiralis muscle larvae were harvested from 3 month old CD1 female mice weighing 25 g which had been infected with 500 muscle larvae each 40 days before, by HCl–pepsin digestion and then maintained in culture for 18 h. Five hundred thousand larvae were washed three times for 20 min each time by sedimentation in a sterile 50 ml conical tube with 45 ml of warm sterile phosphate buffered saline (PBS), pH 7.3, supplemented with Penicillin and Streptomycin (25,000 mg/mL and 25,000 U/mL, respectively). At each change of the washing solution, larvae were gently shaken to dislodge adherent bacteria. The washing solution was removed after the final sedimentation of larvae and washed an additional five times by sedimentation in a sterile 50 ml conical tube with 45 ml of warm RPMI 1640 media supplemented with Penicillin and Streptomycin, as above. The larvae were resuspended in warm RPMI 1640 media supplemented with 5,000 mg/mL of Penicillin, 5,000 U/mL of Streptomycin, 200 mM Glutamine and 100 mM sodium pyruvate at a concentration of 5,000 larvae/ml and placed in 25 ml tissue culture flasks. The flasks were incubated in 5% CO2 at 37°C for 16–18 h. The larvae were separated from the medium by sedimentation in 50 ml conical tubes. The medium was filtered through a 0.2 µm filter and the larvae were discarded. The filtered medium was then concentrated (100x) in a pressure concentrating chamber using a YM-3 filter at 4°C and clarified by washing with PBS in the same chamber or by dialysing in PBS for at least 4 h. The protein concentration was checked by spectrophotometer at 260 nm and 280 nm, and each batch was aliquoted and lyophilized at 500 µg of total proteins per vial. The ESA and control sera were reconstituted in analytical grade water, aliquoted and stored at −20°C according to the manufacturer's instructions immediately prior to use. Reconstituted ES antigens were further diluted in carbonate buffered saline (pH 9.6) to a working concentration of 5 mg/ml and the reconstituted positive control sera were further diluted 1/200 in blocking solution (0.5% bovine serum albumin, 0.05% Tween 20 in PBS) for use in the assay. A panel of eight negative control serum samples were sourced from Lao people with no reported history of trichinellosis or of consuming raw pork. Negative control serum and test serum were diluted 1/200 in blocking solution for use in the assay. The ES ELISA was performed in Laos at the National Centre for Laboratory and Epidemiology (NCLE) using a validated protocol [23], [24] with some minor modifications. Two positive control serum samples, 40 test serum samples and conjugate and substrate controls were added in duplicate to each plate; eight negative control serum samples were added to single wells of each plate. The optical density (OD) was measured at a wavelength of 450 nm using a microtiter plate reader (HumaReader, Germany). The cut-off on each plate was calculated as the mean OD of the eight negative control reference sera plus three standard deviations; a test ratio was calculated by dividing the OD of the test sample by the plate cut-off value and a test ratio ≥1 was considered reactive in the ES ELISA. Sixty-eight samples that had an ES ELISA test ratio ≥1 were randomly selected from the pool of positive samples and sent to the EURLP for confirmatory testing. Samples were tested by the ES ELISA and western blot according to methods described elsewhere [23]. The abattoir survey design has been described elsewhere [20]. In brief, pig surveys were conducted at three slaughter-points in Xiengkhuang and Oudomxay Provinces from May–September 2008 and at two collection points in Huaphan and Luangprabang Provinces from October 2008–January 2009. The survey team consisted of trained district and provincial agricultural and forestry government staff who visited the slaughter points approximately every two weeks. The tongue and diaphragm pillar muscles were excised from all pigs brought for slaughter on the nights the survey team visited. Muscle samples were collected into labelled plastic containers and stored at 4°C before transport on ice to the National Animal Health Laboratory in Vientiane where samples were stored at 4°C prior to artificial muscle digestion. Tongue and diaphragm muscle samples were artificially digested by the magnetic stirrer method in 1% pepsin (1∶10,000 US National Standard Formulary) and 1% hydrochloric acid (HCl) after removal of fat and fascia [25], [26]. Samples were tested in pools by muscle type with a maximum of 100 g per pool using 10 g of tissue per animal (if the animal was small >5 g of tissue was processed per animal). Tongue samples from positive pools were artificially digested as per the above protocol using 20 g muscle tissue (>10 g for small animals). Larvae were counted, transferred to 100% ethanol and sent to the EURLP for molecular species identification by multiplex PCR as previously described [27]. The prevalence of human serum reactivity with Trichinella ES antigens was calculated for three diagnostic cut-offs in the ES ELISA, test ratios ≥1, ≥1.2 and ≥1.4. The level of agreement between the ES ELISA results from Laos and EURLP, and the level of agreement between the ES ELISA results from Laos and the western blot test were calculated for the three diagnostic cut-offs using the Kappa statistic. Sensitivity and specificity could not be calculated since no ES ELISA negative samples from Laos were subjected to further testing at the EURLP. The questionnaire and laboratory test data were entered into a spreadsheet (Excel; Microsoft, USA) and subsequent analysis was carried out in STATA/IC version 10 (Stata Corp LP, USA). The socioeconomic status of each household was calculated by use of principal component analysis of household assets [28], [29] after replacement of missing values with the mean of the respective asset for that ethnic group. All assets were dichotomous. The households were ranked into wealth quintiles according to their cumulative standardized asset scores. Univariate logistic regression without adjustment was used to test associations between ES ELISA reactivity and gender, location, ethnicity, age, wealth status and uncooked meat consumption habits. Risk factors significant or borderline significant (P≤0.20) in the univariate analyses were included in a multivariate random effects logistic regression model adjusting for the effect of household clustering. The results are reported as adjusted odds ratios and 95% confidence intervals (CIs). The final analysis only considered persons with serologic and questionnaire data. In the pig study, the Pearson's chi-square test was used to explore associations between infection status (larvae detected by artificial digestion) and age, breed, sex and production system at last point of sale. A total of 1,582 persons in 332 households were eligible to participate in this survey. Of these persons, 1,419 (89.7%) individuals from 324 households aged 6–91 years provided a blood sample, a completed questionnaire, and had valid laboratory test results. The final survey population consisted of 583 Lao-Tai (93.6% compliance), 564 Mon-Khmer (95.4% compliance), and 272 Hmong-Mien (73.4% compliance). No Sino-Tibetan persons were recruited into this study. Survey population structures stratified by province are shown in Table 1. Significant differences in the survey population structure were observed for ethnicity and wealth status. Lao-Tai people made up the majority of the population surveyed in Huaphan province (95.6%), Mon-Khmer people made of the majority of the survey population in Oudomxay and Lauangprabang provinces (78.6% and 64.3%, respectively) and Hmong-Mien people made up the majority of the survey population in Xiengkhuang province (58.5%) (Table 1). Oudomxay province had the greatest proportion of participants who were very poor or most poor (68.5%) and this was reflected in the finding that Mon-Khmer people were the most impoverished ethnic group and Lao-Tai people were the least poor overall. Sixty-eight samples with a test ratio ≥1 in the ES ELISA in Laos were tested at the EURLP by the ES ELISA method and all but one were confirmed positive, corresponding to 98.5% agreement. In comparison with the western blot test, only 35.3% (24/68) of the Lao samples had three diagnostic bands detected, a banding profile consistent with clinically confirmed trichinellosis (Figure S1) [23], [24]. Using a diagnostic cut-off test ratio ≥1.0, ≥1.2 and ≥1.4 for the ES ELISA, the level of agreement with the western blot test was 35.3%, 50.0% and 62.3%, respectively. The two-by-two tables comparing the western blot test results and ES ELISA at different diagnostic cut offs are presented in Table 2. Using a diagnostic cut-off test ratio ≥1.0, ≥1.2 and ≥1.4, the prevalence of Trichinella antibodies detected by ES ELISA were 19.1%, 12.7% and 7.5%, respectively (Table 3). The prevalence of antibody detection by ES ELISA was highest in males, increased with increasing age to a peak in 35–49 year olds, increased with increasing wealth and was highest in the Lao-Tai ethnic group (Table 3). Prevalence was highest in Oudomxay province when a cut-off test ratio ≥1.2 and ≥1.40 were applied, and was highest in Xienghuang province when a cut-off test ratio ≥1.0 was applied (Table 3). The proportion of people reporting the consumption of uncooked beef, pork and fermented sausage peaked in older age groups for all ethnic groups (Figure 1) with the exception of Hmong-Mien people consuming fermented pork sausage, which was comparatively low for all age groups. The prevalence of antibody detection using a cut-off test ratio ≥1.0 was highest in people reporting consumption of raw pork (26.2% verses 18.0%), raw beef (27.1% verses 14.7%) and fermented pork sausage (29.8% verses 15.0%). Similarly, the prevalence of antibody detection using a cut-off test ratio ≥1.2 was highest in people reporting consumption of raw pork (18.5% verses 11.8%), raw beef (19.3% verses 9.1%) and fermented pork sausage (21.7% verses 9.2%). Using a cut-off test ratio ≥1.4, prevalence was highest in people reporting consumption of raw pork (10.8% verses 7.0%), raw beef (12.3% verses 4.6%) and fermented pork sausage (13.3% verses 5.2%). The prevalence of antibody detection stratified by frequency of raw meat consumption are summarised in Table 3. After controlling for clustering at the household level, the risk of having Trichinella antibodies detected in the ES ELISA was significantly greater for people residing in Oudomxay province, people of Lao-Tai ethnicity, increasing age and being male. Increasing wealth was no longer associated with increased risk of having Trichinella antibodies detected after controlling for other risk factors (Figure 2). The frequency of consuming raw pork and beef were not associated with increased risk of having antibodies detected. Only consumption of fermented pork sausage on a weekly basis was significantly associated with increased risk of having antibodies detected when the diagnostic cut-off test ratio ≥1.2 (Odds ratio = 3.29 (95% CI = 1.35–7.99); Figure 2). Tongue and diaphragm muscle samples were tested by the artificial digestion method from 728 pigs sampled from all four northern provinces included in the study. Trichinella larvae were isolated from 15 pigs (2.1%) of which 13 were identified as T. spiralis by molecular typing. Two isolates were not identified to the species level due to damaged DNA that may have occurred during tissue storage and muscle digestion. Prevalence of Trichinella spp. infection in pigs varied significantly (P<0.05) by province whereby the highest prevalence was recorded in Xiengkhuang (4.8%) and Oudomxay (2.8%) provinces (Table 4). No samples collected in Luangprabang province were infected with Trichinella larvae at the time of this survey. There was no significant difference in prevalence by breed, sex or the production system where the pigs were purchased immediately prior to slaughter. Of the 15 pigs infected with Trichinella larvae, 0.1–0.9 larvae per gram (lpg) of tongue tissue was detected in 10 pigs, 1–10 lpg was detected in three pigs and greater than 10 lpg was detected in two pigs. The highest recorded intensity of infection was 69 lpg in a pig slaughtered in Xiengkhuang province. Trichinellosis is endemic in Southeast Asia with a concentration of outbreaks occurring in the ethnically diverse regions of northern Thailand, northern Vietnam and Laos [8], [9], [10], [11]. Our study confirms endemicity of T. spiralis in the pig population of Laos together with a spatial difference in prevalence of T. spiralis infection in pigs with worm burdens sufficient to cause severe human disease. One of the principle aims of the present study was to determine population and individual level risk factors associated with human exposure to Trichinella spp. larvae. Therefore we conducted a randomised cross-sectional survey of the human population in four northern provinces of Laos and used the ES ELISA as a serological measure of exposure. A high prevalence of Trichinella antibodies was detected by ES ELISA, with significant increased risk being associated with increasing age, Lao-Tai ethnicity, residing in Oudomxay province, being male and regular consumption of fermented pork sausage. An important limitation of this study was the low participation rate of people from the Hmong-Mien ethnic group. The reasons for this low participation rate have been discussed elsewhere [20], [22]. Overall, the prevalence of ES ELISA reactivity across all ethnic groups increased with increasing age and prevalence was highest for males. In the Hmong-Mien ethnic group the ratio of females to males was similar for all age groups except the youngest group, where boys represented 60% of the age group. This discrepancy indicates that older males were over represented in the survey and our prevalence estimates are possibly higher than would otherwise have been the case if participation rates were higher. In addition, the highest non-participation rates in the Hmong-Mien group were observed in Huaphan province and this may have led to an over-estimation of prevalence of ES ELISA reactivity in this province. No diagnostic test for trichinellosis, in any host species, has been validated for cross-sectional studies in Southeast Asia. The Trichinella ES ELISA lacks specificity owing to the relatively large population of antigens resulting in the detection of non-specific cross-reacting antibodies [23], [24]. The western blot test described by Gomez-Morales et al. [23], [24] was used as the gold-standard comparator for a small subset of ES ELISA positive samples in this study. The full spectrum of exposures, from subclinical infection, exposure to inactivated or injured larvae, old exposures through to acute and chronic clinical disease would lead to a varied serological spectrum at a population level. The ES ELISA results presented here are therefore imperfect but provide a measure of exposure at the population level. Despite these limitations, we were able to demonstrate widespread serological evidence of exposure to Trichinella larvae in the human population. Increasing the diagnostic cut-off in the ES ELISA resulted in improved agreement with the western blot test, likely as a consequence of improved specificity at the expense of sensitivity. For this reason, prevalence was calculated for a range of diagnostic cut-offs in the ES ELISA to test the effect on the subsequent risk factor analysis. The pattern of risk for age, province of residence, gender, ethnicity and wealth status remained essentially the same as the diagnostic cut-off increased. For raw meat consumption, only the self-reported consumption of fermented pork sausage on a weekly basis was significantly associated with antibody detection in the ES ELISA at a cut-off test ratio ≥1.2. The consumption of fermented pork has previously been linked with an outbreak of trichinellosis in Bolikhamxay province in central Laos [18]. The lack of association with raw meat consumption, particularly raw pork, was somewhat unexpected since previously reported outbreaks of trichinellosis in Oudomxay and Bolikhamxay provinces have been linked with consumption of raw pork at festivals [8], [18]. This might be explained by the limitations of using single point-in-time self-reporting as opposed to asking the survey participants to keep a more detailed food diary. In general, methods of assessing dietary intake are imperfect and subject to error [30], especially in an ethnically diverse population with high rates of illiteracy and where Lao may have been the second language. From the data collected we were unable to estimate or correct for recall bias and the possibility that some survey participants may have misinterpreted the questionnaire cannot be ruled out. Future studies assessing risk associated with consuming uncooked pork should consider the use of a food diary to better estimate the prevalence of consuming uncooked meat. In all risk models, the risk of having Trichinella antibodies detected by ES ELISA was highest in Oudomxay province compared to all other provinces. This finding may be an artefact of the large and widespread outbreak of trichinellosis that occurred in this province in 2005 [8] and be indicative of more widespread exposure over and above the clinical cases that were reported. More research, using more statistically powered surveys, is warranted in Oudomxay province to further investigate the risk of trichinellosis in this province. The majority of pigs in which T. spiralis larvae were detected had a worm burden of less than 1 lpg and the infecting dose for clinically apparent trichinellosis has been estimated to range from ∼70-150 larvae [3]. The serological results together with the meat consumption habits and the abattoir survey results suggests that subclinical exposure may be common in Laos. Barennes and others (2008) reported apparently low morbidity associated with the 2005 outbreak in Oudomxay province and hypothesised that alcohol consumption may have diminished the severity of disease. Our results indicate that population immunity may have had a protective role. The risk of Trichinella antibody detection in the ES ELISA increased significantly with increasing age and Lao-Tai people were at significantly greater risk. Regionally, trichinellosis has been associated with ethnically diverse mountainous areas of northern Vietnam and Thailand [9], [10], whereas in Laos we found the greatest risk associated with the majority lowland Lao-Tai population and the lowest risk was associated with people from the minority upland Mon-Khmer and Hmong-Mien ethnic groups. A high proportion of people from all ethnic groups reported consuming uncooked meat. Public health interventions, including a detailed assessment of the risks posed by ceremonial food preparation and the development of food safety education and awareness programs, could potentially reduce the transmission of Trichinella and other foodborne pathogens in Laos.

10.1371/journal.pbio.1001477

Strength of Gamma Rhythm Depends on Normalization

Neuronal assemblies often exhibit stimulus-induced rhythmic activity in the gamma range (30–80 Hz), whose magnitude depends on the attentional load. This has led to the suggestion that gamma rhythms form dynamic communication channels across cortical areas processing the features of behaviorally relevant stimuli. Recently, attention has been linked to a normalization mechanism, in which the response of a neuron is suppressed (normalized) by the overall activity of a large pool of neighboring neurons. In this model, attention increases the excitatory drive received by the neuron, which in turn also increases the strength of normalization, thereby changing the balance of excitation and inhibition. Recent studies have shown that gamma power also depends on such excitatory–inhibitory interactions. Could modulation in gamma power during an attention task be a reflection of the changes in the underlying excitation–inhibition interactions? By manipulating the normalization strength independent of attentional load in macaque monkeys, we show that gamma power increases with increasing normalization, even when the attentional load is fixed. Further, manipulations of attention that increase normalization increase gamma power, even when they decrease the firing rate. Thus, gamma rhythms could be a reflection of changes in the relative strengths of excitation and normalization rather than playing a functional role in communication or control.

Brain signals often show a stimulus-induced rhythm in the “gamma” band (30–80 Hz) whose magnitude depends on attentional load, leading to suggestions that gamma rhythm plays a functional role in routing signals across cortical areas. However, gamma power also depends on simple stimulus features such as size or contrast, which suggests that gamma could arise from basic cortical processes involving excitation–inhibition interactions. One such process is divisive normalization, a mechanism that suppresses the response of a neuron by the overall activity of a large pool of neighboring neurons. Recent studies have shown that attention increases the strength of both excitation and normalization. We hypothesized that the increase in gamma power in an attention task is due to the effect of attention on excitation and normalization. By manipulating the normalization strength independent of attentional load in macaque monkeys, we show that gamma power increases with increasing normalization, even when attentional load is held fixed. Thus, gamma rhythms could be a reflection of changes in the relative strengths of excitation and normalization rather than playing a functional role in communication or control.

Modulations in gamma rhythms have consistently been observed during high-level cognitive processes such as attention [1]–[5], memory [6], feature-binding [7],[8], or conscious perception [9], leading to the suggestion that these rhythms play a functional role in high-level cognitive processing [7],[10]. However, several studies have shown that the magnitude and center frequency of the gamma rhythm depend on stimulus features such as contrast [11]–[13], orientation [14],[15], size [15],[16], and direction [12],[17], irrespective of the cognitive state, suggesting that gamma rhythms could be a reflection of basic cortical processes such as the interaction between excitation and inhibition [18]. Recent studies have suggested that selective attention, a high-level cognitive function often associated with gamma rhythms [1]–[5], is mediated through a sensory mechanism called normalization [19],[20]. Normalization is a form of gain control in which neuronal responses are reduced in proportion to the activity of a large pool of neighboring neurons [21],[22]. In the normalization model of attention, attention increases the excitatory drive to a neuron processing the attended stimulus. However, the increased excitatory drive also increases the strength of the normalization pool. The relative increase in the strength of normalization compared to excitation depends on several factors, such as the stimulus size and the focus of attention [20],[23], as well as tuning properties of the normalization pool [24], and these factors determine the overall effect of attention on the firing rate of the neuron. The normalization model of attention, as well as other models (see Discussion), therefore predict that attention changes the relative strengths of excitation and inhibition. We hypothesized that the changes in gamma power observed with attention reflect the effect of attention on the underlying excitation and normalization strengths. In particular, we hypothesized that gamma power should increase with increasing normalization, even if attentional load is held fixed. We tested this hypothesis by recording single units and local field potentials (LFPs) from the middle temporal area (MT) of two macaque monkeys while they performed a task in which normalization and spatial attention were varied independently, and studying the effects of these manipulations on gamma power. To manipulate the strength of normalization, we cued the monkeys to attend to a stimulus outside the receptive field of an MT neuron while presenting two stimuli inside the receptive field—one moving in the cell's preferred direction and the second in the opposite (null) direction (“Normalization Protocol,” Figure 1A). The addition of a null stimulus, which by itself produces little excitation, decreases the response produced by the preferred stimulus alone, a phenomenon that has been explained using normalization [21],[22]. The addition of a null stimulus does not appreciably increase the excitatory drive received by the recorded neuron, but it increases the normalization strength considerably because other neurons in the normalization pool have different direction selectivities and therefore some neurons in the pool respond to the null stimulus also. Therefore, addition of a null stimulus increases normalization strength without any appreciable increase in excitation, and consequently decreases the firing rate. We manipulated normalization by varying the contrasts of the preferred and null stimuli inside the receptive field (each could take one of three contrasts: 0%, 50%, or 100%) while keeping the animal's attention directed away from the receptive field. We label each condition as PxNy, where x and y are the contrasts of the preferred and null stimuli. The stimuli were presented rapidly (200 ms) with a short interstimulus interval (158–293 ms; Figure 1C), which made it unlikely that the animals could adjust their attention in response to the variable contrast of stimuli within the duration of the presentations. Figure 2A shows the average time-frequency power (on a log scale) of 96 recording sites in the area MT of two monkeys (55 from Monkey 1 and 41 from Monkey 2; results were similar and individually significant for the two monkeys and hence the data were pooled) for the P100N0 condition (a single stimulus at 100% contrast moving in the preferred direction). Time-frequency analysis was done using the Matching Pursuit algorithm, which provided sufficient resolution to resolve any oscillatory activity related to normalization/attention as well as transient activity due to fast stimulus presentation rates (see Materials and Methods for details). Line noise and monitor refresh rate caused a sustained increase in power in the LFP, visible as two narrow horizontal lines at 60 and 75 Hz in Figure 2A. In addition, there was a prominent increase in power between 65 and 80 Hz starting around ∼100 ms after stimulus onset. Figure 2B shows the power spectrum (on a log scale) of the LFP, obtained by averaging the time-frequency power between 50 and 250 ms (red trace). For comparison, we also include the power spectrum when no stimulus was presented (P0N0 condition; orange trace) and the “baseline” spectrum obtained by averaging the power between 100 and 0 ms before stimulus onset for all nine normalization conditions (black trace). The baseline spectrum had slightly more power than the P0N0 spectrum (black curve is slightly above orange), which was expected because the baseline period contained some residual activity from the previous stimulus. The localized increase in gamma power between 65 and 80 Hz was reflected as a “bump” in the P100N0 spectrum, which was missing in both baseline and P0N0 spectra. The gamma band increase observed between 65 and 80 Hz is not an artifact of the monitor refresh. Because the monitor refresh occurs at a fixed frequency, phase-locking of neurons to the monitor refresh rate is typically limited to a very narrow frequency band around the refresh rate, and in particular there is no evidence in the literature of such artifacts spreading to a broad frequency band. Further, even if the activity related to the monitor refresh rate varied with time (because the stimulus changed with time), it would cause an amplitude modulation of the 75 Hz sinusoid. The Fourier Transform of an amplitude modulated sinusoid is equal to the convolution of the Fourier Transform of the sinusoid (which produces a delta function at 75 Hz) and the Fourier Transform of the amplitude modulation. This is simply the Fourier Transform of the amplitude modulation centered at 75 Hz. Irrespective of the type of amplitude modulation introduced by the time-varying stimulus, the spread should be symmetric around 75 Hz, which was not the case. For the P100N0 condition, the artifact related to monitor refresh rate was visible as a narrow peak at 75 Hz that was distinct from the gamma band increase (the spectrum for the P100N0 condition around 75 Hz is enlarged in the inset). Further, gamma modulation was observed for the attention condition even when the stimulus conditions were identical (see below), which rules out the monitor refresh rate–related noise as the sole source of gamma power. Although the use of Matching Pursuit resolved the line and monitor-related noise from ongoing oscillatory activity in the gamma band at high resolution, the results obtained using a traditional multitaper method [25],[26] were comparable and showed a prominent increase in power in the gamma range (Figure S1). Figure 3A shows the average firing rates when a stimulus moving in the neuron's preferred direction was presented at 0% (left), 50% (middle), and 100% (right) contrast, together with a null stimulus at 0% (red traces; lower preferred stimulus contrast is shown in a lighter shade), 50% (green), and 100% (blue) contrast. As expected from normalization, addition of a null stimulus decreased the firing rates. Figure 3B shows the change in LFP power relative to a common baseline period (Figure 2B, black trace) for different pairings of preferred (different columns) and null contrasts (different rows). Gamma rhythm was observed between 65 and 80 Hz, and its strength increased when a null stimulus was added (first versus second/third row). This increase was specific to the gamma band—for example, power did not increase in the high-gamma band (>80 Hz) with increasing normalization (Figure 3B, also see Figure 4B for comparison as a function of frequency). To study these effects in more detail, we plotted the power between 50 and 250 ms as a function of frequency (Figure 4A) as well as the gamma power (between 65 and 80 Hz; excluding 74–76 Hz) as a function of time (Figure 4C), for all nine normalization conditions. Figure 4B and 4D show the change in power (in dB) between the P100N100 and P100N0 conditions as a function of frequency and time, respectively. In Figure 4B, the change was significant only in the gamma range and at very low frequencies (which was due to differences in transient activity; see Figure 3B). The change in gamma power started ∼50 ms after stimulus onset and persisted throughout the duration of the stimulus (Figure 4D). To quantify the effect of normalization, we computed the total power in the gamma range (65–80 Hz, excluding 74–76 Hz; the analysis window is indicated by a black box in the panels of Figure 3B) and high-gamma range (80–135 Hz), for each normalization condition. Figure 5A shows the mean change in gamma power for different stimulus conditions relative to the P100N0 condition. Neurons in area MT typically have a low semi-saturation constant (σ in Text S1) and tend to saturate even for contrasts much less than 100% [27], so the results were similar for stimuli at 50% and 100% contrast (gamma power was not significantly different between P50N0 and P100N0 conditions; difference: 1.7%±2.0%, p = 0.39, N = 96, t test). However, gamma power increased significantly when a null stimulus at 50% or 100% contrast was added to a preferred stimulus at 50% or 100% contrast: relative changes in gamma power from P100N0 condition for P50N50, P50N100, P100N50, and P100N100 conditions were 11.1%±2.8%, 11.3%±3.0%, 19.6%±2.8%, and 18.8%±3.1%, respectively (p = 1.6×10−4, p = 2.9×10−4, p = 2.9×10−10 and p = 3.2×10−8, N = 96, t test). When analyzed separately for the two monkeys, the corresponding values were 10.5%±3.4%, 12.6%±4.3%, 25.6%±3.8%, and 27.3%±4.5% for Monkey 1 (p = 3.7×10−3, p = 4.6×10−3, p = 1.3×10−8, and p = 1.0×10−7, N = 55, t test) and 11.8%±4.7%, 9.5%±4.1%, 11.5%±3.7%, and 7.3%±3.5% for Monkey 2 (p = 0.02, p = 0.03, p = 0.003, and p = 0.04, N = 41, t test). On the other hand, the increases in high-gamma power (Figure 5B) for corresponding conditions were −0.4%±1.3%, −1.5%±1.4%, 3.4%±1.4%, and 2.5%±1.5%, respectively (p = 0.76, p = 0.3, p = 0.02, and p = 0.09, N = 96, t test). Thus, addition of a second stimulus inside the receptive field of a neuron, which increased normalization, increased the magnitude of the gamma rhythm even when attention was fixed outside the receptive field. However, increasing normalization had negligible effect at high-gamma frequencies. Similar results were obtained using the multitaper method. Relative changes in gamma power (sum of power at 65, 70, and 80 Hz) from P100N0 condition for P50N50, P50N100, P100N50, and P100N100 conditions were 5.3%±2.4%, 8.0%±3.0%, 13.0%±2.7%, and 16.2%±3.4%, respectively (p = 0.03, p = 0.009, p = 5.2×10−6 and p = 7.3×10−6, N = 96, t test). For high-gamma power, the corresponding values were −0.2%±1.4%, −2.4%±1.3%, 2.2%±1.5%, and 1.0%±1.5%, respectively (p = 0.87, p = 0.07, p = 0.14, and p = 0.50, N = 96, t test). Interestingly, while normalization is generally thought to be largely un-tuned for orientation [21],[22], the gamma rhythm was much stronger when a preferred stimulus was presented instead of a null stimulus (compare P0N100 versus P100N0 in Figures 3B; both should involve the same normalization signal). This suggests that the gamma rhythm depends not only on the suppressive normalization signal, but on the incoming excitatory drive as well, and could be a resonant phenomenon arising from the excitation–inhibition interaction [13],[18],[28],[29]. However, differences in the levels of excitation alone across stimulus conditions cannot explain these results, because changes in excitation modulate power in a broad frequency band including the high-gamma band (see Discussion for more details). Next, we studied the effect of shifting the focus of attention under identical stimulus conditions (Figure 1B, “Spatial Attention Protocol”). Figure 6A shows the average firing rates of the 96 neurons when two stimuli at 100% contrast moving in the preferred and null directions were presented inside the receptive field, while the animal focused on a stimulus outside the receptive field (P100N100; dark blue trace) or on the null (P100N100Att; magenta) or preferred (P100AttN100; violet) stimulus inside the receptive field. This attentional manipulation allowed us to dissociate the dependence of gamma power on normalization versus firing rate modulations. This is because the response of the neuron shifted toward the response elicited when the attended stimulus was presented alone, and therefore decreased when attending to null (P100N100Att) and increased when attending to preferred (P100AttN100) compared to the P100N100 condition [30],[31]. In contrast, the strength of normalization increased for both P100N100Att and P100AttN100 conditions (compared to the P100N100 condition) because attention was directed to a stimulus inside the receptive field instead of outside. This was indeed reflected in the gamma power, whose strength increased when attention was directed inside the receptive field for both the P100N100Att and P100AttN100 conditions (Figure 6B; compare first versus second/third row). Figure 6C shows the normalized firing rate (Firing), gamma power (γ), and high-gamma power (Hi-γ) for the P100N100, P100N100Att, and P100AttN100 conditions (normalized with respect to P100N0 as before). The firing rate decreased by 28.6%±1.8% (dark blue bar) when a null stimulus was added to the receptive field and decreased by 37.1%±2.3% when attention was directed to that null stimulus (magenta bar). Attention to the preferred stimulus largely counteracted the presence of the null stimulus, leaving a decrease of only 3.3%±2.6% from the preferred only stimulus (violet bar). On the other hand, gamma power increased by 18.8%±3.1% when the null stimulus was added, 33.6%±4.8% when this null stimulus was attended, and 40.1%±4.3% when the preferred stimulus was attended (all changes compared to the P100N0 condition). The increase of 12.9% in the gamma power from P100N100 to P100N100Att was highly significant (p = 3.5×10−5, N = 96, t test). When analyzed separately, the increase was 9.0% (p = 0.0017, N = 55, t test) for Monkey 1 and 18.2% (p = 0.005, N = 41, t test) for Monkey 2. The increase from P100N100Att to P100AttN100 was 8.1% for the pooled data (p = 0.02, N = 96, t test), 4.4% for Monkey 1 (p = 0.35, N = 55, t test), and 13.3% for Monkey 2 (p = 0.04, N = 41, t test). Thus, manipulations of attention that increased normalization increased gamma power even when they decreased the firing rate, suggesting that the effects of attention on gamma power may be an indirect consequence of its direct effect on normalization. Unlike manipulations of normalization, manipulations of attention changed the power at non-gamma frequencies also. For example, power in the high-gamma range increased by 2.5%±1.5% when the null stimulus was added, 9.6%±3.1% when this null stimulus was attended, and 15.0%±1.8% when the preferred stimulus was attended (Figure 6C, “Hi-γ”). The increases of 6.9% from P100N100 to P100N100Att and 4.9% from P100N100Att to P100AttN100 were both significant (p = 0.03 and p = 0.02, N = 96, t test). To study the effect of attention at different frequencies in more detail, we plotted the power between 50 and 250 ms as a function of frequency (Figure 6D; left column) and the gamma power as a function of time (Figure 6D, right column) for different attention conditions. The top row shows the raw power, while the middle and bottom rows show the change in power for the P100N100Att versus P100N100 condition and P100AttN100 versus P100N100 conditions, respectively. Attention increased the power in a broad frequency band above 50 Hz and decreased power below 30 Hz (left column, middle and bottom rows). As a function of time, gamma power was elevated throughout the duration of the trial irrespective of stimulus onset for the P100N100Att versus P100N100 condition (middle row, right column), but showed a larger increase after stimulus onset for the P100AttN100 versus P100N100 condition (bottom row, right column). Results obtained from multitaper analysis were very similar (not shown). We observed a pronounced suppression at low frequencies (<30 Hz) with attention, as shown in Figure 6B and 6D. To study the effects of normalization and attention at low frequencies, we plotted the change in power from baseline for different normalization and attention conditions (Figure 7A). From the time-frequency difference plots (Figures 3B and 6B), two prominent features were observed at low frequencies. First, we observed an increase in power at ∼10 Hz at ∼100 ms, probably reflecting the stimulus-induced transient. Second, we observed a pronounced suppression in power between 20 and 30 Hz. Figure 7B shows the change in power (from the P100N0 condition as before) in the alpha (8–12 Hz; left panel) and beta2 (20–30 Hz; right) bands for different normalization and attention conditions. For the Normalization conditions (from P0N0 through P100N100), alpha power increased with the strength of normalization, probably because the stimulus-induced transient reflected the overall population activity that increased with increasing normalization (Figure 3B). The beta2 band did not show any significant modulation with normalization (Figure 7B, right panel). This can also be seen in Figure 3B, where the blue patches reflecting the beta2 decrease have approximately the same intensity. Even though this patch appears missing in the P0N0 condition, it is only because power at other frequencies changes by a similar proportion—that is, other frequencies also have a similar shade of blue, so the color contrast is not salient (compare the orange trace in Figure 7A that has no dip in the beta2 range with other traces that show a prominent dip). On the other hand, attention decreased the power in both alpha and beta2 ranges (Figures 6B and 7), consistent with a large number of prior studies [5],[12],[32],[33]. Finally, we studied whether the increase in gamma power due to attention can be explained through normalization on a neuron-by-neuron basis. Neurons in area MT have a variable change in firing rate when a null stimulus is added to a preferred stimulus in their receptive field—for some neurons, the firing rate decreases substantially, while for others there is hardly any decrease, which can be explained by the variability in the strength of the normalization (the tuned normalization model is summarized in Text S1) [24]. The strength of normalization can be approximated as α = (firing rate(P100N0)/firing rate(P100N100))−1 (Text S1). Previous studies have shown that α is strongly correlated with the overall attentional modulation in firing rates [measured as (P100AttN100−P100N100Att)/(P100AttN100+P100N100Att)] [19],[24]. We therefore studied whether α can also predict the attentional modulation in gamma power. Figure 8A plots the relationship between the increase in gamma power (measured in dB) when attention was directed to the preferred stimulus versus outside (P100AttN100 versus P100N100), as a function of the normalization strength (α). Neurons demonstrating a stronger normalization signal (α) should show a greater attentional modulation in gamma power. However, these two parameters were not correlated (ρ = 0.01, p = 0.9, Spearman Rank test). This is because gamma power depends not only on the strength of normalization but also on the strength of the incoming excitation, and attention increases both these quantities. This issue can be partially resolved by studying the correlation between α and the increase in gamma power when attention was directed to the null stimulus (Figure 8B), because in this case attention increases the strength of normalization but does not substantially increase the strength of incoming excitation (because the null stimulus produces almost no response in neurons in area MT). In this case, the increase in gamma power was weakly but significantly correlated with α (ρ = 0.3, p = 0.003, N = 96, Spearman Rank test), although the correlation did not reach significance for Monkey 1 when the analysis was done separately for each monkey (Monkey 1: ρ = 0.21, p = 0.13, N = 55; Monkey 2: ρ = 0.37, p = 0.02, N = 41, Spearman Rank test). Thus, changes in firing rates from a pure manipulation of normalization (which were used to estimate α) were a weak but significant predictor of the changes in gamma power during a manipulation of attention, but only when attention modulated the normalization strength alone. Differences between the effects of normalization and attention on the power spectrum are addressed in more detail in the Discussion. This study integrates a number of other results to directly link normalization strength and gamma power and provides an alternate explanation for the increase in gamma typically observed in higher cortical areas due to attention. Prior studies have shown that gamma power is modulated by incoming excitation and inhibition and could be a resonant phenomenon arising from their interaction [13],[16],[18],[28],[29],[34]. Some models of normalization are based on such excitation–inhibition interactions [21],[22], although other models of normalization may operate without inhibition, as described below. Finally, previous studies have shown that effects of attention and normalization on a particular neuron are tightly correlated [18],[28], suggesting that attention could change the strengths of excitation and normalization [19],[20]. The present study integrates these results—we first link gamma power to normalization strength while keeping attention constant, and then use an attention paradigm to show that the increase in gamma power due to attention could be explained at least in part by the effect of attention on normalization strength. Early models of attention such as the biased competition model [35]–[37] suggested that when multiple stimuli are presented inside the receptive field of a neuron, they activate different neural assemblies that compete for high-level representation, and attention biases the competition in favor of the attended stimulus. These models, however, fail to explain the effect of attention on neural responses when a single stimulus is present inside the receptive field [38]. Other types of models such as the flexible input gain model [23],[39] operate by changing the relative weights of inputs into a neuron, without changing the rules by which these inputs are integrated together. The input gain model can explain the increase in firing rates observed when a single stimulus is presented, as well as the competitive behavior when multiple stimuli are presented [23],[39]. In this model, the response of a neuron when a preferred and a null stimulus are both presented is given by RP,N = λ((βPRP)n+(βNRN)n)1/n, where RP and RN are the responses when the preferred and null stimuli are presented alone, βP and βN are the attentional gains applied to each input, n incorporates nonlinear summation (n = 1 for linear; n = infinity for winner-take-all), while λ is a scaling term. However, input gain or biased competition models cannot easily explain the decrease in firing rates when a null stimulus is attended if the null stimulus produces no response to begin with, which was the case in our dataset (Figure 3A, left panel). Specifically, if RN = 0, the input gain model reduces to RP,N = λβPRP, which cannot explain the decrease in firing rate observed when attention is directed to the null stimulus unless the scaling parameter λ changes with the direction of attention (preferred versus null). The normalization model of attention (Text S1) also acts by multiplying the inputs by a gain term and, in this regard, is similar to the input gain model. In addition, the responses are divided (normalized) by a term that depends on the null stimulus contrast and null attentional gain, even if the null stimulus produces no response. The normalization model can effectively change the scaling term of the gain model (λ) with changing attention, and therefore can explain a wider range of experimental results [19],[20],[24]. Several studies have shown that increasing the strength of incoming excitation increases the power in a broad frequency band above ∼30 Hz, including the gamma and high-gamma band, and this broad-band increase in power is correlated with the firing rate of the neural population near the microelectrode [40],[41]. This is different from “band-limited” gamma rhythm that is often visible in the power spectrum as a distinct “bump” with a bandwidth of ∼20 Hz, which is sustained by a inhibitory network [28],[42],[43], and may not be correlated with spiking activity [14],[34],[41]. Our results show that normalization increases band-limited gamma, while attention increases both excitation and normalization and therefore affects the power over a broader frequency range. Band-limited gamma may not always be observed during an attention task. For example, Khayat and colleagues [12] recorded from area MT of monkeys engaged in an attention task while presenting two random dot patterns—one moving in the null direction at 100% contrast paired with another moving in the preferred direction at varying contrasts, thus changing both excitation and normalization across stimulus conditions. The authors observed a broadband change in power in the gamma and high-gamma range, but no band-limited gamma. A similar spectral profile was observed in another recording from area MT where random dot patterns were used [17]. Indeed, most early studies that showed a salient band-limited gamma used one of two types of stimuli—gratings or oriented bars [44]–[46]. Most studies showing an effect of attention on band-limited gamma have also used either gratings or bars [1],[4],[5],[32],[47]. The absence of a prominent band-limited gamma rhythm in a demanding attention task [12] suggests that band-limited gamma may not play a functional role in attention and may not even be a fundamental marker of normalization or excitatory–inhibitory interactions. Instead, it could be a rhythm that is generated under special stimulus conditions and may reflect excitatory–inhibitory interactions within those restricted conditions. In this paper we have only considered a specific type of normalization, which is due to the addition of a nonoverlapping null stimulus inside the receptive field. Response suppression also occurs when an overlapping null stimulus is added to a preferred stimulus inside the receptive field, or when the stimulus size exceeds the classical receptive field (surround suppression). Whether these forms of suppression involve the same normalization circuit is unclear. For example, although earlier models of suppression produced by overlapping orthogonal gratings were based on inhibition [21],[22], recent models have explained this suppression without inhibition (for a review, see [48]). Consistent with this, a recent paper has shown that superimposing a null grating on a preferred grating decreases the gamma power in the primary visual cortex (V1), and surprisingly, also increases the gamma center frequency [49]. It is possible that superimposed and nonoverlapping orthogonal gratings produce suppression by different mechanisms, with only the latter requiring inhibition. Similarly, the presentation of a stimulus that is larger than the classical receptive field suppresses the response, although this manipulation increases the gamma power and decreases the gamma oscillation frequency in V1 [16]. The mechanism of surround suppression is unclear, with some studies showing an increase in incoming excitation and inhibition [50] and others showing the opposite effect [51]. Similarly, the cortical sites where normalization acts are also unclear. Earlier models assumed that normalization occurred simultaneously in multiple areas (V1 and MT; [52],[53]). However, properties of some types of opponent motion suppression differ between V1 and MT, which has been explained by a mechanism in which suppression arises in area MT [54]. On the other hand, responses of MT neurons that respond to the global motion of plaids (but not to the constituent component motion) were explained by a model where divisive normalization instead occurred in V1 [55]. Chalk and colleagues [5] have recently shown that gamma power decreases in area V1 with increasing attention, although under identical conditions gamma increases in V4. The differences could be due to the ways normalization is implemented in different cortical areas (see [5] for a more detailed discussion). In summary, the normalization signal that is involved in response suppression could be computed using different mechanisms, depending on the specific stimulus properties and cortical area. At present, it is unclear how universal the relationship between gamma and normalization described in this article is; that is, whether other forms of normalization would also modulate gamma power in a similar way. Similarly, although the stimulus configuration used in this article (nonoverlapping orthogonal stimuli inside the receptive field) is a common design used in several attention studies [24],[30],[31],[35],[37], the relationship between attention and gamma when other forms of normalization may be operating remains an open question. In our data, manipulations of normalization strength affected only the gamma range (and very low frequencies that likely reflected a stimulus transient). Attention, on the other hand, decreased power at low frequencies, consistent with prior studies [12],[32],[33] and increased power in the gamma and high-gamma ranges. As described above, a broadband increase in gamma and high-gamma power is correlated with the firing rate of the neural population near the microelectrode [40],[41]. However, in this study we observed an increase in gamma and high-gamma power even when attention was directed to the null stimulus. This is at odds with a previous study where gamma and high-gamma power decreased, consistent with the decrease in firing rate [12]. There are several factors that may have contributed to this difference. First, Khayat and colleagues [12] measured gamma power 510–1,010 ms after stimulus onset, while we measured gamma power between 50 and 250 ms after stimulus onset. It is possible that stimulus onset excites the entire population transiently, before suppressive and attention-related mechanisms take over to modify the responses of the neural population. The effect would be a transient increase in overall firing followed by a reduction in firing of the population, which may explain why high-gamma power is high initially (when we recorded) but lower in the steady state (when Khayat and colleagues recorded). Another factor may be the spatial spread of attention. As described earlier, high-gamma power depends on the firing rate of the overall population near the microelectrode, not just of the neuron being recorded from the microelectrode. Directing attention to the null stimulus inside the receptive field has two opposing effects: an increase in the firing rate of most neurons in the attended cortical region, and a reduction in the firing rate of neurons whose receptive fields contained both the preferred and null stimuli (such as the neurons shown in Figure 6A). Depending on the focus of attention, the overall population activity could either increase or decrease. Importantly, the changes in high-gamma power with attention do not influence the main result of this article, which is the increase in band-limited gamma power with increasing normalization strength. Because the stimuli used by Khayat and colleagues did not produce a salient band-limited gamma rhythm (see above), the results between the two studies cannot be compared directly. The lack of change in high-gamma power with increasing normalization strength (Figures 3B and 5B) can be explained similarly. A single stimulus activates a population of neurons, whose firing rate decreases when a second orthogonal stimulus is added (due to normalization and surround suppression). However, the second stimulus also activates another population of neurons. The overall population firing recorded by the microelectrode depends on the stimulus size, the size of the receptive field, suppressive surround and normalization pool, as well as the cortical spread of the population activity that is picked up by the microelectrode. It is possible that the overall population firing rate did not change appreciably when a second stimulus was added in our normalization protocol, so that high-gamma power did not change. The gamma peak was observed between 65 and 80 Hz, a frequency range that is slightly above the traditional gamma range (30–60 Hz) and that overlaps with the high-gamma band [41],[56]. This could be due to the early time window for analysis (because the stimulus presentation was for a short duration), because gamma peak frequency is higher after stimulus onset and decreases with time (for example, see Figure 1H of [41]). This is also consistent with a previous report that showed gamma oscillations at ∼50 Hz when analysis was done at a late interval (>300 ms) but a peak at 65 Hz when analysis was done at an early period ([1], compare their Figure 1 versus 4). In addition, gamma center frequency varies from subject to subject depending on the resting GABA concentration [57], and also depends on stimulus parameters such as size [16],[34] and contrast [13]. Although the center frequency of the gamma rhythm was relatively high, it could be dissociated from high-gamma activity (related to population firing) based on the spectral profile because gamma rhythm between 65 and 80 Hz had a distinct bump in the power spectrum while the high-gamma activity had a broadband profile with no distinct peak. Nonetheless, because the effect of spiking activity is detectable above ∼50 Hz in the LFP and becomes progressively more significant with increasing frequency [41], the increase in gamma power due to attention could partly be due to the increase in the population firing rate. In addition, as discussed above, gamma power depends not only on suppressive normalization, but also on the strength of the incoming excitation, and its precise relation with excitation and inhibition is unknown. Consequently, the increases in gamma power due to attention and to normalization were not tightly correlated in our data (unlike the tight correlation observed in firing rates as described in [19],[24]). Only when attention was directed to the null stimulus, for which the increase in the incoming excitation was less (although not zero, because the high-gamma power increased significantly), could we observe a weak correlation between attention and normalization (Figure 8B). In summary, our study shows that changes in the strength of normalization, which occur during attentional modulation, can also change the gamma power, although the precise nature of the relationship between normalization and gamma remains to be established. Changes in gamma power in an attention task due to changes in the underlying normalization strength must be accounted for before a more advanced functional role for gamma in the formation of communication channels [3],[10] or binding of stimulus features [7],[8] can be unequivocally established. All procedures related to animal subjects were approved by the Institutional Animal Care and Use Committee of Harvard Medical School. This study uses the same dataset as used by Ni and colleagues [24]. Data were collected from two male rhesus monkeys (Macaca mulatta) that weighed 8 and 12 kg. A scleral search coil and a head post were implanted under general anesthesia. After recovery, each animal was trained to do an orientation change detection task. The animal was required to hold its gaze within 1.0° from the center of a small fixation target while a series of drifting Gabor stimuli were flashed at three locations: two within the receptive field of the MT neuron being recorded and one at a symmetric location on the opposite side of the fixation point from the receptive field. All three Gabors were centered at the same eccentricity from the fixation point, and the Gabors were identical except for their contrast and drift direction. The two stimulus locations in the receptive field were separated by at least 5 times the SD of the Gabors (mean Gabor SD, 0.45°; SD of Gabor SD, 0.04°; range, 0.42–0.50°; mean separation of Gabor centers, 4.2°; SD, 0.86°; range, 2.2–6.9°). The stimuli were presented on a gray background (42 cd/m2), which had the same mean luminance with the Gabors, on a gamma-corrected video monitor (1024×768 pixels, 75 Hz refresh rate). The animal was cued to attend to one of the three locations in blocks of trials and to respond when a Gabor with a different orientation appeared there (the target), ignoring any orientation changes at uncued locations (distractors), which occurred with the same probability as changes at the cued location. The animal indicated its response by making a saccade directly to the target location within 100–600 ms of its appearance. Correct responses were rewarded with a drop of juice or water. The target location was cued by a yellow annulus at the beginning of each trial as well as by instruction trials. Instruction trials consisted of a series of Gabor stimuli that appeared in only one location. Two instruction trials were inserted each time the cued location changed. Gabors were presented synchronously in all three locations for 200 ms, with successive stimuli separated by periods with pseudorandom durations of 158–293 ms. During each presentation, one Gabor inside the receptive field moved in the preferred direction of the neuron, while the other Gabor inside the receptive field moved in the opposite (null) direction. The Gabor outside the receptive field moved in an orthogonal (intermediate) direction. The “Normalization” and “Spatial Attention” protocols differed in the location of the cue (outside versus inside the receptive field) and the number of contrasts used for each stimulus (three versus two). For the Normalization protocol (Figure 1A), the monkey attended to the stimulus outside the receptive field, and all Gabors could take one of three contrast values: 0%, 50%, or 100% (the target stimulus had either 50% or 100% contrast). This created nine different stimulus conditions inside the receptive field, as shown in Figure 3 (for each condition, we pooled data for the three different contrast levels for the Gabor outside the receptive field). For the Spatial Attention protocol (Figure 1B), the monkey attended to one of the locations inside the receptive field (which could have either the preferred or null stimulus in different presentations). All Gabors had either 0% or 100% contrast (target stimulus always had 100% contrast). We only used the stimulus condition for which both the preferred and null stimuli inside the receptive field had 100% contrast because that configuration showed the largest effect of attention. The stimulus at a given location inside the receptive field could either be the preferred or null stimulus across presentations within the same trial (Figure 1). For a subset of data recorded from Monkey 1 (45 out of 68 neurons), the stimulus direction was fixed for a given location, so that the preferred stimulus always appeared in the bottom half of the receptive field while the null stimulus always appeared on top. The results shown in the article were similar for this modified version of the task; the data were pooled. The timing of the target appearance in each trial was selected from an exponential distribution (flat hazard function for orientation change) to encourage the animal to maintain constant vigilance throughout each trial. However, trials were truncated at 6 s if the target had not appeared (∼20% of trials), in which case the animal was rewarded for maintaining fixation up to that time. The orientation change was adjusted for each stimulus configuration using an adaptive staircase procedure (QUEST; [58]) to maintain a behavioral performance of 82% correct [hits/(hits+misses); range, 57%–93%] across all target locations [the average orientation change for targets and distractors were 50±12° and 52±7° for Monkeys 1 and 2 (mean±SD)]. Both monkeys had fast reaction times (245±13 and 195±7 ms; mean ± SD), which, coupled with the large attentional modulation observed in the firing rates, suggested that they were paying close attention to the stimuli. Recordings were made using glass-insulated Pt-Ir microelectrodes (∼1 MΩ at 1 kHz) in area MT (axis ∼22–40° from horizontal in a parasagittal plane). A guide tube and grid system [59] was used to penetrate the dura. Spikes and LFP were recorded simultaneously using a Multichannel Acquisition Processor system by Plexon Inc. with a head-stage with gain 20 (Plexon Inc. HST/8o50-G20). Signals were filtered between 250 Hz and 8 kHz, amplified and digitized at 40 kHz to obtain spike data. For the LFP, the signals were filtered between 0.7 and 170 Hz, amplified and digitized at 1 kHz. We used the FPAlign utility program provided by Plexon Inc. to correct for the filter induced time delays (http://www.plexon.com/downloads). The headstage HST/8o50-G20 has low input impedance, which can lead to a voltage divider effect at low frequencies (Figure 2B shows this effect at frequencies below ∼5 Hz) [60]. This is unlikely to affect our results because this effect is much less prominent in the frequency range of interest (65–80 Hz) and we always compared data across different stimulus conditions that had the same filter settings. Once a single unit was isolated, the receptive field location was estimated using a hand-controlled visual stimulus. Computer-controlled presentations of Gabor stimuli were used to measure tuning for direction (eight directions) and temporal frequency (five frequencies) while the animal performed a fixation task. The temporal frequency that produced the strongest response was used for all of the Gabors. The temporal frequency was rounded to a value that produced an integral number of cycles of drift during each stimulus presentation, so that the Gabors started and ended with odd spatial symmetry, such that the spatiotemporal integral of the luminance of each stimulus was the same as the background. Spatial frequency was set to one cycle per degree for all of the Gabors. The preferred Gabor was used to quantitatively map the receptive field (three eccentricities and five polar angles) while the animal performed a fixation task. The two stimulus locations within the receptive field were chosen to be at equal eccentricities from the fixation point and to give approximately equal responses, and the third location was 180° from the center point between the two receptive field locations, at an equal eccentricity from the fixation point as the other locations. Cells were included in the analysis if they were held for at least nine repetitions (mean 41 repetitions) of each stimulus/attention combination used in this article. The response for each condition was taken as the average rate of firing in a period 50–250 ms after stimulus onset. Target stimuli and stimuli presented with a distractor were excluded from analysis, as were stimuli that appeared after the target. Additionally, the first stimulus presentation in each trial was excluded from analysis to reduce variance arising from stronger responses to the start of a stimulus series. Instruction trials were excluded from data analysis. Spikes and LFP were collected from 68 sites from Monkey 1 and 50 from Monkey 2. Out of these, 13 and 9 sites were discarded because either the LFP signal was too large and saturated frequently or was too weak (<10 µV). The results were similar (and individually significant) for the two monkeys, and the gamma oscillations were also in the same frequency range; the data were pooled. Time-frequency analysis was performed using the Matching Pursuit algorithm [61]. Due to the rapid presentation of the stimuli (duration of 200 ms with interstimulus interval of 158–293 ms), the LFP signal had transient activity associated with stimulus onset/offset. This required time-frequency analysis over short intervals (i.e., good temporal but poor spectral resolution). On the other hand, line noise at 60 Hz and the monitor refresh rate at 75 Hz produced signals at constant frequency (60 and 75 Hz), which were sustained for long periods (Figure 2). To represent such signals, time-frequency analysis should be done over long intervals (to achieve good spectral resolution at an expense of temporal resolution). These requirements are difficult to fulfill using traditional signal processing techniques such as short-time Fourier Transform or multi-tapering, but can be addressed using multiscale analysis techniques such as Matching Pursuit [61]. In this method, we start with an overcomplete dictionary of Gabor functions that have a wide range of time-frequency resolutions, including delta functions and sinusoids. The functions that best represent the signal are chosen for representation using an iterative procedure [26]. In this article, Matching Pursuit analysis was done on 1-s-long LFP segments, so the line noise at 60 Hz and the weaker noise at the monitor refresh rate of 75 Hz were captured by sinusoidal functions, which had a spectral resolution of ∼1 Hz, resulting in sharp lines at 60 and 75 Hz (Figure 2). Although Matching Pursuit algorithm provides better resolution to resolve transient and sustained activity, the results obtained using the multitaper method were similar (Figure S1). For each site, first a common “baseline power spectrum” was computed by averaging the power between 100 to 0 ms before stimulus onset for all nine normalization conditions (denoted by Baseline(ω); Figure 2B, black line). For Figure 3B and 6B, the time-frequency power spectra were normalized by this baseline power [10.(log(Power(t,ω)−log(Baseline(ω))]. Note that all the plots were normalized by the same baseline power (average of the baseline power obtained from the nine normalization conditions), which eliminates the possible effects of differences in baseline power across conditions. We showed changes in LFP power instead of raw power because LFP has a prominent “1/f” structure with more energy at low frequencies, which makes it difficult to observe any changes at higher frequencies in the raw time-frequency power spectra. Further, the difference spectra do not show the line and refresh-rate-related noise because this noise is present before stimulus onset also. The difference spectra were smoothed by averaging the power in every 4 time and frequency bins (essentially downsampling by a factor of 4 in both dimensions). This smoothing was done only for better visual display; all the power versus frequency/time plots (Figures 4, 5D, and 7A) as well as the power difference calculations (Figures 5, 6C, 7B, and 8) were done using raw data. The gamma power was computed by summing the power between 65 and 80 Hz, but excluding the monitor refresh rate (between 74 and 76 Hz). Power from each condition was divided by the power for the P100N0 condition before averaging across neurons. High-gamma power was taken between 80 and 135 Hz because we observed a noise peak between 140 and 150 Hz, possibly arising from the stepper motor used to drive the microelectrodes when it was not moving, and the power above 150 Hz was attenuated by the low pass filter in the Plexon recording system.

10.1371/journal.pntd.0007303

Comparative accuracy of typhoid diagnostic tools: A Bayesian latent-class network analysis

Typhoid fevers are infections caused by the bacteria Salmonella enterica serovar Typhi (Salmonella Typhi) and Paratyphi A, B and C (Salmonella Paratyphi). Approximately 17.8 million incident cases of typhoid fever occur annually, and incidence is highest in children. The accuracy of current diagnostic tests of typhoid fever is poorly understood. We aimed to determine the comparative accuracy of available tests for the pediatric population. We first conducted a systematic literature review to identify studies that compared diagnostic tests for typhoid fever in children (aged ≤15 years) to blood culture results. We applied a Bayesian latent-class extension to a network meta-analysis model. We modelled known diagnostic properties of bone marrow culture and the relationship between bone marrow and blood culture as informative priors in a Bayesian framework. We tested sensitivities for the proportion of negative blood samples that were false as well as bone marrow sensitivity and specificity. We found 510 comparisons from 196 studies and 57 specific to the pediatric population. IgM-based tests outperformed their IgG-based counterparts for ELISA and Typhidot tests. The lateral flow IgG test performed comparatively well with 92% sensitivity (72% to 98% across scenario analyses) and 94% specificity. The most sensitive test of those investigated for the South Asian pediatric population was the Reverse Passive Hemagglutination Assay with 99% sensitivity (98% - 100% across scenario analyses). Adding a Widal slide test to other typhoid diagnostics did not substantially improve diagnostic performance beyond the single test alone, however, a lateral flow-based IgG rapid test combined with the typhoid/paratyphoid (TPT) assay yielded improvements in sensitivity without substantial declines in specificity and was the best performing combination test in this setting. In the pediatric population, lateral-flow IgG, TPT and Reverse Passive Hemagglutination tests had high diagnostic accuracy compared to other diagnostics. Combinations of tests may provide a feasible option to increase diagnostic sensitivity. South Asia has the most informed set of data on typhoid diagnostic testing accuracy, and the evidence base in other important regions needs to be expanded.

Typhoid fever is an infection caused by the bacterium Salmonella Typhi. Typhoid fever is rare in developed countries but remains high in the developing world. Effective treatment is available but accurate diagnosis of typhoid fever is challenging as typhoid fever can be difficult to distinguish from other infections. Bone marrow culture is the most accurate diagnostic test for typhoid fever however is invasive and not feasible in many settings. New vaccines for typhoid and the need for improved estimates of burden increases the demand for improved understanding of diagnostic accuracy. Comparing the diagnostic accuracy of tests for typhoid fever is challenging as head-to-head studies are few. We applied newly developed methods for comparative evaluation of diagnostic tests for typhoid fever in children using statistical approaches that allowed for the proper incorporation of uncertainty and comparison of tests that had not been compared directly. The lateral-flow IgG, TPT and Reverse Passive Hemagglutination tests all had good diagnostic accuracy compared to other diagnostics. Combinations of tests may provide a feasible option to increase diagnostic sensitivity. Finally, while South Asia has the most informed set of data on typhoid diagnostic testing accuracy, the evidence base in other important regions needs to be expanded.

Typhoid fever (also known as enteric fever) is a systemic infection caused by the Gram-negative bacteria Salmonella enterica serotypes Typhi or Paratyphi A,B and C[1],[2]. While rare in developed countries, the burden of typhoid remains high in developing countries. Recent annual estimates of typhoid fever cases in low- and middle-income countries range from approximately 17.8 million[3] to 26.9 million[4] cases worldwide and most of these are in South Asia. The pediatric population is of particular interest as most cases occur in those between 3 and 19 years of age[1], the highest incidence of typhoid occurs in those less than 5 years of age[5]. Recent modelling work reported a higher incidence among children aged two to four years compared to those less than two years.[3] With the recent World Health Organization pre-qualification of, and GAVI commitments towards, a typhoid conjugate vaccine for use in routine immunization programs, there is a need for better data on typhoid burden in young children, which requires better understanding of diagnostic accuracy. Prior meta-analyses have focused on all age groups without distinguishing performance in children; however, we hypothesize that diagnostic accuracy may differ between children and adults due to a greater degree of prior exposure to Salmonella and other pathogens in adults, leading to serologic cross reactivity. If diagnosed promptly, typhoid can be successfully treated with antibiotics. [1, 2] Accurate diagnosis of typhoid fever has proved a major challenge. Clinical signs and symptoms are often non-specific, and typhoid can be difficult to distinguish from other acute febrile illnesses, including dengue, malaria, influenza, leptospirosis, and Rickettsial infections[6–8]. The definitive diagnosis for typhoid fever is via isolation of S. Typhi from blood, bone marrow or other sterile sites.[1] The most sensitive and specific diagnostic test for typhoid fever is bone marrow culture; however, as this test is invasive, carries risks of medical complications, and requires technical expertise and specialized equipment, it is not widely performed in endemic settings as a routine diagnostic procedure. Among culture-based methods, blood culture is the most commonly used typhoid diagnostic method, but results are not available for days, and many settings lack the resources required for proper culturing techniques. Furthermore, it has limited sensitivity (40–75% in most settings)[9, 10], which may be further diminished by prior antibiotic use. The Widal test, developed in the late 19th century to measure antibodies against the O and H antigens of Salmonella, remains perhaps the most widely used typhoid diagnostic in the world. However, the Widal test only has moderate sensitivity and specificity, particularly in endemic settings, and there remains a challenge of determining a proper cut-off point for a positive result[5, 11]. Indeed, rapid and reliable (>90% sensitivity and specificity) diagnostics do not yet exist for invasive salmonellosis. The Reverse Passive Hemagglutination (RPHA) Test, that detects the S. Typhi antigen, was found to have a sensitivity and specificity that is comparable with the Widal test leading to suggestion that it could be used as an alternative to the Widal test in busy microbiology laboratories[12, 13]. Newer diagnostic tests, such as the antibody tests Typhidot and Tubex, have demonstrated moderate accuracy[14]. The typhoid/paratyphoid diagnostic assay (TPT test) has shown promising results.[15] Polymerase chain reaction (PCR) and other molecular, transcriptomic and metabolomic methods have been developed, but they have yet to be evaluated in large scale settings. Assessing the comparative performance of diagnostic testing is challenging as few head-to-head evaluations exist and previous reviews of diagnostic testing have found a high level of variation in testing methods for typhoid fever globally and a lack of a single applicable gold standard, a challenge that is particularly acute given the low sensitivity of the most common reference standard, blood culture.[3, 9] We aimed to assess the comparative performance of typhoid diagnostics using newly developed methods for comparative evaluations [16]. In particular, we combined a Bayesian network meta-analysis (NMA) procedure with latent class analysis. [16]. We developed a comprehensive search strategy to identify relevant studies comparing diagnostic tests for typhoid disease. We particularly considered typhoid fever to include Salmonella Typhi and S. Paratyphi A. We searched the following databases: EMBASE, MEDLINE, ISI Web of Science and the Cochrane Central Register of Controlled Trials from inception to December 26, 2016. We also scanned references from systematic reviews on typhoid diagnostic tools identified via the above search. We conducted a grey literature search of Google Scholar and the National Institutes of Health Research Portfolio Online Reporting Tools (NIH RePORT). We searched conference proceedings of the International Conference on Typhoid and Other Invasive Salmonelloses and the American Society of Tropical Medicine and Hygiene Conference, and unpublished data submitted by the originator companies to the US Food and Drug Administration and the European Medicines Agency as part of diagnostic registration applications. Additionally, we performed manual searches of clinicaltrials.gov and the WHO International Clinical Trials Registry Platform to identify studies that have not yet been published but have results and were potentially eligible for inclusion. Specific search terms and results by database are provided in S1 Table. We also engaged key leaders from disparate agencies that conduct research in diagnostic development, including, but not limited to the U.S. Department of Defense (Walter Reed Army Institute of Research and Defense Advanced Research Projects Agency) and non-profit research institutions and diagnostic development organizations. All abstract and full-text screening of studies was done in duplicate. Data extraction was completed using a standardized data extraction form. The extraction form was designed for this study and pilot tested by the authors. A copy of our extraction form is included in S2 Table. We extracted all comparisons across diagnostic tests as well as within any relevant subgroups presented in the included studies. Study characteristics of interest for extraction included: detailed description of diagnostic tests used including the details of any commercial tests used, types and volume of biological specimen, study location (detailed location, country and coded into World Bank region), broad age group of study population, duration of illness (most often reported as duration of fever), patient reported antibiotic self-treatment/use prior to study entry. For studies where subgroup data were not reported, study authors were contacted for age-specific contingency tables. Data were analyzed at the study level and at the level of individual test comparison (index test versus reference test) with both test result and disease status dichotomized. Pair-wise meta-analysis or network meta-analysis was only done in a subset of studies. This subset was in populations of children, approximately aged 15 or younger (in some cases, it was clear that most subjects were children, but we could not be certain that teenagers and those over 15 years of age were not included) that used blood culture alone as the diagnostic reference test and were conducted in one of three World Bank regions: South Asia, East Asia & Pacific (EAP) and sub-Saharan Africa. These restrictions were introduced to reduce heterogeneity across studies, make synthesis results more interpretable, and focus on pediatric cases in typhoid endemic regions. From a combined 1,749 records identified, there were 196 studies included for full-extraction (See Fig 1 for flow diagram). From these studies, 57 comparisons between tests from 32 studies were included for the NMA (studies listed in Table 1). Full datasets for study level characteristics and comparison level data are presented in S3 and S4 Tables. A glossary of terms is provided in S5 Table. The summary results of the search are presented in Tables 2 and 3 separated by the full set of studies and the subset of studies included the NMA. The full set of studies includes all 196 identified studies in our search that represented 510 pairwise comparisons between two typhoid diagnostic tests. The subset of 32 studies used in NMA represented 57 comparisons. Study level characteristics for 196 included studies are presented in S3 Table and summarized in Table 2. Among the full set of studies, the majority were conducted in areas of high typhoid endemicity (68.4%), and 72.4% of studies were conducted in either South or East Asia (World Bank Regions classification). There was a relatively even distribution of patient age mixes between adults and children in the studies. However, many studies did not report age, and among the 62 studies that included both adults and children, no subgroup results were reported by age. Just over half of the studies (60.4%) included less than 200 patients with few studies containing more than 1000 patients. There was a slightly higher proportion of newer (post 2000) studies in the full dataset with the majority of studies in the network analysis set being conducted in 2010 or later. In both the full set of studies and the network, the majority of studies (59.2%) did not provide details on the volume of biological specimen collected for the tests or the duration of symptoms (58.7%). Prior antibiotic use can greatly influence the sensitivity of blood culture; however, 72.3% of studies did not report on this characteristic. For those studies that did provide these data we have presented these in Table 2. Pairwise, summary estimates for meta-analysis of testing characteristics are presented in Table 3 and as forest plots in S1–S4 Figs. For our network, the numbers of comparisons across each of the six types of index and reference diagnostic tests categorized by date of publication is presented in Fig 2 and summarized in Table 4. The most common comparisons in the full set of 510 comparisons were index tests using antibody, Widal and molecular diagnostics contrasted to viable bacteria culture tests. While the Widal test is the most widely used diagnostic test for typhoid in endemic regions, the majority of the literature focused on evaluating the performance of other antibody tests. The graphical network of comparisons with the NMA set across all index and reference tests for is presented in a network structure in Fig 3A. A network of evidence was generated overall (Fig 3A) and for each World Bank Region under study (Fig 3B–3D). The testing characteristics generated from Bayesian analysis are presented in Tables 5–8. Across all regions combined (Fig 3A and Table 5), rapid tests had both high sensitivity and specificity estimates. Among rapid tests, the reverse passive hemagluttination antigen test had 99% sensitivity (72% to 100% across scenario analyses) and 92% specificity; Typhidot IgM outperformed Typhidot IgG with 80% sensitivity (70% to 85% in scenario analyses) and 95% specificity; and Typhidot IgM or IgG had 91% sensitivity (86% to 93% in scenario analyses), however with specificity of 86%. ELISA IgM outperformed its IgG counterpart and the TPT test also performed very well with 94% sensitivity (76% to 100% in scenario analysis) and a specificity of 97%. The best Widal test appeared to be a 1:160 titer for the H-antigen slide test, yielding a sensitivity of 79% and a specificity of 98%. Lastly, the most sensitive test of all tests investigated for the pediatric population was the reverse passive hemagluttination antigen test however scenario analyses did yield fairly large model variability. For EAP (Fig 3B and Table 6), the rapid test lateral flow IgM and PCR had very low sensitivity compared to the latent class bone marrow reference test (13% and 7% respectively). TUBEX TP, O12 was associated with a sensitivity of 79%, which was the highest among the investigated tests, and a specificity of 99%. ELISA IgG was inferior to ELISA IgM. The scenario analyses yielded modest sensitivity with ELISA IgM possibly yielding sensitivity up to 67%. For Sub-Saharan Africa (Fig 3C and Table 7), ELISA Total Ig appeared superior to the other investigated tests with a sensitivity of 85% (81% to 88% in scenario analyses) and 92% specificity, which was the lowest specificity observed in the network analysis. Both Widal tests had very low sensitivity (<25% across all scenario analyses). For South Asia (Fig 3D and Table 8), several rapid tests had both high sensitivity and specificity estimates. Among the rapid tests, the lateral-flow immunochromatographic dipstick IgG assay had 92% sensitivity (72% to 98% across scenario analyses) and 94% specificity; Typhidot IgM outperformed Typhidot IgG with 74% sensitivity (65% to 80% in scenario analyses) and 97% specificity; and Typhidot IgM or IgG had 79% sensitivity (76% to 91% in scenario analyses), however with specificity of 90%. ELISA IgM outperformed its IgG counterpart and the TPT test also performed very well with 90% sensitivity (72% to 99% in scenario analysis) and a specificity of 93%. The best Widal test appeared to be a 1:80 titer for the H-antigen slide test, yielding a sensitivity of 76% and a specificity of 99%. Lastly, the most sensitive test of all tests investigated for the South Asian pediatric population was Reverse Passive Hemagglutination with 99% sensitivity and scenario analyses did not yield large model variability. Sensitivity and specificity of hypothetical combination tests are presented in Table 9 and were estimated for the South Asian population only, since none of the rapid tests in our subset of data were associated with good test performance characteristics in the two other World Bank regions. For acute care pediatric subjects tested in the South Asian setting, adding the ‘best’ Widal test (i.e., H-antigen slide test with cut-off 1:80) to any of the three highest performing rapid tests (reference tests: lateral flow IgG, TPT, and Typhidot IgM or IgG) did not yield marked improvements. Conversely, adding a lateral flow-based IgG rapid test to the TPT approach yielded improvements in sensitivity without substantial declines in specificity and was the best performing test combination. The results of this analysis builds the evidence base for typhoid diagnostics and is the first attempt to apply newly developed comparative methods for diagnostics testing accuracy.[16] This review and approach yielded several key insights. First, the body of studies on typhoid diagnostics and within study estimates of diagnostic accuracy were highly heterogeneous, even when restricting to studies with similar populations and study designs. Second, despite this heterogeneity, certain diagnostics consistently outperformed others; in particular, IgM-based ELISA and Typhidot outperformed their IgG-based counterparts, and the IgA-based TPT Test performed well in South Asia. Finally, the analytic methods allowed us to generate estimates for test performance based on combinations of tests. We found that combinations of existing sensitive and specific diagnostics may overcome the accuracy limitations inherent in single diagnostics, achieving what may be sufficient accuracy for use in certain clinical settings. Applying these methods allows us to generate estimates for test performance based on combinations of tests. This analysis has also provided comparative estimates of diagnostic testing accuracy for specific tests and targets across a more homogenous set of studies with similar age ranges, geographies and reference tests. This is an important addition because of the wide variety of test types within a family of targets such as antibody or antigen. Though there is an issue of regional variation in antibody response, the majority of our studies were from typhoid endemic regions likely with similar diagnostic titer cut-offs. This expanded and more detailed evidence base allows for more precise comparative assessments of diagnostic testing accuracy via indirect comparisons or network analysis. The methods and results of this meta-analysis differ from previous meta-analyses of typhoid diagnostics, including those of Storey et al[9] and Wijedorou et al[51] in several ways. First, previous studies have focused on specific products rather than antigen/antibody combinations and performed single comparisons against a reference standard (a composite reference standard or blood culture), without performing between study comparisons through a network framework. We used latent class analysis to account for imperfect reference standards, which is critical given the low sensitivity of blood culture. Additionally, prior analyses focused on single diagnostics without examining their performance in combination and concluded that accuracy was insufficient. By focusing on diagnostic types and their combinations, and utilizing a network meta-analytic framework, we found that certain combinations of diagnostics exceeded 90% sensitivity and specificity. Our analysis provides evidence that IgM-based ELISA and Typhidot assays diagnostics outperformed their IgG counterparts. Thriemer et al[14] performed a SLR and meta-analysis of the performance of Tubex TF and Typhidot in typhoid endemic countries and concluded that neither test was exclusively reliable for the diagnosis of the disease. Storey et al.[9] also concluded that no single test has sufficiently good performance but suggested that some existing diagnostics could be useful as part of a composite reference standard. Our exploration of combination tests found, in the South Asian pediatric setting, combining a lateral flow IgG assay with the IgA-focused TPT test yields a high performing diagnostic combination. Combinations of the widely used Widal test and tests with good performance characteristics in Bayesian latent class analysis (lateral flow IgG or TPT test) did not yield substantial improvements to the individual tests alone. We found that DNA-based tests, whether nested or not, performed similarly with limited sensitivity but high specificity. DNA diagnostic tests were few in our selected group of studies in children, likely due to the small blood volumes drawn from children and the need for substantial volumes for direct molecular diagnostics. The appeal of molecular diagnostics is that they can be more specific than serologies, more rapid than culture, and potentially less affected by prior antibiotic use. The main limitation is that the organism burden in blood during typhoid fever has been estimated at 0.1–1 CFU/ml[52]. For detection to be possible, a large volume of blood is needed, together with highly efficient DNA extraction, concentration and amplification. As a result, in practice, sensitivity is variable but often modest. There are strengths and limitations to our analysis. Strengths include the extensive searching and identification of published and unpublished data. A further strength is the application of hierarchical modelling using the latent class analysis as it examines the strength of statistical relationships among variables. The analysis was also strengthened by our efforts to limit between-study heterogeneity through only including studies where: a reference test was included, the patient population consisted of children, and select geographical regions were examined. We assessed the potential for regional differences in diagnostic performance by dividing countries into World Bank regions; while these divisions are imperfect and the epidemiology may vary substantially within regions, there was not substantial variation in results in the NMA dataset, with few countries providing the majority of data. Our results were derived from data among children, who may be less likely to have prior exposure to typhoid and other infections compared with adults. It is possible that serologic cross reactivity to other pathogens may be more common in adults, and diagnostic accuracy may be lower. Therefore, we caution against extrapolating these findings to other age groups. This study had several limitations. These were predominantly related to lack of studies in populations of interest to us. The majority of studies have been small, with over half of studies having less than 200 patients. In these studies–the risk of bias is high due to lack of statistical power and the higher chance of sampling bias. Furthermore, many of the studies were done using convenience sampling which leads to undefined study populations as whomever presented with index symptoms were included. Our results suggest there is a need for additional large sample studies of new methods/technologies to be confidently judged for their diagnostic accuracy. This echoes the conclusions of previous reviews and meta-analyses despite an enlarged and enhanced evidence base.[9] Further, in studies where a composite reference is used–there is a need for additional standardization of techniques and what constitutes a composite standard. In our attempt to extract specific data reference tests, different combinations of tests were used as the composite standard which complicates comparison across studies. One of the challenges in summarizing evidence across diagnostic tests, such as serologic tests and molecular tests, is that very few studies used the same diagnostic approaches. The studies evaluating serologies used various combinations of antigens (e.g. Vi, Omp, LPS), antibody isotypes (IgG, IgM, IgA), and assay formats (commercial versus in-house ELISA, immunoblot, lateral flow), while studies evaluating molecular diagnostics used varying gene targets, extraction methods and PCR platforms. We therefore aggregated these diagnostics into “antibody”, “antigen” and “PCR” based tests to facilitate analysis of overall accuracy by general broad method; however, this precluded a more nuanced synthesis of evidence on which specific approaches and targets perform better. A fundamental challenge with evaluating the accuracy of typhoid diagnostics is the lack of perfect reference standards. Bone marrow culture has the highest sensitivity, but was not used in most studies due to its invasiveness. Blood cultures, widely used due to their near perfect specificity, are only 50–65% sensitive. As a result, studies may inaccurately classify individuals with negative cultures as not having typhoid, which can in turn lead to under-estimates of the specificity of serologic diagnostics. To address this challenge and obtain comparative estimates of sensitivity and specificity with respect to bone marrow culture, we therefore applied a latent class extension to the conventional network meta-analysis model. The Bayesian framework allowed us to implement known diagnostics properties of bone marrow culture and the relationship between bone marrow and blood culture as informative priors to more accurately estimate the performance of various diagnostics. Serologic tests for S. Typhi pose a particular challenge because, while surface antigens for typhoidal Salmonella are generally conserved, they are also shared with many other Enterobacteriaceae.[53] This means that diagnostic kits aimed at a general mix of S. Typhi antigens frequently suffer from low specificity.[53] Further the titres and specificities of antibodies to the classical typhoidal antigens O, H and Vi, vary a great deal, as demonstrated by studies of typhoidal antibody titres in endemic settings[54]. These issues pose challenges to the development of serologic assays built on these targets. In conclusion, our analysis found a heterogeneous body of evidence for typhoid diagnostics. There is a high degree of variability in diagnostic testing characteristics across tests and regions even after restricting on patient population age, geographic region and reference test. Nevertheless, there are good combinations of existing tests that may provide opportunities in both for individual diagnosis as well as population-based surveillance. South Asia has the most informed set of data on typhoid diagnostic testing accuracy and the evidence base in other important regions needs to be expanded as the performance of diagnostics could vary by region and specific setting. In South Asia, there is evidence for good test performance of some rapid tests, but the evidence is variable due to limited numbers of studies once the data is stratified down by test type. Further work, particularly in the area of novel antigen detection, enhanced molecular diagnostic techniques, host transcriptional assays, metabolomic profiling and low-cost culture techniques all hold potential to drive real gains in the typhoid diagnostics space. Novel antigens specific for S. Typhi, as proposed by Baker et al[53], remains an exciting area of work given the variability of typhoid presentation. An important challenge would be the development of a panel of specific S. Typhi antigens that identify different stages of infection. These could be generated by testing cohorts of patients with protein microarrays in various specimen types to identify specific patterns of infection. Such studies, if fruitful, could lead to the development of low-cost assays. Novel culture techniques that are efficient and require minimal laboratory infrastructure would allow for improved burden estimation and a more accurate diagnosis, and therefore appropriate treatment.[55] To advance the evaluation of these new diagnostics, standardized clinical specimen biobanks representing multiple countries, populations and age groups should be established to facilitate direct comparison of multiple diagnostics against one another. Such a collaborative effort could help further overcome the limitations of population and diagnostic heterogeneity and imperfect reference standards that have limited diagnostic evaluation thus far, and accelerate the identification of accurate diagnostics for typhoid fever.

10.1371/journal.pntd.0002072

Activities of Daily Living Associated with Acquisition of Melioidosis in Northeast Thailand: A Matched Case-Control Study

Melioidosis is a serious infectious disease caused by the Category B select agent and environmental saprophyte, Burkholderia pseudomallei. Most cases of naturally acquired infection are assumed to result from skin inoculation after exposure to soil or water. The aim of this study was to provide evidence for inoculation, inhalation and ingestion as routes of infection, and develop preventive guidelines based on this evidence. A prospective hospital-based 1∶2 matched case-control study was conducted in Northeast Thailand. Cases were patients with culture-confirmed melioidosis, and controls were patients admitted with non-infectious conditions during the same period, matched for gender, age, and diabetes mellitus. Activities of daily living were recorded for the 30-day period before onset of symptoms, and home visits were performed to obtain drinking water and culture this for B. pseudomallei. Multivariable conditional logistic regression analysis based on 286 cases and 512 controls showed that activities associated with a risk of melioidosis included working in a rice field (conditional odds ratio [cOR] = 2.1; 95% confidence interval [CI] 1.4–3.3), other activities associated with exposure to soil or water (cOR = 1.4; 95%CI 0.8–2.6), an open wound (cOR = 2.0; 95%CI 1.2–3.3), eating food contaminated with soil or dust (cOR = 1.5; 95%CI 1.0–2.2), drinking untreated water (cOR = 1.7; 95%CI 1.1–2.6), outdoor exposure to rain (cOR = 2.1; 95%CI 1.4–3.2), water inhalation (cOR = 2.4; 95%CI 1.5–3.9), current smoking (cOR = 1.5; 95%CI 1.0–2.3) and steroid intake (cOR = 3.1; 95%CI 1.4–6.9). B. pseudomallei was detected in water source(s) consumed by 7% of cases and 3% of controls (cOR = 2.2; 95%CI 0.8–5.8). We used these findings to develop the first evidence-based guidelines for the prevention of melioidosis. These are suitable for people in melioidosis-endemic areas, travelers and military personnel. Public health campaigns based on our recommendations are under development in Thailand.

Melioidosis is a serious infectious disease caused by the environmental saprophyte, Burkholderia pseudomallei. The infection is potentially preventable, but developing prevention guidelines is hampered by a lack of evidence on which to base them. The purpose of this study was to provide evidence for inoculation, inhalation and ingestion as routes of infection. To achieve this, we undertook a matched case-control study and performed home visits to obtain drinking water and culture this for B. pseudomallei. We found that activities associated with increased risk of developing melioidosis included working in a rice field, other activities associated with exposure to soil or water, an open wound, eating food contaminated with soil or dust, drinking untreated water, outdoor exposure to rain, water inhalation, current smoking and steroid intake. Presence of B. pseudomallei in drinking water source(s) doubled the odds of acquiring melioidosis. This is the first study to show that ingestion is an important route of human B. pseudomallei infection, and that exposure to rain is an independent risk factor for melioidosis. We used this finding to develop the first evidence-based guidelines for the prevention of melioidosis. These are suitable for people in melioidosis-endemic areas, travelers and military personnel.

Burkholderia pseudomallei is a Category B select agent and the cause of naturally acquired melioidosis in South and East Asia, Northern Australia, the Indian subcontinent and areas of South America [1]–[3]. Northeast Thailand is a hotspot for this infection, with an annual incidence of 21.0 per 100,000 population and a crude mortality rate of 40% [4]. This rate is comparable to that for deaths from tuberculosis in this region, where melioidosis is the third most common cause of death from infectious diseases [4]. Visitors to areas where melioidosis is endemic are also at risk of acquiring this infection. Melioidosis is readily misdiagnosed in returning travelers because of a lack of familiarity with the clinical and microbiological features, compounded by a highly variable incubation period that may extend to many decades [5], [6]. The largest transient population to have been affected in living memory was US combatants in the conflict with Vietnam, when the disease acquired the nickname ‘Vietnamese time bomb’ [7]. B. pseudomallei is present in soil and surface water in areas where melioidosis is endemic, and most cases are thought to result from bacterial inoculation [8]. This is based on the observations that people at high risk of melioidosis such as agricultural workers in Thailand and indigenous people in Australia are regularly exposed to soil and water without protective clothing and may suffer repeated minor injuries [9], [10]. The role of other routes of infection is uncertain. Inhalation may have been a route of infection for US combatants during the Vietnam conflict [11], and several studies from northern Australia have reported a shift towards a higher frequency of pneumonia and severe disease during the rainy season or following heavy monsoon rains and winds [12]–[14]. Recent evidence also suggests that ingestion might be an important route of B. pseudomallei infection. West et al. showed that gastric inoculation of B. pseudomallei led to melioidosis in an experimental mouse model [15]. Several clusters of melioidosis cases have been reported from Australia in which a strain of B. pseudomallei isolated from a common water source was a genetic match for the strain causing disease in the cluster [16], [17], although it is not clear whether these cases were infected through ingestion rather than inoculation. Melioidosis is potentially preventable, but developing prevention guidelines is hampered by a lack of evidence on which to base them. Advice in Northern Australia is based on common sense and includes avoidance of direct contact with soil and standing water and washing after exposure [18]. There are no recommendations to prevent melioidosis via inhalation or ingestion. No advice is given in Asia or other places where melioidosis is endemic, and no advice is given to tourists despite the steady trickle of cases in returning travelers. Here, we describe a matched case-control study in which we identify activities associated with an increased risk of disease acquisition, define the importance of three routes of melioidosis infection, and describe the first evidence-based guidelines for the prevention of melioidosis. A prospective 1∶2 matched case-control study was performed at Sappasithiprasong Hospital between Jul 2010 and Dec 2011. This 1,100-bed hospital is situated in the provincial town of Ubon Ratchathani in northeast Thailand, 70 km west of Laos and 95 km north of Cambodia, and serves around 2 million people. Cases were initially identified through daily contact with the hospital diagnostic microbiology laboratory, and were defined as patients aged ≥18 years with culture-proven melioidosis (isolation of B. pseudomallei from any clinical sample and compatible clinical features). Controls were identified through the hospital computerized admission records, and were defined as patients admitted with non-infectious conditions during the same period (+/−2 weeks, and therefore season), matched for gender, age (+/−5 years), and presence or absence of diabetes mellitus. Patients admitted with infectious conditions were not eligible as controls, as the sensitivity of culture for the diagnosis of melioidosis is not perfect [19]. As a result, culture-negative melioidosis patients were not enrolled as controls. Matching was performed for known predisposing factors (diabetes, gender, age and time of presentation) to control for confounding. Target enrollment numbers were at least 250 cases and 500 controls, which would allow the detection of an approximate odds ratio of 2.0 with 90% power using a two-sided 1% test [19]. Each case and control was interviewed and information collected on specified activities of daily living during the 30 days preceding the onset of symptoms using a standardized study form. Relatives were interviewed if patients were not capable of answering questions. Patients with melioidosis are often severely unwell, and complete data capture via relatives was considered important to avoid the bias associated with exclusion of this group (Text S1). Trained study staff administered the questionnaire. The study was approved by the research ethics committees of Sappasithiprasong Hospital, and the Faculty of Tropical Medicine, Mahidol University. Written informed consent was obtained from all participants. Further details of the study design, definitions of cases and controls, assessment of exposure and statistical methods are provided in the supporting information. A home visit was performed for case and control patients who resided within 100 km of the hospital, a distance limit imposed by the feasibility of travel, sample and data collection in the course of a single day. Five liters was collected from each source of drinking water, and tap water regardless of consumption. If the water was filtered or boiled by the householder before consumption, samples of these were collected for culture. Water samples were transported on the same day to our research laboratory at Sappasithiprasong Hospital and cultured for the presence of B. pseudomallei. In brief, for each 5 liter sample, 1 liter was passed through two 0.45 µm filters and 4 liters was passed through 2.5 g of sterile diatomaceous earth (Celite, World Minerals, USA) [20]. Filters were cultured on Ashdown agar to provide a quantitative bacterial count, and diatomaceous earth was cultured in selective broth (TBSS-C50) [21] to provide a sensitive, qualitative method. Broth was incubated at 40°C in air for 48 hours, after which 10 µl of the upper layer was streaked onto an Ashdown agar plate to achieve single colonies, incubated at 40°C in air and examined every 24 hours for 7 days. In the event that enrichment broth was positive but filters on Ashdown agar were negative, the quantitative count was defined as <1 CFU/L. Identification of bacterial colonies was performed as described previously [22]. Univariable and multivariable conditional logistic regression analyses were performed. All variables that were either statistically significant in univariable analyses (with 0.25 significance level) or that were selected a priori based on current knowledge were included in multivariable analyses. The final multivariable model was developed using a purposeful selection method [23]. Data were analyzed using Stata12.0 (StataCorp, Texas, US). Conditional odds ratios are presented, and all p-values are two-tailed. A total of 414 patients presenting to Sappasithiprasong Hospital with culture confirmed melioidosis between July 2010 and December 2012 were assessed for eligibility (Figure 1). Of these, 84 patients were excluded because they were less than 18 years of age (n = 50), had recurrent melioidosis (n = 33), or declined to participate (n = 1). Of the 330 cases enrolled into the study, two matched controls were identified for each of 226 cases (69%) and one matched control for 61 cases (18%). No matched control could be identified for the remaining 43 cases (13%) who were excluded from further analysis, giving a total of 287 cases and 513 controls. A history of activities of daily living prior to the onset of infective symptoms was obtained from relatives for a total of 92 cases (32%) and 26 controls (5%). A diagnosis of diabetes was more common in patients with melioidosis who were excluded because of failure to find a control, compared with those enrolled as cases (72% v.s. 42%, Table S1). The median age of cases was 54 years (interquartile range 46–64 years, range 18–88 years), 181 (63%) were male, 120 (42%) were diabetic, and 100 (35%) died within 28 days of the admission date. The 513 controls were enrolled from various departments including surgery, orthopedics, general internal medicine and ophthalmology (Table S2). Common causes of illness were cancer (n = 58), bone fracture (n = 35), corneal ulcer (n = 26), cerebrovascular diseases (n = 20), cataract (n = 17), glaucoma (n = 17), calculus formation in the kidney or ureter (n = 16), and intervertebral disc disorder (n = 15). Working in a rice field in the month prior to the onset of infective symptoms was reported in 72% of cases and 48% of controls (Table S3), and almost tripled the odds of acquiring melioidosis (conditional odds ratio [cOR] 2.9, 95% confidence interval [CI] 2.1–4.0). The odds of having melioidosis increased by approximately 10% for each 10 working hours/week increase (cOR 1.1, 95%CI 1.0–1.1), and by approximately 20% for each 10 centimeter increase in depth that the legs were submerged in soil or water (cOR 1.2, 95%CI 1.0–1.3). Conversely, there was a decreased risk associated with wearing long trousers or rubber boots. There was no significant reduction in risk associated with wearing cloth gloves, and wearing rubber gloves was not reported. Washing after working in the rice field was associated with a decreased risk, but washing with water pooled in the rice field was associated with an increased risk of melioidosis (Table S3). Other activities leading to exposure to soil or water were also strongly associated with a risk of melioidosis (cOR 1.8, 95%CI 1.3–2.5 and cOR 2.3, 95%CI 1.7–3.3, respectively). People who walked barefoot everyday had nearly 2.5 times the odds of developing melioidosis compared to those who never walked barefoot (cOR 2.4, 95%CI 1.1–5.5). People who bathed in pond water had 11 times the odds of having melioidosis (cOR 11.1, 95%CI 1.3–92.5). Having an open wound was strongly associated with risk (cOR 2.4, 95%CI 1.4–4.1), and the risk increased if herbal remedies or an organic substance was applied directly onto an open wound (cOR 2.9, 95%CI 1.6–5.3). Eating food contaminated with soil or dust was reported by 40% of cases and 22% of controls (cOR 2.4, 95%CI 1.7–3.3). Table S4 shows the water drinking habits of the 800 participants. Overall, 17% filtered and 13% boiled water before drinking. Following inspection of filtration machines, these were not considered to represent adequate treatment because of poor machine maintenance. Therefore, only bottled and boiled water was considered treated, while unboiled water from wells, boreholes, collected rainwater and tap water was considered untreated. Drinking untreated water was reported by 85% of cases and 72% of controls, and was associated with a doubling in the odds of acquiring melioidosis (cOR 2.3, 95%CI 1.5–3.3). Outdoor exposure to a dust cloud or rain was associated with increased risk (cOR 1.6, 95%CI 1.2–2.2 and cOR 2.9, 95%CI 2.0–4.1, respectively). The use of a protective item (a mask or umbrella) was associated with a lower risk of infection although this did not reach significance (Table S4). Water inhalation of untreated water (accidental choking during drinking or swimming, associated with vigorous coughing) was reported by 23% of cases compared with 9% of controls (cOR 3.0, 95%CI 2.0–4.5). Being an active smoker was associated with increased risk (cOR 2.2, 95%CI 1.4–3.7), but being an ex-smoker was not (Table S4). Consumption of any oral steroid medication was associated with a cOR of 3.2 (95%CI 1.6–6.3). Education beyond primary school was reported by 15% of cases and 26% of controls. A monthly income of greater than 5,000 baht per month was reported by 24% of cases and 37% of controls. Both factors were associated with halving the odds of acquiring melioidosis in the univariable model (Table S4). The final multivariable conditional logistic regression model included 286 cases and 512 controls (1 case and 1 control were excluded because of missing values). The findings of this analysis indicated that activities associated with an increased risk of melioidosis involved all three routes of acquisition. Working in a rice field, other activities leading to exposure to soil or water, eating contaminated food, drinking untreated water, outdoor exposure to rain, an open wound, water inhalation and taking steroids were independent risk factors in the final model (Table 1). There was borderline evidence that active smoking was associated with acquiring melioidosis (cOR 1.5, 95%CI 1.0–2.3, p = 0.069). Home visits and sampling of drinking water was performed in 142/287 cases (49%) and 228/513 controls (44%) who resided within 100 kilometers of the hospital. B. pseudomallei was detected in 12% (10/84) of borehole water samples, 12% (32/273) of tap water samples, and 4% (1/27) of well water samples. B. pseudomallei was not detected in rain water (which is collected into a closed earthenware containers), or bottled water (0/160 and 0/32, respectively) (Table S5). The median quantitative count of B. pseudomallei in culture-positive samples was 1 CFU/L (interquartile range [IQR] <1 to 13; range <1 to 65 CFU/L). Two out of 53 samples of water that had been treated by a household member using filtration were culture positive for B. pseudomallei. Combining the results from the interview and microbiological data, we found that 7% (10/142) of cases and 3% (7/228) of controls drank water from sources that were demonstrated to contain B. pseudomallei (cOR 2.2, 95%CI 0.8–5.8). On the basis of our findings, we propose that protection is required against all three routes of B. pseudomallei acquisition. We recommend that residents and visitors to melioidosis-endemic areas avoid direct contact with soil and water, outdoor exposure to heavy rain or dust clouds, do not consume untreated water, and wash food to be eaten raw using boiled or bottled water (Table 2). If direct contact with soil or water is necessary, we recommend that protective gear such as rubber gloves and boots or waders should be worn. We encourage cessation of smoking (particularly in those with underlying conditions such as diabetes that are known predisposing factors for melioidosis), and discourage the application of herbal remedies or organic substances to wounds. This study has provided evidence to indicate that ingestion and inhalation, together with inoculation, are important routes for the development of melioidosis in Thailand. A range of activities were found to be independently associated with melioidosis, including presumed inoculation during unprotected occupational exposure to soil or environmental water, ingestion by eating contaminated food or drinking untreated water, and inhalation by outdoor exposure to rain. We also confirmed the presence of B. pseudomallei in water obtained from wells and boreholes and from piped water supplies, and recorded that a number of cases had consumed untreated water from these sources prior to presentation with melioidosis. This is the first study to show that ingestion is an important route of human B. pseudomallei infection. Based on data obtained from The Provincial Waterworks Authority of Ubon Ratchathani province, only people living in the town of Ubon Ratchathani (3% of the provincial population) receive piped chlorinated water [24]. Tap water quality control does not include assessment for the presence of B. pseudomallei, which we found in 12% of tap water samples (32/273) and which a number of cases had consumed without adequate treatment. Unlike observations made in Hong Kong (14), none of the collected rainwater samples tested positive for B. pseudomallei. It is customary for the drinking rain water to be collected in large earthenware pots situated close to the house, the water in which can reach temperatures in excess of 40°C. This may explain the negative culture results in our setting, although it is also possible that the bacterial count was below the level of detection of our methodology. We recommend that all non-bottled water should be boiled prior to consumption. Although filtration is an alternative method of water purification, we observed that filters were poorly maintained and detected B. pseudomallei in some filtered water samples. In view of this, we do not recommend the use of filtration. This is also the first evidence to indicate that exposure to rain is an independent risk factor for melioidosis. Exposure to dust clouds was a significant risk on univariable but not multivariable analysis. This is the first study to identify that smoking may be associated with acquiring melioidosis. Smoking could decrease the effectiveness of the local inflammatory response and increase the risk of infection by inhalation. Although microbiological confirmation of aerosolized B. pseudomallei has not been published, this could be due to poor sensitivity of the techniques used or a very low bacterial concentration. In experimental mice, inhalation of only 5 CFU can result in death within a few days [25]. Our study has several limitations. Relatives were asked about activities of daily living when cases or controls were not capable of providing this information, and it is possible that they were not aware of the full spectrum of activities undertaken. It is also possible that cases who were aware of their diagnosis of melioidosis might mention risk factors more readily than controls. This would only be the case if people were knowledgeable about melioidosis, but in a recent survey most Thai people (72%) had not heard of melioidosis, and the remainder had heard of the word but did not know what it meant (unpublished data). The education about melioidosis was given to all participants after the interview. There may be other factors associated with a risk of melioidosis that we failed to examine, and we cannot evaluate the relative risk of a matched variable. The study was not powered to identify risk factors with a relative risk less than 2.0. The criteria specified for matching were stringent and we were unable to find controls for some patients. A diagnosis of diabetes was more common in patients with melioidosis who were excluded because of failure to find a control, compared with those enrolled as cases (72% v.s. 42%). This is because the prevalence of diabetes in patients admitted to the hospital with non-infectious conditions (potential controls) was low, and finding matched controls for diabetic cases was more difficult than that for non-diabetic cases. However, diabetes is the strongest risk factor for melioidosis [8], and matching for diabetes is very important to control the possible confounding effect. We consider it likely that our findings are applicable to similar settings in neighboring Asia but may be less applicable to more distant geographical settings including Australia. Current efforts are being directed toward increasing public awareness and implementing preventive measures for melioidosis in endemic areas, particularly Thailand. A vaccine that protects against B. pseudomallei infection is not available and there is no prospect of one being developed and ready for use in the near future [26]. There is, therefore, every reason to look for alterative solutions to prevent melioidosis, both in people living in regions of the world that are endemic for melioidosis, and visitors (e.g. travelers and military personnel) to these regions. The Ministry of Public Health in Thailand has included melioidosis on a priority list of emerging diseases in Thailand, and public health campaigns for melioidosis prevention based on the knowledge of this work are being developed. These actions will include the implementation of an education programme based on the recommendations provided here, the improvement of infrastructure relating to effective treatment of public water supplies and access to protective clothing, tractors and other machinery to reduce contact time of farmers with soil and environmental water. Further studies are required to model the cost-benefit of guidelines for the prevention of melioidosis, together with their acceptability, up-take and impact on rates of melioidosis.

10.1371/journal.pgen.1000239

Genetic and Linguistic Coevolution in Northern Island Melanesia

Recent studies have detailed a remarkable degree of genetic and linguistic diversity in Northern Island Melanesia. Here we utilize that diversity to examine two models of genetic and linguistic coevolution. The first model predicts that genetic and linguistic correspondences formed following population splits and isolation at the time of early range expansions into the region. The second is analogous to the genetic model of isolation by distance, and it predicts that genetic and linguistic correspondences formed through continuing genetic and linguistic exchange between neighboring populations. We tested the predictions of the two models by comparing observed and simulated patterns of genetic variation, genetic and linguistic trees, and matrices of genetic, linguistic, and geographic distances. The data consist of 751 autosomal microsatellites and 108 structural linguistic features collected from 33 Northern Island Melanesian populations. The results of the tests indicate that linguistic and genetic exchange have erased any evidence of a splitting and isolation process that might have occurred early in the settlement history of the region. The correlation patterns are also inconsistent with the predictions of the isolation by distance coevolutionary process in the larger Northern Island Melanesian region, but there is strong evidence for the process in the rugged interior of the largest island in the region (New Britain). There we found some of the strongest recorded correlations between genetic, linguistic, and geographic distances. We also found that, throughout the region, linguistic features have generally been less likely to diffuse across population boundaries than genes. The results from our study, based on exceptionally fine-grained data, show that local genetic and linguistic exchange are likely to obscure evidence of the early history of a region, and that language barriers do not particularly hinder genetic exchange. In contrast, global patterns may emphasize more ancient demographic events, including population splits associated with the early colonization of major world regions.

The coevolution of genes and languages has been a subject of enduring interest among geneticists and linguists. Progress has been limited by the available data and by the methods employed to compare patterns of genetic and linguistic variation. Here, we use high-quality data and novel methods to test two models of genetic and linguistic coevolution in Northern Island Melanesia, a region known for its complex history and remarkable biological and linguistic diversity. The first model predicts that congruent genetic and linguistic trees formed following serial population splits and isolation that occurred early in the settlement history of the region. The second model emphasizes the role of post-settlement exchange among neighboring groups in determining genetic and linguistic affinities. We rejected both models for the larger region, but found strong evidence for the post-settlement exchange model in the rugged interior of its largest island, where people have maintained close ties to their ancestral lands. The exchange (particularly genetic exchange) has obscured but not completely erased signals of early migrations into Island Melanesia, and such exchange has probably obscured early prehistory within other regions. In contrast, local exchange is less likely to have obscured evidence of population history at larger geographic scales.

In On the Origin of Species [1] and The Descent of Man [2], Darwin suggested that patterns of global biological and linguistic variation might correspond because of their parallel evolution in isolated human groups. Recently, Cavalli-Sforza and colleagues [3]–[5] described a more formal version of this process in which congruent genetic and linguistic trees form as a result of serial population splits and isolation that occur during range expansions into new territories. Anthropologists [e.g., 6],[7] have long been skeptical of this “branching” model of genetic and linguistic coevolution, being wary of conflating biological evolution and cultural change, and because any tight link between the two forms of variation could only occur if past human populations remained isolated following the splits. While it is conceivable that they did so for short periods as they expanded to fill unoccupied regions [8], the prolonged isolation required for congruent evolution seems unlikely. Genetic and linguistic correspondence may also form through a process that is analogous to the genetic model of isolation by distance [9]–[11]. In this process, populations are arrayed evenly over a geographic landscape and neighboring populations exchange both genetic and linguistic features. Genetic and linguistic features may move independently of one another, in which case a correlation will form between genetic and linguistic distances that is purely the result of the underlying correlation of both with geographic distance [3],[12],[13]. Genetic and linguistic features may also move between groups together, in which case their underlying correlation will be independent of geographic distance [13]. Earlier studies have not provided convincing support for either the branching or isolation by distance processes for gene-language coevolution. Cavalli-Sforza and colleagues [5] found some congruence between global gene and language trees, but their informal method of tree comparison was subsequently challenged [14]. With a more formal test, Hunley and colleagues rejected the branching model in Native North America [15] and Native South America [16], though they found some superficial congruence between gene and language trees. The isolation by distance coevolution process has seldom been explicitly tested, but studies in several world regions have either failed to identify genetic and linguistic correlations of any kind or have identified only weak correlations [13], [17]–[25]. Several factors may account for the lack of evidence for gene-language coevolution. First, genes and languages may disperse in very different ways simply because biological transmission is solely vertical but linguistic transmission is both vertical and horizontal [7],[26]. The differing modes of biological vs. linguistic transmission might, in the long term, disrupt correspondences that initially formed through the branching process. Second, differing rates of neutral genetic and linguistic evolution, or differing selective pressures, may prevent the formation of stable genetic and linguistic correspondences [3],[19],[27],[28]. Third, the large geographic scale of many of these studies might prevent the detection of linguistic and genetic correspondences that form at more local levels [16],[29]. Finally, gene-language correspondences could be blurred by the combination of continual group movements and inter-group exchange. The lack of strong support for coevolution may also reflect deficiencies in the methods used to examine linguistic variation. Many studies employ controversial language classifications estimated from cognate data [30]–[32] and estimate linguistic distances simply by counting nodes in these classifications [16], [33]–[36]. Even if a classification is correct, node counting may produce particularly inaccurate distances for long-separated languages [4],[37]. In this study, we compared detailed genetic and linguistic patterns from data collected across a set of particularly diverse populations in the Southwest Pacific. To construct a linguistic classification and estimate linguistic distances, we used data from over 100 structural linguistic features (i.e., aspects of sound systems and grammar) that may avoid some of the limitations associated with cognate data [37]–[39]. These linguistic data, and high-quality autosomal microsatellite data, were used to test predictions of the two coevolutionary models. The datasets come from Northern Island Melanesia, a region well-known for its complex history and remarkable biological and linguistic diversity [40]. The earliest inhabitants of the region arrived at least 40,000 years ago and are thought to have diversified in place in relative isolation from the rest of humanity for the following 30,000 years [41], but there is clear evidence of at least one additional population movement into the region from farther west about 3,300 years ago [42]. The region is geographically complex, with a set of neighboring islands varying in size and ruggedness. As a result, it is a particularly informative region to analyze factors mediating or inhibiting the formation of genetic and linguistic correspondences. The languages of Northern Island Melanesia (NIM) belong to two major groups: Oceanic and Papuan. Oceanic is a major branch of the widespread Austronesian language family that appeared in the region about 3,300 years ago [43], almost certainly associated with the Lapita cultural complex [42],[44]. In NIM, Oceanic languages are found mainly on the smaller offshore islands and along the coasts of the major islands (see Figure 1), though they are spoken in some large island interiors as well. Our sample includes populations that speak 14 of the more than 150 Oceanic languages spoken in the region today. The Papuan languages are likely descendents of languages spoken by people who began arriving in the region more than 40,000 years ago [38],[45]. As a result of their antiquity, they do not form a coherent language family according to conventional historical linguistic criteria, but are rather a residual category of non-Austronesian languages [37]. The Papuan languages in NIM tend to be restricted to the interior highlands of New Britain and Bougainville (Figure 1). Our sample includes populations that speak 9 of the 20 or so Papuan languages spoken in the region today. The standard method of constructing the historical relationships between languages, called the Comparative Method, is a tree-building technique that relies on recognizing sets of words in different languages that are related in meaning and form (cognates) and which show regular sound changes (i.e., shared innovations) demonstrating that they derive from a single ancestral language. Because cognates change relatively rapidly, reconstructions using the Comparative Method cannot generally be made beyond 8,000 years [32]. In NIM, the Papuan languages share no clearly related cognates, possibly because they have been isolated from one another for so long, making the Comparative Method inapplicable for examining their relationships [37],[46],[47]. Recently, Dunn and colleagues [37] proposed the use of abstract structural linguistic features to address the time-depth constraint. These features could provide an independent phylogenetic measure, not related to the lexical evidence. Structural features include syntactic patterns such as constituent order in clauses and noun phrases, paradigmatic structures of pronouns, and the structure of verbal morphology [38]. It is an open question whether structural features are in general more resistant to exchange between different languages, but in contrast to cognate data, the Papuan languages of NIM do show some structural similarity, suggesting that, at least in this case, structural features are more stable [37]. However, structural features are not without their problems, including possible non-independence and homoplasy. To examine their utility and consistency for historical linguistic reconstruction, Dunn and colleagues [37] compared an Oceanic language classification constructed with structural data to one constructed using the Comparative Method. The topologies of the two trees were quite similar. Their structural classification of Papuan languages in NIM also captured the geography of the region fairly well, with its major branches representing the languages of different islands and its more terminal branches joining geographic neighbors within islands. These results were confirmed in subsequent analyses [48],[49] and suggest that structural linguistic features may well produce reliable language trees and linguistic distances estimates, at least in NIM. The branching model predicts that the patterns of linguistic and genetic variation will be treelike, so that for our datasets, the Oceanic- and Papuan-speaking populations will cluster on separate branches of the language and genetic trees, and it also predicts that the topologies within the separate Oceanic and Papuan clusters will be similar in both trees. We tested these predictions by comparing simulated and observed patterns of genetic variation and the topologies of gene and language trees. The isolation by distance model predicts that genetic and linguistic distances will be correlated with one another not because of congruent tree-like evolution but because of ongoing genetic and linguistic exchange between neighboring populations. If genetic and linguistic exchange have occurred independently of one another, the genetic-linguistic distance correlation will lose statistical significance when geographic distance is held constant. If they have moved largely in concert with one another, the genetic-linguistic distance correlation will remain significant when geographic distance is held constant. These predictions were tested using computer simulations, matrix correlation and partial correlation tests, and by examining plots of genetic, linguistic and geographic distances. The detailed genetic and linguistic datasets were recently collected from 33 populations located on the major islands of the Bismarck Archipelago and Bougainville in NIM [38],[39],[50] (Figure 1, Table 1). The genetic data consist of 751 autosomal microsatellite loci drawn from Marshfield Screening sets # 16 and # 54, and the loci were typed in 776 individuals. The linguistic data consist of 108 abstract structural features scored as present or absent in 23 Northern Island Melanesian languages. The features provide broad typological coverage of the known linguistic variation of the region and represent features typically described in a published sketch grammar. Three language groups covered in the genetic survey had not been analyzed (see Table 1), and for them, we substituted data from very closely related languages. The population names are linguistically based. Where genetic data were collected from more than one group in a language area, we added a distinguishing letter (e.g., Anêm-K and Anêm-P for the two Anêm-speaking groups from the Keraiai and Purailing areas). Table 1 lists each population name, island, language affiliation, geographic coordinates, genetic sample size and allelic identity (by which the populations are ordered). Because of recent movements, three populations could not be clearly classified as coastal or interior, and they were therefore classified as “intermediate”. The linguistic and genetic data are available from the authors upon request. Our basic unit of genetic similarity is the allelic identity between individuals, defined as the probability that two alleles of the same locus drawn from two random individuals, either within the same population or from two different populations, are identical [51]. Heat plots were employed to examine the geographic and linguistic patterns of the within- and between-population allelic identities. The last column of Table 1 shows that the Oceanic-speaking populations generally have lower allelic identities than the Papuan-speaking populations. The mtDNA and Y-chromosome data in the same populations have a similar pattern [65]–[67], and the mtDNA and Y-chromosome distances are also much higher between Papuan-speaking populations. This was taken to show the primary action of genetic drift in small isolated groups of Papuan speakers that arrived very early in the region. The Oceanic-speaking populations arrived much more recently, lived in larger groups, and/or were less isolated from one another. However, the allelic identities show an even more pronounced relationship to the coastal/inland residential distinction. Without exception, the coastally-located populations have lower allelic identities than the inland populations. Two of the coastally-located Papuan-speaking groups (Sulka and Kuot) fall in this lower allelic identity coastal grouping, and two of the inland Oceanic-speaking groups (Mamusi and Nakanai-S) fall in the higher allelic identity interior grouping. These linguistic “outlier” populations probably reflect recent population movements between the New Britain coast and interior. As mentioned, Figure 2 shows the presumed history of population splits used as the basis for the simulated branching model. Figure 3A shows the simulated heat plot derived from the simulations of this branching history. The simulated allelic identities in Figure 3A are lowest between the Oceanic and Papuan populations, higher between populations on different islands, higher still between populations within islands, and highest within populations. The level of allelic identity is also uniform between populations at different levels in the hierarchy, reflecting the isolation of branches following ancient population splits. The hierarchical organization and the uniformity of allelic identity within major clusters are fundamental properties of the branching process. Figure 3B shows the observed allelic identity heat plot, with the populations arranged in the same order as in 3A (i.e., clustered first by language group, then by island). The poor fit with the predicted properties of the branching model in 3A is obvious. The Oceanic-Papuan comparisons do not have low and uniform allelic identities. For example, the allelic identities between the Oceanic-speaking Mamusi and Nakanai-S on the one hand and the Papuan-speaking Ata on the other are high compared to the identities between same-language-speaking populations (Figure 3B, circled squares). These are three neighboring groups in the interior of central New Britain. Identities are also high between the four Bougainville populations, even though two of them speak Oceanic languages (Saposa and Teop) and two speak Papuan languages (Aita and Nasioi). Figure 3C shows the same allelic identities arranged simply by island and neighborhood (i.e., not by language). While the fit to the expected pattern is still poor, this reordering shows that allelic identities are relatively high between populations on the same island, and relatively low and uniform between populations on different islands. It also underlines the high identities between the linguistically diverse Mamusi, Nakanai-S, and Ata in the New Britain interior, and between the different language speaking populations on Bougainville. In sum, the observed pattern of allelic identity variation is not consistent with the branching model. It shows that significant genetic exchange has occurred between local populations within islands whether they belong to the same major language group or not, but that genetic exchange between islands may have been relatively restricted for some time. The language and genetic trees in Figure 4 reinforce this scenario. Neither tree completely separates the Oceanic- from the Papuan-speaking populations. Instead, the trees tend to group populations from the same island. The island grouping is particularly strong for the genetic tree, which also clusters geographic neighbors within islands better than the language tree, e.g., it contains the Mamusi/Nakanai-S/Ata cluster from inland New Britain. The language tree does not contain this cluster, but instead groups the geographically distant Ata and Anêm together, both of which speak Papuan languages. Overall, the language tree has a stronger tendency than the genetic tree to group Papuan-speaking populations separately from Oceanic-speaking populations, suggesting that structural linguistic features are more resistant to exchange than genes between the major language groups, or that linguistic exchange has been comparatively more common within the language groups than between them. The results may also reflect relatively low information content in the linguistic data. The bootstrap values of the language tree are low, and the linguistic data contain only 108 features compared to the 6,437 alleles for the microsatellite loci. The results of the model-fitting procedure are shown in Tables 2 and 3. The Λ values for the fitted baseline and language trees are reported in Table 2. Λ for the baseline tree is very high relative to the degrees of freedom, indicating that it does not capture the genetic structure of the NIM populations very well. The lack of fit is also shown by the plot of the observed genetic distances vs. the expected genetic distances for the baseline tree shown in Figure 5A. This result is not surprising given the lack of similarity between the structure-less baseline tree and the topologically complex genetic tree. However, even though the observed and expected genetic distances are not perfectly congruent, the correlation coefficient for the plot is fairly high, indicating that even the baseline tree captures some of the genetic structure of NIM populations. The reason for the high correlation is that the model-fitting procedure estimates the individual population allelic identities fairly accurately for the baseline tree, and this identity is one of the two parameters used to estimate genetic distance. The reason the correlation is not even higher is that the other parameter used to estimate genetic distance is the between-population allelic identity, and, since the baseline tree has only one internal node, the model-fitting procedure estimates only one value for this between-population identity. In the observed data, there are many different values for the between-population identities, causing the discrepant results. Λ is much lower for the fitted language tree than it is for the fitted baseline tree (Table 2). The F-test indicates that the superior fit is statistically significant (Table 3). This superior fit may not be because of any deep congruence between the linguistic and genetic structures, but only because of a few superficial internal nodes (tips) shared by the language and genetic trees (e.g., Aita - Nasioi). To test this possibility, we used the model-fitting method to fit a tree that contained only these shared tips. Λ for this tips-only tree was much lower than it was for the baseline tree (Table 2), but it was still not nearly as low as it was for the complete language tree. This result suggests that the language tree captures more than just some superficial aspects of the genetic structure. Figure 5B is the plot of the observed genetic distances vs. the expected genetic distances based on the language tree. The relatively high squared correlation for the plot also confirms that the language tree captures more of the genetic structure than the baseline tree. There are, however, several clear outlier points in the plot, and Λ is still very high for the language tree relative to its degrees of freedom, meaning that its fit is far from perfect. The lower plot in Figure 5B shows that of all of the groups, the Kol contribute most to the high Λ of the language tree. Λ for the language tree reconstructed after removing the Kol is 5,777 compared to 8,593 for the full language tree (see Table 4). The plot shows that the Kol are generally closer to neighboring populations than the language tree would predict, reflecting the greater tendency of the genetic tree to group neighboring populations on the same island. For example, in the genetic tree, the Kol, who speak a Papuan language, cluster with the nearby Oceanic-speaking Mengen, whereas in the language tree, they cluster with other Papuan-speaking populations who are more distant geographically. These different tree patterns confirm the greater tendency of genes to move between Papuan- and Oceanic-speaking populations than structural linguistic features. The contributions of other populations to the lack of correspondence between the observed and expected genetic distances are shown in Table 4. Methods described in Text S1 were used to identify four additional populations that contributed disproportionately to the lack of correspondence. Three of these four outliers also involved neighboring Oceanic- and Papuan-speaking populations that clustered together in the genetic tree but not in the language tree. Λ for the language tree lacking the Kol and these other four outlier populations is 1,992 (Table 2), which represents a dramatic reduction compared to the full 23 population language tree (F-test p<0.0001, Table 4). The revised 18-population language tree is shown in Figure 4C, and the plot of the observed genetic distances vs. the expected genetic distances for this revised tree is shown in Figure 5C. The very high squared correlation coefficient in 5C confirms its superior fit relative to the full 23-population language tree. However, Λ is still high for this revised language tree, indicating that even it does not fully capture the genetic structure of NIM populations. The lower plot in Figure 5C shows that the Mali are the largest outlier in this comparison. The Mali are closer to other New Britain populations in the genetic tree, regardless of the language they speak, than they are in the language tree. Overall, the results show the pervasive pattern of closer genetic than linguistic proximity between populations on the same island. Figure 6 shows the heat plot for the simulated isolation by distance model allelic identities. The simulated identities are highest within populations and then fall off steadily as the geographic distance between populations increases (indicated by the change in color moving horizontally or vertically away from the diagonal). There is some hint of this fall-off for some populations in the observed matrix, but, overall, the observed pattern diverges from the predicted. In the simulations, the populations are arrayed next to one another in a linear stepping stone pattern, but the 33 sampled NIM populations are not located next to one another in a simple linear fashion. However, the lack of congruence between the heat plots is not because of this difference. Isolation by distance predicts decreasing allelic identity with increasing geographic distance regardless of the actual sampling locations, and this pattern does not occur for the observed allelic identities. This conclusion is supported by additional simulations reported in the last section of Text S1. Table 5 shows the matrix correlation results. Waypoints did not improve the correlations, so we report only the results for the direct great circle distances. The correlations listed for the full sample are suggestive of an isolation by distance coevolutionary process in the region, but several of the correlations are not statistically significant at the multiple tests-adjusted level. However, when the correlation coefficients are calculated for localized geographic and linguistic comparisons, many of them increase in magnitude and cross the threshold of statistical significance. Figure 7 shows plots of the genetic, linguistic and geographic correlations and highlights the localized geographic and linguistic comparisons. Figure 7A and 7B shows the genetic-geographic distance correlation, with different localized sets highlighted. In Figure 7A, the interior and coastal sets are highlighted in red and blue. The lack of mixing of the colors suggests that there has been limited genetic exchange between island interiors and coasts. Figure 7B highlights the Papuan and Oceanic sets. The mixing of the colors shows that Papuan and Oceanic-speaking populations have exchanged genes. This exchange has occurred primarily between the interior Oceanic-speaking Mamusi and Nakanai-S with interior Papuan-speaking populations, and between the coastal Papuan-speaking Kuot and Sulka with coastal Oceanic-speaking populations. Table 6 shows how the Oceanic and Papuan genetic-geographic distance correlations improve when these four outlier populations are removed. Plots 7C and 7D show the linguistic-geographic distance correlations, with the different sets highlighted as before. As one might expect for the linguistic correlations, the coastal and interior strata are less clearly distinguished than the Oceanic and Papuan strata. This is again consistent with the argument that there has been little linguistic exchange between Oceanic and Papuan languages where they occur in neighboring groups (e.g., the four outliers). The poorer distinction for the interior and coastal strata is caused by these outliers. Table 6 shows that the interior and coastal linguistic-geographic distance correlations improve dramatically when the four outliers are removed. Plots 7E and 7F show the genetic-linguistic distance correlations with similar highlighting. They suggest that any linguistic-genetic correlation is driven solely by the Papuan-speaking populations, but as Table 6 shows, when the four outliers are removed, the correlation for the Oceanic comparisons increases dramatically and becomes statistically significant. These results provide further support for the conclusion that linguistic exchange has been comparatively limited between Oceanic- and Papuan-speaking populations where they overlap geographically. The plots also show that for any given geographic distance, the interior/Papuan-speaking populations have higher genetic and linguistic distances among them than do the coastal/Oceanic-speaking populations. The correlation coefficients are also generally larger between interior/Papuan populations than they are between coastal/Oceanic populations. This distinction is the result of the comparatively restricted movement in the rugged highland interiors [68], coupled with the much longer tenure of Papuan-speaking populations. The correlations are particularly high in the New Britain interior (Table 5, blue squares in Figure 7). The genetic-geographic distance correlation is 0.94 (p<0.0000), which, to our knowledge, is the highest such correlation reported for any region worldwide. The high linguistic-geographic (0.59) and genetic-linguistic correlations (0.67) for the New Britain interior are also significant at a high level of probability, but the partial correlation, in which geographic distance is held constant, is not. As mentioned, the correlation and partial correlation patterns are consistent with an isolation by distance process where genetic and linguistic exchange have occurred largely independently of one another. The results on the New Britain coast suggest a separate isolation by distance pattern there as well. All of the correlation coefficients there are high, but only the genetic-linguistic distance correlation is statistically significant (Table 5). The p-values for the other correlations are low (genetic-geographic = 0.0066; linguistic-geographic = 0.0099), but they are above the multiple tests adjusted significance level (p = 0.0024). When the two Papuan-speaking populations are removed from the coastal New Britain sample, the correlations increase in magnitude and the partial correlation also crosses the threshold of statistical significance (Table 6), despite the fact that the sample contains only six populations. We suspect that a larger sample would reveal an even more robust isolation by distance pattern on the coast and on the other islands in the region. The tests of the branching model in Northern Island Melanesia show that genetic and linguistic exchange between local populations has erased evidence that may have once existed for a branching process there. Genes have tended to move freely between nearby populations, regardless of the languages they speak. On the other hand, structural linguistic exchange has been particularly limited between neighboring Oceanic and Papuan languages. In these instances, the Oceanic-speaking populations have become very similar genetically to their Papuan-speaking neighbors (the best example of this is the high allelic identity between the Ata, Mamusi and Nakanai-S shown in the heat plot in Figure 3B). Although an alternate explanation for this situation is that Oceanic languages have simply been adopted by formerly Papuan-speaking groups [c.f., 50], this now appears most unlikely, because the general tendency in Northern Island Melanesia is for neighboring populations, regardless of their languages, to become genetically similar (other clear examples are the Kove/Anêm and also the Kuot and their neighbors on New Ireland). Previous analyses of the autosomal microsatellites [50] as well as Y-chromosome data [67] suggest that Papuan-speaking groups, who entered NIM first and expanded there long before the arrival of the early Oceanic-speakers, have contributed much more genetically to Oceanic-speaking groups than vice versa over the last three millennia. The genetic, linguistic and geographic distance correlations are consistent with an isolation by distance coevolutionary process in the interior of the largest island in the region, New Britain. For the correlations to be so strong, the patterns of ancestral residence and local migration must have persisted for a considerable period. It is remarkable that the patterns have persisted in the face of the destabilizing influence of European contact [42],[69] and also of displacements caused by major volcanic eruptions [70]. One reason for the persistence is the continuing ties of the people to their land. Even today, most people in our sample remain in small villages and continue to farm their local gardens, or they maintain dual residences there and in larger population centers [68]. The matrix correlation results show that studies of prehistory and coevolution at the regional level must take into account the geographic and linguistic heterogeneity of a region, since ecological and sociocultural variation are likely to strongly influence biological and cultural patterning. Parallels to the heterogeneity found in NIM probably exist, in many cases unidentified, in every major world region and in various locations within each region [71]–[74]. Our results are apparently at odds with the studies of Cavalli-Sforza et al. [4],[5] that identified a strong correspondence between global gene and language trees. One explanation is that global patterns are more likely to emphasize ancient demographic events, such as population splits associated with the colonization of major world regions, while local patterns will generally emphasize more recent demographic events. Wilkins and Marlowe [75], for example, showed that genetic data collected from local populations are more likely to reveal recent changes in migration associated with the rise of agriculture than data collected from a global sample. However, it is also possible that the differences between the global results of Cavalli-Sforza and colleagues and ours are not so pronounced. In their studies, they identified several instances of disagreement between the language and genetic trees caused by different patterns of genetic and linguistic exchange and language shift, so the global pattern may also reflect, to a substantial degree, the types of local population interactions we identified in NIM. The structural linguistic data used in this study [48],[76] have recently come under attack, both in terms of their quality and what they capture (i.e., just more recent contacts, or mainly ancient language splits). Our results certainly suggest that structural features may well be more resistant to dynamics of diffusion than genes, and therefore likely contain considerable information about language splits as well as language contacts. The structural features may also be more resistant to diffusion than lexical items, making them more suitable than cognate data for examining linguistic splits in NIM, and probably in other regions as well. Dunn et al. [48],[49] have addressed the criticisms of data quality in detail, but they acknowledge that there are some problems. The linguistic features are not completely independent of one another, the data may contain substantial homoplasy [37],[49], and for the NIM dataset, there are 8.7% missing data. Despite these shortcomings, the significant correlations between the linguistic, genetic, and geographic distances certainly show that the structural linguistic data contain important information about the relationships between NIM languages. In particular, the separation of the Oceanic and Papuan groupings in the plots of linguistic vs. geographic distances (Figure 7D) suggests that, even if the data only reveal linguistic contacts, the contacts have been stronger between populations within each major language group than between populations in different language groups [see also 39]. Another relevant point is that the linguistic data and methods typically used in studies of coevolution have usually been of comparatively poor quality. To illustrate the higher quality of our structural linguistic dataset, we employed the commonly used method of node counting to estimate linguistic distances between NIM languages in a classification constructed using the Ethnologue (http://www.ethnologue.com/), and we then examined the correlation between these distances and the genetic and geographic distances. None of the correlations were statistically significant. If not for the structural linguistic data, we would have failed to identify any linguistic relationship to genetic or geographic patterns at all. The limitations of these sorts of data are not restricted to Northern Island Melanesia. Hunley et al. [16] tested the branching and isolation by distance models in South America, where linguistic divergence has been occurring for a considerably shorter period. They examined the fit of language and gene trees constructed from linguistic cognate data and mtDNA sequences, and identified correspondences only between the tips of the language and genetic trees, i.e., only between very recently diverged groups. In the current study, the language and genetic structures shared more than just a few superficial similarities, clearly suggesting the results are indicative of more ancient relationships. Studies of coevolution will clearly benefit greatly from using similar structural linguistic datasets. The highly informative nature of the genetic data available to us (i.e., the 751 microsatellite loci with 6,437 different alleles) also undoubtedly led to our finding of comparatively high correlations in our various analyses. Many recent studies have used mitochondrial d-loop data and Y-chromosome data to investigate genetic and linguistic correspondence in various world regions [15], [16], [20], [77]–[81], but these data are comparatively uninformative. The Y-chromosome data typically contain only a few loci, and the mitochondrial d-loop data are plagued by homoplasy, which confounds the construction of genetic classifications and limits the accuracy of genetic distance estimation [82]. In an earlier publication, information content issues prevented us from successfully fitting our structural language tree to mtDNA and Y-chromosome data collected from most of the same populations [66]. The mitochondrial d-loop data were able to recreate some of the same correlation patterns we found using the autosomal microsatellite data, but the correlations were always weaker than those we have reported here. The implications of our results for broader issues in Pacific prehistory are important but must be interpreted carefully. While our results provide little support for the branching model in Northern Island Melanesia, this is different from arguing that branching did not occur in very early periods there, or elsewhere in the Pacific, and it does not mean that our microsatellite data lack important information about the deeper prehistory of the entire region. For example, two contrasting scenarios for the origins of the Polynesians have persisted in recent Pacific prehistory debates, and they bear a very close relationship to the two models examined in this paper. The first has been called the phylogenetic model [83],[84], which is essentially identical to the branching model, and the second, called a reticulate model [85], is essentially identical to the isolation by distance model [see also rebuttal by 86]. A number of mixed models, perhaps more realistic than either of these, have also been proposed [87]. Bellwood [83] also argued that phylogenetic differentiation should be expected to occur primarily during or shortly after the early rapid range expansions in new territories, while the reticulate model, which stresses a continuous and relatively uncoordinated shifting of linguistic, cultural, and biological boundaries through assimilation, intermarriage, borrowing, and diffusion, may become more evident in subsequent periods. The genetic data have been interpreted to support several of these Polynesian origin scenarios. Some have indicated that a clear phylogenetic signal exists between Taiwan Aborigines and Polynesians, with little intermixture taking place in Near Oceania, while other datasets have been interpreted to suggest heavy intermixture with, or major contributions from, Near Oceanic and Wallacean populations [50], [65], [88]–[93]. While the results of our present study are broadly inconsistent with phylogenetic models in Northern Island Melanesia, our group did identify in the same microsatellite data a small but clear genetic coancestry between certain Taiwanese populations and Oceanic-speaking groups in Island Melanesia, as well as a much stronger Taiwan Aboriginal signal in Polynesia, indicating that intermixture over the past 3,000 years has not completely erased genetic signals of early Oceanic origins in either NIM or Polynesia [50]. The more comprehensive nature of our genetic and linguistic coverage in this region has now allowed a more complete, if complex, picture of ancient population dynamics to emerge.

10.1371/journal.ppat.1001330

SLO-1-Channels of Parasitic Nematodes Reconstitute Locomotor Behaviour and Emodepside Sensitivity in Caenorhabditis elegans slo-1 Loss of Function Mutants

The calcium-gated potassium channel SLO-1 in Caenorhabditis elegans was recently identified as key component for action of emodepside, a new anthelmintic drug with broad spectrum activity. In this study we identified orthologues of slo-1 in Ancylostoma caninum, Cooperia oncophora, and Haemonchus contortus, all important parasitic nematodes in veterinary medicine. Furthermore, functional analyses of these slo-1 orthologues were performed using heterologous expression in C. elegans. We expressed A. caninum and C. oncophora slo-1 in the emodepside-resistant genetic background of the slo-1 loss-of-function mutant NM1968 slo-1(js379). Transformants expressing A. caninum slo-1 from C. elegans slo-1 promoter were highly susceptible (compared to the fully emodepside-resistant slo-1(js379)) and showed no significant difference in their emodepside susceptibility compared to wild-type C. elegans (p = 0.831). Therefore, the SLO-1 channels of A. caninum and C. elegans appear to be completely functionally interchangeable in terms of emodepside sensitivity. Furthermore, we tested the ability of the 5′ flanking regions of A. caninum and C. oncophora slo-1 to drive expression of SLO-1 in C. elegans and confirmed functionality of the putative promoters in this heterologous system. For all transgenic lines tested, expression of either native C. elegans slo-1 or the parasite-derived orthologue rescued emodepside sensitivity in slo-1(js379) and the locomotor phenotype of increased reversal frequency confirming the reconstitution of SLO-1 function in the locomotor circuits. A potent mammalian SLO-1 channel inhibitor, penitrem A, showed emodepside antagonising effects in A. caninum and C. elegans. The study combined the investigation of new anthelmintic targets from parasitic nematodes and experimental use of the respective target genes in C. elegans, therefore closing the gap between research approaches using model nematodes and those using target organisms. Considering the still scarcely advanced techniques for genetic engineering of parasitic nematodes, the presented method provides an excellent opportunity for examining the pharmacofunction of anthelmintic targets derived from parasitic nematodes.

In parasitic nematodes, experiments at the molecular level are currently not feasible, since in vitro culture and genetic engineering are still in their infancy. In the present study we chose the model organism Caenorhabditis elegans not only as a mere expression system for genes from parasitic nematodes, but used the transformants to examine the functionality of the expressed proteins for mediating anthelmintic effects in vivo. The results of our experiments confirmed that SLO-1 channels mediate the activity of the new anthelmintic drug emodepside and showed that the mode of action is conserved through several nematode species. The chosen method allowed us to examine the functionality of proteins from parasitic nematodes in a defined genetic background. Notably, expression of the parasitic nematode gene in anthelmintic-resistant C. elegans completely restored drug susceptibility. As C. elegans is highly tractable to molecular genetic and pharmacological approaches, the generation of lines expressing the parasite drug target will greatly facilitate structure-function analysis of the interaction between emodepside and ion channels with direct relevance to its anthelmintic properties. In a broader context, the demonstration of C. elegans as a heterologous expression system for functional analysis of parasite proteins further strengthens this as a model for anthelmintic studies.

Infections with parasitic nematodes heavily affect the well-being, health, and productivity of humans and animals worldwide [1]. Since the 1960s several broad spectrum anthelmintic compounds have been available. During decades of frequent and sometimes inappropriate use of these anthelmintics, resistance to currently available drugs has developed and is an increasing problem in parasitic nematodes, especially in livestock [2]. In human medicine, where mass anthelmintic treatment programmes were employed during recent years in countries with endemic gastro-intestinal nematode infections, there is also growing concern regarding anthelmintic resistance, and several reports of treatment failure were published during recent years [3]-[6]. In livestock non-chemical worm control procedures such as pasture management, feeding, and breeding are being tested, but they are cost- and labour-intensive and often not practical [7]. In parasites of companion animals, resistance is less common. Nevertheless, populations of the canine hookworm Ancylostoma caninum were recently reported to show high degrees of resistance to pyrantel [8]. Therefore, the need for anthelmintic compounds with new modes of action is urgent. Recently, three groups of anthelmintic compounds employing new mechanisms of action have been introduced. The oxindole alkaloid paraherquamide was described first in 1981 [9]. Paraherquamide and its derivative 2-deoxoparaherquamide (derquantel) are anthelmintically active by blocking acetylcholine receptors and therefore inhibiting neurotransmission [10], [11]. Derquantel has been launched in combination with abamectin as a drench for sheep in New Zealand in 2010. The combination showed high efficacies against field infections with strongyles in sheep [12]. The second group, comprising the amino-acetonitrile derivatives (AAD), was recently reported to act mainly through the nicotinic acetylcholine receptor ACR-23. This receptor is not present in mammals and is not involved in the action of levamisole, ivermectin, benomyl, dimethyl-4-phenylpiperazinium, and aldicarb. The derivative AAD 1470 was shown to have good efficacy against different species of gastrointestinal nematodes [13]. The first available AAD on the market was AAD 1566 (monepantel), which has been launched as a sheep drench. The third group are the cyclooctadepsipeptides. The parent compound of this class is PF1022A, which was discovered as a fermentation product of the fungus Mycelia sterilia [14]. The semi-synthetic derivative emodepside has a broad spectrum of anthelmintic activity [15], indicating that the mechanism of action might be conserved throughout nematode clades. Emodepside and PF1022A were also shown to be effective against anthelmintic-resistant populations of the sheep nematode Haemonchus contortus and the cattle nematode Cooperia oncophora [16]. Commercially, emodepside was first available as a spot-on preparation in combination with praziquantel for cats. Recently, emodepside has been launched as a tablet for dogs, also in combination with praziquantel. In Caenorhabditis elegans, emodepside potently inhibits locomotion, egg-laying, and pharyngeal pumping [17]. Previous studies identified nematode latrophilin (LAT-1) as a target for emodepside [18], [19], but LAT-1 is not required for the inhibitory effects of emodepside on locomotion [19], [20]. Indeed, a mutagenesis screen revealed the large conductance calcium-gated potassium channel SLO-1 as a key component for the mechanisms of action of emodepside [20]. SLO-1 channels are regulated by voltage and by intracellular concentration of calcium ions [21]–[24]. They were first identified in experiments with the slowpoke mutant of Drosophila melanogaster, which exhibits abnormal locomotory behaviour and decreased flight ability [22], [24]. In C. elegans, SLO-1 was previously shown to control excitatory neurotransmitter release. It is expressed in the nerve ring and in the body wall muscle [21]. The slo-1 loss-of-function mutants show a characteristic locomotor phenotype consisting of an increase in locomotor reversal frequency [20], [21]. The mutagenesis screen for emodepside-resistant C. elegans mentioned above revealed nine independent lines that were able to move and to reproduce on agar plates with an emodepside concentration as high as 1 µM, a concentration that immobilises wild-type C. elegans. All nine lines fell into a single complementation group that mapped closely to the slo-1 locus on chromosome V. Four of them were sequenced and showed mutations in the slo-1 locus predicted to lead to a reduced or abolished function of the channel. In locomotion assays, the slo-1 mutants had different degrees of resistance to emodepside. Reduction-of-function mutants showed reduced susceptibility to emodepside whilst loss-of-function mutants were not at all inhibited after exposure to emodepside [20]. The putative slo-1 null allele reference strain NM1968 slo-1(js379)V [21] was also highly resistant to emodepside. The expression of slo-1 in slo-1(js379) animals from the pan-neuronal promoter snb-1 [21], [25] and the muscle cell-specific promoter myo-3 [21], [26], either in combination or separately, restored emodepside susceptibility to different degrees [20]. In this study, we identified slo-1 orthologues in H. contortus, A. caninum and C. oncophora. The slo-1 coding sequences of A. caninum and C. oncophora were subsequently expressed in the emodepside-resistant C. elegans strain slo-1(js379) to investigate their ability to rescue emodepside susceptibility of slo-1 loss-of-function mutants. Furthermore, we compared the ability of different C. elegans promoters as well as the slo-1 5′ flanking regions of A. caninum and C. oncophora to drive expression of slo-1 in slo-1 loss-of-function mutants and examined the locomotor phenotype as well as the degree of emodepside susceptibility in the transformants. Finally, we showed that penitrem A, an inhibitor of mammalian SLO-channels [27], is able to antagonise the paralysing effect of emodepside on infective A. caninum larvae as well as on the locomotion of young C. elegans adults in a dose-dependent manner. The animals used for the maintenance of the parasitic nematode strains were helminth-free prior to infection. All animals used in this study were handled in strict accordance with good animal practice as defined by the relevant national and local animal welfare bodies, and all animal work was approved by the appropriate committee. Calves were infected with approx. 30,000 C. oncophora third-stage larvae, and sheep were infected with 6,000-8,000 infective larvae of H. contortus. After 21 to 30 days, the animals were necropsied, and the small intestine or the abomasum, respectively, was removed. The worms were either washed off or picked directly from the mucosa. Dogs were infected with 400-500 infective A. caninum larvae. After reaching patency, the dogs were treated with 4 mg/kg arecoline. The subsequently deposited faeces were collected and sieved through a 100 µm mesh sieve. The adult A. caninum were picked directly from the sieve. The recovered parasites were sorted according to sex, washed in 0.9% NaCl solution and subsequently in DEPC-treated water. The worms were frozen at -80°C in sterile GIT buffer (4 M guanidine; 0.1 M Tris, pH 7.5; 1% β-mercapto-ethanol). All experiments with animals were performed in strict accordance to the German law for animal welfare (Tierschutzgesetz) and with the approval of the respective local authority, the Niedersächsisches Landesamt für Verbraucherschutz und Lebensmittelsicherheit (LAVES) under the reference numbers 01A38, 01A48 and 06A395. All efforts were made to avoid and minimize suffering of the animals. Total RNA was isolated using Trizol reagent (Invitrogen, Karlsruhe, Germany), according to the manufacturer's recommendations. For cDNA synthesis and Rapid Amplification of cDNA Ends (RACE), the BD SMART RACE cDNA Amplification Kit (Clontech, St-Germain-en-Laye, France) was used following the manual. For isolation of genomic DNA, a standard phenol-chloroform method was used [28]. The GenomeWalker Universal Kit (Clontech) was used to amplify the putative slo-1 promoter regions of A. caninum and C. oncophora. Primers to amplify the putative C. elegans slo-1 promoter region were designed based on the sequence of YAC clone Y51A2D (GenBank Acc. No. AL021497). The first primers for fragments of the slo-1 coding sequence of H. contortus were designed based on EST (Expressed Sequence Tag) sequences revealed by the H. contortus EST Basic Local Alignment Search Tool (BLAST) of the Wellcome Trust Sanger Institute server, using the C. elegans slo-1 sequence (GenBank Acc. No. NM_001029089, accordant with slo-1 splice variant b) as template. The same primers were used to amplify a partial slo-1 coding sequence of C. oncophora. Primers for A. caninum slo-1 were designed based on a partial coding sequence detected in the whole genome shotgun library AIAAGSS 001 using the BLAST application of the Nematode Net [29]. Sequences of primers are given as supporting data, Table S1. PCR products were cloned into the pCR4-TOPO vector, using the TOPO TA Cloning Kit (Invitrogen) or into the pCR-Blunt vector, using the Zero Blunt PCR Cloning Kit (Invitrogen) and transformed into TOP10 Escherichia coli cells (Invitrogen). Vectors containing full-length slo-1 coding sequences were transformed into JM109 E. coli cells (Stratagene, La Jolla, CA, USA). Plasmid DNA preparation was performed using the NucleoSpin Plasmid Kit or the NucleoBond AX 100 Kit (Macherey and Nagel, Düren, Germany). To introduce the required restriction sites, PCR was performed using primers carrying the restriction sites (refer to supporting data, Table S1) with a plasmid, containing the respective full-length sequence, or with cDNA as template. The PCR products were cloned as described above and subcloned into the respective expression vector using T4 DNA ligase (Invitrogen). The basis of the expression plasmids was pBK3.1 [20], [21] (kindly provided by Lawrence Salkoff, Washington University School of Medicine, St. Louis), carrying the C. elegans slo-1 coding sequence downstream of the C. elegans snb-1 promoter, leading to neuron specific expression [21], [25]. The expression plasmids were propagated in XL10-Gold Ultracompetent Cells (Stratagene). The coding sequences of A. caninum and C. oncophora slo-1, respectively, were cloned between the XbaI and BamHI restriction sites within the pBK3.1, thus replacing the C. elegans slo-1 coding sequence. To test the functionality of the slo-1 coding sequences to be analysed in as natural an expression pattern as possible, constructs were built carrying the slo-1 coding sequences downstream of the C. elegans slo-1 promoter. To achieve a construct carrying the C. elegans slo-1 promoter and the C. elegans slo-1 coding sequence, a ligation was set up with three DNA fragments, since the coding sequence of C. elegans slo-1 contained an additional HindIII restriction site: 1) the vector backbone of pBK3.1 digested with BamHI/HindIII, 2) the promoter sequence (HindIII/partial XbaI digest), and 3) the coding sequence of pBK3.1 digested with XbaI/BamHI. The plasmids carrying the parasite slo-1 coding sequences downstream of the C. elegans slo-1 promoter were derived by modifying pBK3.1 constructs which already carried the slo-1 coding sequence of the parasitic nematodes. The snb-1 promoter was excised and replaced by the C. elegans slo-1 promoter sequence using the HindIII and XbaI restriction sites flanking the promoter region. For this purpose, the plasmid carrying the C. elegans slo-1 promoter sequence had to be digested completely with HindIII, but only partially with XbaI, since the promoter sequence had an additional XbaI restriction site. The plasmid carrying the C. oncophora slo-1 coding sequence downstream of the C. elegans slo-1 promoter was not used for functional analysis but as a starting point to construct a plasmid with the C. oncophora slo-1 coding sequence downstream of the C. oncophora slo-1 promoter region (see below). To test the functionality of the parasite promoter sequences, the parasite promoters were used to drive expression of the respective parasite slo-1 in C. elegans. For this purpose, the putative promoters were inserted between the HindIII and XbaI restriction sites in the modified pBK3.1 as described above, replacing the C. elegans slo-1 promoter. Due to additional HindIII and XbaI restriction sites in the C. oncophora slo-1 promoter sequence, the plasmid construction was done by blunt end ligation. All plasmids used for expression experiments in C. elegans were sequenced by custom sequencing (SeqLab Laboratories Goettingen, Germany), ensuring that the coding sequences and the ligation sites were intact. For an overview of constructs used for the transformation experiments refer to Table S2 (supporting data). Sequences were analysed using the Sci Ed Central Align Plus 5 software, version 5.04 (Scientific and Educational Software; Cary, NC, USA), and the NCBI BLAST [30]. The predicted SLO-1 amino acid sequences and selected sequences of potassium channels of other species revealed by the BLAST search were aligned using the ClustalX2 [31] software package with default settings except that the alignment parameters were changed to BLOSUM. ClustalX2 calculates scores as percentages of the number of identities in the best alignment divided by the number of residues compared, excluding gap positions. The alignment constructed was manually edited and, after elimination of all positions containing gaps, a phylogenetic tree was built using bootstrap analysis (1000 replicates) and the Neighbour Joining method by the Mega4 software package [32] using the default Poisson correction model for multiple substitutions at the same site and assuming homogenous substitution rates for all sites. The analysis of the putative promoter regions was performed using the Genome2Promoter and MatInspector software packages (Genomatix, Munich, Germany). The putative slo-1 promoters of the three nematode species were compared by alignments using the BLAST bl2seq (filter inactivated for low complexity regions) [30]. The C. elegans strains were grown on nematode growth medium (NGM) agar plates containing 50 µl of E. coli (OP50) overnight culture as a food source at 20°C or room temperature. Strains employed were Bristol N2 and NM1968 slo-1(js379)V [21]. The latter contains a mutation within the transmembrane region of the SLO-1 channel which leads to the early termination of the protein and is therefore predicted to encode a non-functional ion-channel. Thus, slo-1(js379) animals show a slo-1 null phenotype due to a translational knock-out. Emodepside was prepared as five different stock solutions (2 mM to 200 nM) in ethanol. 0.5 ml of stock solution was added to 100 ml NGM agar after autoclaving and at a temperature of 42°C. Accordingly, control plates were prepared containing 0.5 ml ethanol per 100 ml NGM agar, leading to a final concentration of 85 mM ethanol. This ethanol concentration does not significantly impair C. elegans locomotion [33], [34]. All plates were seeded with 50 µl E. coli OP50. In some of the experiments, agar plates also contained 1 µM penitrem A (Enzo Life Sciences, Lörrach, Germany) in 28 mM DMSO (final concentration) or only the DMSO vehicle as control. For the body bend counts, experiments were performed in the absence of E. coli, i.e. on plain un-seeded NGM plates. Hermaphrodite C. elegans were transformed by microinjection of plasmids into the gonads. Transformation with the differentially modified pBK3.1 plasmids (30 ng/µl) was accomplished by co-injecting the pPD118.33 (Addgene plasmid: 1596; 50 ng/µl) GFP-expressing marker. Successful transformation was determined by identification of the selection marker. For the behavioural and pharmacological analysis only worms carrying the selection marker were used as they were predicted to express the transgene of interest as well. To confirm the transcription of the introduced slo-1 coding sequences in transgenic worms, RT-PCR was performed. Total RNA was isolated from a bulk of worms using the TriFast method (PeqLab), and contaminating DNA was removed by a DNase I treatment. 1 µg of total RNA was used for cDNA synthesis (RevertAid First Strand cDNA Synthesis Kit, Fermentas, St.Leon-Rot, Germany), and a -RT control (lacking the Reverse Transcriptase) was performed for each sample. PCR was performed using 1 µl of template in a 25 µl setup (High Fidelity PCR Enzyme Mix, Fermentas, St.Leon-Rot, Germany). Each cDNA was analyzed with all test primer pairs. For primer sequences refer to Table S3 (supporting data). The C. elegans slo-1 knockout strain NM1968 slo-1(js379)V shows an abnormal behaviour of locomotion in terms of increased reversals, i.e. to stop and reverse direction [21]. To analyse the impact of the heterologously expressed SLO-1 on this behaviour, the number of reversals was counted for all lines. Therefore, a total of 10 L4 stage larvae of each line were selected and placed on an OP50 seeded NGM agar plate. After 24 hours the young adult worms were transferred separately away from the bacterial lawn for one minute to allow removal of bacteria adherent to the worm. Then the worm was put on an un-seeded NGM-agar plate, and, after one minute of acclimatisation, the reversals were counted for 3 min. Numbers of body bends per minute and of reversals in different C. elegans lines were compared using One-Way-ANOVA and individual lines were then compared with Tukey's post hoc test implemented in GraphPad Prism. A p-value <0.05 was considered as statistically significant. For locomotion assays L4 stage larvae of stable lines (at least F2 generation) were used. For each strain (transformed and control strains) ten worms were analysed for each concentration of emodepside (1 nM, 100 nM, 1 µM, and 10 µM, and in case of expression from the parasite promoters also 100 µM) and the ethanol control, respectively. The assays were repeated using two independent stable lines, so that in total 20 worms for each construct and concentration were analysed. The experiments were not repeated for the worms expressing the A. caninum slo-1 from the C. elegans slo-1 and snb-1 promoters due to the lack of sufficient numbers of transformants. The setup for the locomotion assay was as follows: L4 stage larvae of N2, slo-1(js379) and the transformed slo-1(js379) lines were transferred to NGM plates containing E. coli OP50 and either different concentrations of emodepside (10 µM to 1 nM) or ethanol vehicle. Worms were maintained on emodepside or control plates for 24 hours at 20°C and locomotion was examined afterwards. For that purpose, worms were transferred for one minute to plain un-seeded NGM plates to remove bacteria. Subsequently, the worms were transferred to a fresh un-seeded NGM plate and, after one minute, body bends were counted for each worm for another minute. A single body bend is defined as one full sinusoidal movement of the worm. For analysis of a transformant line at a certain concentration of emodepside, N2 and slo-1(js379) worms were tested on the same day as parallel controls. For statistical comparisons, four-parameter logistic concentration-response-curves with variable slope were fitted using GraphPad Prism 5.0 after plotting the log10 of the emodepside concentration vs. the relative body bend activity at that concentration (percentage of maximum number of body bends in each data set). Bottom values were always constrained to greater than 0. Top values, Hill slopes and EC50 were not constrained. Calculation of means and 90% confidence intervals and statistical tests for differences in 1) EC50, 2) bottom or 3) all four parameters (top, bottom, Hill slope, and EC50) were also done using GraphPad Prism. For slo-1(js379), linear regression including testing for linearity and a significance test for a slope differing from 0 was performed with the same software. Statistical significance was assumed for p<0.05. Infective larvae of A. caninum (non-exsheathed) were incubated for 24 h at room temperature in 1×PBS buffer containing either penitrem A or the vehicle dimethylsulfoxide (DMSO) in combination with different concentrations of emodepside or the respective vehicle ethanol. Penitrem A (500 µM stock solution in DMSO) was used in a final concentration of 1 µM penitrem A, resulting in a final DMSO concentration of 28 mM (0.2%). Emodepside (1 mM stock solution in ethanol) was used in final concentrations of 1 µM, 5 µM, and 10 µM, respectively. The final ethanol concentration was 170 mM (1%) in these experiments. The concentration of the vehicles was adjusted to the same final concentration in all setups by adding DMSO and/or ethanol. Furthermore, one control was performed without vehicles to estimate the impact of the vehicles. After 24 h, the larvae were used for a modified larval migration inhibition test (LMIT), similar to that described by Demeler et al. [35]. Briefly, 1800 µl containing approximately 100 larvae was pipetted onto precision sieves (mesh size 20 µm) in a 24 well plate. The volume of 1800 µl was sufficient that the sieves were hanging in the liquid and motile larvae were able to penetrate the meshes. After further incubation for 24 h at room temperature, the sieves were removed and the bottom side was carefully rinsed with approximately 300 µl 1×PBS to gather any adherent larvae. Thus, this well contained the migrated larvae. Then, the sieves were turned upside down, and each sieve was rinsed by carefully pipetting 1000 µl 1×PBS through the sieve meshes and collecting the buffer in a so far empty well to recover the non-migrated larvae. For each setup, migrated and non-migrated larvae were counted individually, and the percentage of migrated larvae was calculated as follows: Each setup was performed in triplicate, and the whole experiment was performed three times in total. The results were compared to each other using a One-Way-ANOVA followed by a Tukey's post hoc test (GraphPad Prism) A p-value <0.05 was considered to be statistically significant. Nucleotide sequences: C. elegans YAC clone Y51A2D containing the putative slo-1 promoter region (AL021497); C. elegans slo-1 splice variant b (NM_001029089); partial coding sequence of A. caninum slo-1 (CW974961); partial coding sequence of H. contortus slo-1 (genome version 20060127: contigs >004261, >0045106, >001213, and >057289); A. caninum slo-1 complete coding sequence (EU828635); C. oncophora slo-1 complete coding sequence (EF494185); H. contortus slo-1 complete coding sequence (EF494184); Proteins sequences: C. elegans SLO-1a (AAL28102); C. elegans SLO-1b (AAL28103); C. elegans SLO-1c (AAL28104); C. briggsae hypothetical protein CBG12923 (XP_001675579.1); A. caninum SLO-1 (EU 828635); C. oncophora SLO-1 (EF494185); H. contortus SLO-1 (EF494184); Ixodes scapularis putative calcium-activated potassium channel (EEC10339.1); Cancer borealis calcium-activated potassium channel (AAZ80093.4); Manduca sexta calcium-activated potassium channel alpha subunit (AAT44358.1); Pediculus humanus corporis putative calcium-activated potassium channel alpha subunit (EEB13088.1); Drosophila melanogaster slowpoke, isoform P (NP_001014652.1); Tribolium castaneum predicted protein similar to slowpoke CG10693-PQ (XP_968651.2); Aplysia californica high conductance calcium-activated potassium channel (AAR27959.1); Xenopus laevis potassium large conductance calcium-activated channel, subfamily M, alpha member 1 (NP_001079159.1); Danio rerio novel calcium activated potassium channel (CAX13266.1); Trachemys scripta calcium-activated potassium channel (AAC41281.1); Gallus gallus calcium-activated potassium channel alpha subunit (AAC35370.1); Monodelphis domestica predicted protein similar to large conductance calcium-activated potassium channel subfamily M alpha member 1 (XP_001367795.1); Mus musculus mSlo (AAA39746.1); Homo sapiens potassium large conductance calcium-activated channel, subfamily M, alpha member 1, isoform CRA_d (EAW54600.1); Bos taurus BK potassium ion channel isoform C (AAK54354.1); Canis familiaris calcium-activated K+ channel, subfamily M subunit alpha-1 (Q28265.2); Strongylocentrotus. purpuratus predicted protein similar to calcium-activated potassium channel alpha subunit (XP_783726.2) The search of the Wellcome Trust Sanger Institute H. contortus EST BLAST server using C. elegans slo-1 as template revealed four short fragments of 83 – 150 bp (from the contigs 004261 (two fragments) and 0045106 and 001213) within the coding sequence and a 599 bp fragment containing the last twenty codons of the coding sequence, the stop codon, and part of the 3′ untranslated region (UTR) (from contig 057289). Based on these sequences, primers were designed to amplify the partial coding sequence of H. contortus slo-1. The same primers were used to amplify the respective fragment of C. oncophora slo-1. A partial coding sequence of A. caninum slo-1 was detected in a whole genome shotgun library fragment (GenBank Acc. No.: CW974961) and primers were designed, according to that sequence. RACE-PCR completed the coding sequences and the 5′ and 3′ UTRs. The full-length coding sequences were 3309 bp (EU828635; 1103 predicted amino acids) for A. caninum slo-1, 3333 bp (EF494185; 1111 predicted amino acids) for C. oncophora slo-1, and 3315 bp (EF494184; 1105 predicted amino acids) for H. contortus. GC-contents of the coding sequences were 47.1 – 51.9%, molecular weight and isoelectric point of the proteins were predicted to be 125.02 - 125.88 kDa and 5.77-5.80, respectively. None of the 5′ UTR sequences contained a spliced leader 1 (SL1) sequence. Compared to the predicted sequences of A. caninum and H. contortus SLO-1, C. oncophora SLO-1 had six additional NH2-terminal amino acids. The identities of the nucleotide sequences within the coding region were 80% between A. caninum and C. oncophora, 79% between A. caninum and H. contortus, and 85% between C. oncophora and H. contortus. Based on the predicted amino acid sequences, the identities were 95% between A. caninum and C. oncophora, 95% between A. caninum and H. contortus, and 98% between C. oncophora and H. contortus. The splice variants slo-1a, b, and c of the C. elegans slo-1 cDNA coding sequence were all 73% identical with A. caninum, C. oncophora, and H. contortus slo-1, respectively. Based on predicted amino acid sequences, the identities were 87-88% between C. elegans SLO-1 (splice variants SLO-1 a, b, and c) and all three newly identified parasitic nematode SLO-1 sequences. A phylogenetic tree (Figure 1) shows the relationship of selected SLO channels on the protein level from several animal genera and species. All known nematode SLO-1 orthologues group together: however, within this nematode SLO-1 group, the predicted SLO-1 proteins of the parasitic nematodes cluster in a group distinct from the non-parasitic nematodes C. elegans and Caenorhabditis briggsae. Analysing EST and genome databases for putative SLO-1 orthologues in other nematodes, fragments of coding sequences were identified for a range of species, including Brugia malayi, Trichinella spiralis, Strongyloides ratti, and Trichuris muris (data not shown). As these sequences were either incomplete or of insufficient quality, they were not included in the phylogenetic analysis. The amplified putative promoter sequences covered approximately 3 kb upstream of the start codon (A. caninum slo-1 promoter 2997 bp, C. oncophora slo-1 promoter 3421 bp, C. elegans slo-1 promoter 3084 bp). The 5′ UTR of A. caninum slo-1 included an intron, which was not present in C. oncophora slo-1. The sequence analysis identified no known promoter elements or transcription factor binding sites in any of the slo-1 promoters employed. Just a few consensus sequences were detected, which might indicate RNA polymerase binding sites. No TATA or CAAT elements could be detected. Comparison of the putative slo-1 promoters of the three nematode species by alignments did not reveal any significant similarities. Comparing the sequences with the respective putative promoter regions of C. briggsae and Caenorhabditis remanei slo-1 (3000 bp upstream of the start codon) also revealed no significant similarities (data not shown). In cDNA samples of all analysed transgenic lines, transcripts of the respective expression construct were detected. The primer pairs targeting the expression constructs containing slo-1 coding sequences of the other species gave no amplicon in PCR. In cDNA samples of the slo-1 null mutant strain slo-1(js379) – representing the genetic background of the transgenic strains – and in the Bristol N2 wild-type strain, no transcript of any expression construct could be detected, confirming the authenticity of the PCR results for the transgenic lines. To ensure that the absence of specific PCR products was not due to insufficient RNA-isolation or cDNA-synthesis, a control primer pair was used and gave a PCR product in all analysed cDNA samples (data not shown). In all transgenic strains expressing functional slo-1 from one of the expression constructs, the phenotype of increased reversals exhibited by the slo-1 null mutant strain slo-1(js379) was completely rescued as the rate of reversals was statistically not significantly different (p = 0.87 in a one-way ANOVA) from that observed in Bristol N2 wild-type worms (Figure 2A) but significantly (p<0.001) lower than in mutant slo-1(js379). It was previously shown that C. elegans slo-1 loss-of-function mutants are highly resistant to the inhibition of locomotion behaviour by emodepside [20]. In our study, we expressed slo-1 orthologues of the parasitic nematodes A. caninum and C. oncophora in the emodepside-resistant slo-1(js379) genetic background in order to rescue sensitivity to emodepside and to investigate involvement of these proteins in the mode of action of emodepside. Locomotion was determined by measuring body bends of the worms in the absence of food. By transformation of C. elegans slo-1(js379), stable transgenic lines were obtained expressing 1) A. caninum slo-1 from the neuronal snb-1 promoter, 2) C. oncophora slo-1 from the snb-1 promoter, 3) A. caninum slo-1 from the C. elegans slo-1 promoter, 4) C. elegans slo-1 from the C. elegans slo-1 promoter 5) A. caninum slo-1 from the A. caninum slo-1 promoter, and 6) C. oncophora slo-1 from the C. oncophora slo-1 promoter (an overview is given in supporting data, Table S2). Transgenic lines were analysed for their susceptibility to emodepside. Their locomotion behaviour was compared to that of the wild-type strain Bristol N2 and to that of the loss-of-function mutant slo-1(js379) over a wide range of emodepside concentrations and concentration-response-curves were fitted to the data to allow statistical comparisons. Animals of all analysed lines showed a comparable basic locomotion, measured as body bends per minute, on the control plates without emodepside (Figure 2B). Locomotion of the slo-1(js379) mutant strain was not at all affected by any of the emodepside concentrations tested (Figure 3) as revealed by concentration-response-curves that are not significantly different from a straight line with slope 0 (p = 0.91). In contrast, locomotion of the Bristol N2 wild-type strain was concentration-dependently inhibited by emodepside. The EC50 for this effect varied between 127.3 nM and 144.2 nM (Table 1) in this set of experiments. At the highest concentration used (10 µM), the Bristol N2 wild-type worms were nearly completely paralysed or dead. The transgenic worms expressing A. caninum (Figure 3A) or C. oncophora (Figure 3B) slo-1 from the snb-1 promoter showed significantly different concentration-response-curves (p<0.0001) with increased susceptibility to emodepside compared to the slo-1(js379) mutant but were not as susceptible as Bristol N2 wild-type animals. Although the EC50 values were not altered, the lines expressing parasitic nematode slo-1 from the snb-1 promoter showed significantly increased bottom values (refer to Table 1) indicating that even extremely high emodepside concentrations were not able to cause complete paralysis. At the highest concentration of 10 µM, worms of the transgenic lines were still able to show nearly half the body bend activity as the ethanol control, while the wild-type worms were almost completely immobilised. Expression of A. caninum slo-1 from the C. elegans slo-1 promoter (Figure 3C) showed a marked susceptibility to emodepside that was equivalent to N2 wild-type worms: worms expressing the parasite slo-1 from the C. elegans slo-1 promoter in slo-1(js379) animals fully restored susceptibility to emodepside as revealed by the absence of any significant differences in top and bottom values, Hill slope or EC50 (Table 1). A comparable effect was observed when the emodepside susceptibility of the slo-1(js379) mutant was rescued through the C. elegans slo-1 from the C. elegans slo-1 promoter (Figure 3D and Table 1). Transgenic worms expressing A. caninum or C. oncophora slo-1 from the respective A. caninum or C. oncophora slo-1 promoter showed increased susceptibility to emodepside compared to the slo-1(js379) mutant as well (Figure 4). However, the observed concentration-dependent effects were not as marked as seen for the transgenic worms expressing slo-1 from the C. elegans slo-1 promoter. The lines expressing A. caninum or C. oncophora slo-1 from the A. caninum or C. oncophora slo-1 promoter showed a 62- and 72-fold higher EC50 than the wild type worms. EC50 and 95% confidence intervals and significance levels for comparisons are given in Table 2. In all experiments, the susceptibility appeared not only as a simple reduction of the number of body bends, but also as an altered pattern of movement, since the worms seemed to be stiffened in the forepart of their body. None of the transformed strains showed coiling as was observed previously at 1 µM emodepside after transformation of slo-1(js379) with pBK3.1, the plasmid containing the C. elegans slo-1 coding sequence and the snb-1 promoter [20]. To conclude, a total functional rescue of the wild-type phenotype regarding the inhibitory effect of emodepside on locomotion was achieved with heterologous slo-1 genes expressed under the control of the C. elegans slo-1 promoter in C. elegans, as revealed by our statistical analysis showing no significant differences in the four parameters of the logistic concentration-response curve. These findings provide evidence that the slo-1 genes cloned from A. caninum and C. oncophora are functional, as well as structural, orthologues of C. elegans slo-1. The vehicles DMSO and ethanol in the concentrations used here did not have any statistically significant effect on the migration of A. caninum larvae through 20 µm meshes. In the presence of emodepside, a concentration-dependent inhibition of migration was observed (Figure 5A). The additional presence of 1 µM penitrem A clearly antagonized the effect of emodepside on migration. The difference in migration of larvae incubated with emodepside either with or without penitrem A was statistically highly significant with p-values of <0.001 for all emodepside concentrations tested. Body-bend assays with C. elegans worms produced highly similar results (Figure 5B). In the present study, we identified orthologues of the Ca2+-activated K+ (BK) channel C. elegans slo-1 in the parasitic nematodes H. contortus, C. oncophora, and A. caninum. Subsequently, we analysed the ability of A. caninum and C. oncophora slo-1 to functionally rescue emodepside susceptibility in slo-1 knockout mutant C. elegans. The examination of anthelmintic targets of parasitic nematodes is of great importance, since, in contrast to their orthologues in C. elegans, they are the direct targets for drugs used in veterinary and human medicine. Unfortunately, the parasitic stages of the nematodes, which mainly represent the target population for drugs, cannot be examined easily, and especially functional analysis of gene products in parasitic nematodes is usually not feasible. Up to now, parasitic nematodes cannot be maintained in in vitro cultures for their complete life cycle. Therefore, although it has been occasionally successful in some species such as filaria or Strongyloides spp. [36]-[38], genetic engineering, i.e. expression or knockout of genes, in parasitic nematodes is still an unsolved problem [39]. RNAi experiments in parasitic nematodes had very variable outcomes, depending on the target gene, the delivery method, and the species tested [40]–[44]. This might be due to the fact that parasitic nematodes seem to lack orthologues for a transporter responsible for the systemic spread of RNAi in C. elegans, facilitating the accessibility of cells for RNAi in the latter [45]. Therefore, the use of C. elegans as a model and expression system is currently one of the most powerful tools for the functional analysis of genes of parasitic nematodes, especially if the genes have close orthologues in C. elegans [39]. One approach is the overexpression of a parasitic nematode gene in C. elegans with a wild-type genetic background for the respective gene. This approach can be used if the knockout mutant phenotype for the gene to be examined is lethal or not evident. Couthier et al. [46] expressed the H. contortus transcription factor elt-2 ectopically in C. elegans and found that this expression had similar effects as ectopic expression of the endogenous elt-2. Another experimental setup is exemplified by the experiments described here for slo-1, namely the rescue of the C. elegans loss-of-function mutant by expression of the homologous gene of a parasitic nematode. For that purpose, the mutant should have a clear phenotype and the effects of the rescue should be measurable. Similar experiments examining functionality of parasitic nematode genes in C. elegans have been performed previously. In the study of Kwa et al. [47], ß-tubulin (isotype 1) of H. contortus was expressed in benzimidazole-resistant mutants of C. elegans (TU1054 ben-1(u462)). The benzimidazole-resistance of the ben-1(u462) C. elegans mutants is due to a mutation disrupting the ß-tubulin gene ben-1 [47], [48]. The mutants showed a significantly higher EC50 with regard to the benzimidazole thiabendazole in a larval development inhibition assay compared to the wild-type Bristol N2. In contrast to the resistant ben-1 mutants, H. contortus ß-tubulin expressing ben-1(u462) mutants showed a lower EC50, though not as low as the wild-type larvae [47]. Thus, a total rescue of the wild-type phenotype regarding the effect of thiabendazole on egg-development was not achieved. The effect of expression of H. contortus ß-tubulin on susceptibility of adult ben-1(u462) worms to benzimidazoles has not been reported. Cook et al. [49] expressed the α-subunit of the glutamate-gated chloride channel (GluClα) of H. contortus in C. elegans GluClα mutants, which show a lower sensitivity to ivermectin and a decreased duration of forward movement. Here, a rescue of the wild-type phenotype in respect of the natural locomotion behaviour was observed. However, the effect of ivermectin was not described. Another study showed that expression of the transcription factor of the FOXO/FKH family of Strongyloides stercoralis in C. elegans daf-16 mutants was able to rescue the dauer-forming capability [50]. Very recently, the acetylcholinesterase of the plant-parasitic nematode Globodera pallida was expressed in C. elegans and was shown to functionally rescue the phenotype of the C. elegans double mutant ace-1;ace-2 [51]. In another recent study, Gillan et al. expressed the heat-shock protein 90 (hsp-90) of H. contortus and Brugia pahangi in C. elegans. While expression of H. contortus hsp-90 in C. elegans daf-21 heat shock protein 90 mutants (C. elegans daf-21(nr2081)) partially rescued the phenotype of the mutant, the B. pahangi hsp-90 failed to do so, although the construct was transcribed and translated [52]. The great advantage of C. elegans as an expression system for parasite genes is that posttranslational modifications of recombinantly expressed proteins, which can be necessary for the biological function of the protein, are more conserved between nematodes than between nematodes and standard expression systems [53]. In our experiments, we did not use the recombinantly expressed protein, but the whole transgenic organism to measure the influence of the heterologously expressed proteins on susceptibility to emodepside. The expression of A. caninum slo-1 and C. oncophora slo-1 in the emodepside-resistant C. elegans slo-1(js379) mutant fully rescued the phenotype of worm locomotion: transgenic worms no longer showed increased reversal movement. These findings indicate a complete functional rescue and at least sufficient expression to restore SLO-1 dependent signalling to wild-type levels in the locomotor circuits. The subsequent pharmacological analysis showed that the transgenesis also rescued the phenotypic behaviour of the animals in terms of inhibited locomotion activity in the presence of emodepside. Animals expressing parasitic nematode slo-1 driven by the snb-1 promoter responded to emodepside in a manner qualitatively similar to wild-type animals, although the inhibition of locomotion was significantly weaker than that of the wild-type worms as determined by counting body bends. No complete paralysis was obtained even with an emodepside concentration that completely paralysed the wild-type animals. This phenotype might reflect the fact that expression of slo-1 was only reconstituted in one of its normal compartments, neuronal cells, whereas it was absent from another compartment, the muscle cells. The findings with parasite slo-1 under control of the snb-1 promoter are similar to previous experiments, in which C. elegans slo-1(js379) mutants were rescued by expression of endogenous slo-1 from the snb-1 promoter [20]. Interestingly, the coiled paralysis of the transgenic C. elegans upon exposure to emodepside observed in earlier experiments with the snb-1 promoter driven expression and also with the combination of snb-1 and myo-3 promoter driven expression of the endogenous slo-1 was not observed in our experiments. The coiling previously observed for slo-1(js379) animals expressing slo-1 from the snb-1 promoter in the presence of 1 µM emodepside was supposed to be due to overexpression or to ectopic expression in neurons usually not expressing slo-1 [20]. The most likely reason for the absence of this phenotype in the present study is the altered plasmid used for transformation. Although the linkage between the promoter and the slo-1 coding sequence was identical for the plasmids carrying the parasite slo-1 and the parental pBK3.1 plasmid used in the previous study, the downstream coding sequence may have influenced the level of expression. While the earlier study by Guest et al. [20] aimed to determine whether the mediation of the effects of emodepside is controlled via a neuronal or a muscular pathway, we were now interested in whether the parasitic nematode SLO-1 channels were also able to act as key components for emodepside action. Therefore, we chose to express the parasite slo-1 not only from the neuronal promoter snb-1, which showed a stronger effect in that former study than the muscle-specific promoter myo-3, but also from the putative endogenous C. elegans slo-1 promoter to achieve a pattern resembling the natural expression pattern, and from the putative parasite slo-1 promoters to test their ability to drive expression in C. elegans. The constructs were designed to be comparable to the pBK3.1 construct, which carries the snb-1 promoter sequence, 2987 bp in size. The transgenic animals expressing parasitic nematode slo-1 driven by the C. elegans slo-1 promoter were highly susceptible to emodepside, and since their susceptibility was statistically not different from the susceptibility of the wild-type worms, we considered this phenotype as a full rescue. For some drug targets, such as β-tubulin, a single nucleotide polymorphism can abolish their functionality as a drug target [54]. Therefore, the overall sequence identity between parasite and C. elegans SLO-1 orthologues of 87-88% per se did not ensure a conserved function with regard to emodepside. In the study of Gillan et al. the H. contortus hsp-90 sequence showed 88% identity with the C. elegans orthologue, but its expression rescued the mutant phenotype only partially [52]. The finding that expression of slo-1 from different nematode species restored the susceptibility to emodepside in the slo-1(js379) mutants emphasises that the mode of action is most likely conserved between these species. Generally, SLO-1 channels belong to a relatively conserved ion channel family [23]. This was also confirmed by our BLAST search results, which identified channels in very distantly related genera. The expression of parasite slo-1 under control of the putative slo-1 promoters from A. caninum and C. oncophora aimed to examine the capacity of the parasite-derived promoters to drive expression of the coding sequence of their natural gene within the heterologous background of C. elegans. The transformants showed only partial rescue of emodepside susceptibility. However, in contrast to the lines with snb-1 driven expression, the lines expressing slo-1 from the putative slo-1 promoters of A. caninum and C. oncophora, respectively, did not show increased bottom values. In these experiments the rescued lines clearly had a higher EC50, suggesting that the expression pattern might have been qualitatively restored but that expression levels in general were too low. Since, as was shown in our experiments using the C. elegans slo-1 promoter, the coding sequences of parasite slo-1 appeared to be able to rescue the resistant phenotype completely, the reason for the incomplete rescue is most likely the promoter. The lack of TATA or CAAT elements which we observed for the slo-1 promoters from A. caninum, C. oncophora as well as from C. elegans is consistent with other studies on nematode promoters and strengthens the assumption that the absence of these elements is a characteristic feature of protein-coding genes of this phylum [26], [55]-[59]. Transcriptional regulatory elements can be located at large distances from the start codon, within intron sequences, and also within the 3′ UTR. Furthermore, expression can be influenced by post-transcriptional regulation, e.g. by microRNAs [60]. Nevertheless, most common reporter gene constructs only use upstream intergenic sequence, and it is recommended to include as much of the upstream sequence as possible. Even so, all phenotypes obtained with such reporter constructs must be interpreted with caution as they may not necessarily reflect the endogenous gene expression pattern [61]. We conclude from the present experiments that the parasite slo-1 promoters drive expression in a functionally appropriate pattern, as the parasite slo-1 expressed from the respective parasite slo-1 promoter qualitatively restored emodepside susceptibility in resistant slo-1(js379) C. elegans. The fact that the emodepside susceptibility of the transformants was significantly lower than in transformants expressing parasite slo-1 from the C. elegans snb-1 or slo-1 promoter, respectively, in turn indicates that the expression pattern obtained with the parasite promoters is not equivalent to that obtained with the C. elegans promoters used in this study. Interestingly, the phenotype of slo-1(js379) C. elegans concerning increased reversals was completely rescued by the parasite slo-1 expressed from the parasite slo-1 promoters. The fact that the rescue regarding emodepside susceptibility was less complete again strengthens the assumption that the spatial pattern or some other characteristics of expression such as expression levels in certain cell types might not have been sufficient to completely fill in the function of the wild-type slo-1 expression. An approach to use the slo-1 promoters of C. elegans, A. caninum, and C. oncophora to express GFP for localisation studies in C. elegans was only partially successful. Within the offspring of the microinjected hermaphrodites only single worms were found exhibiting GFP-expression. Fluorescence was detected as punctate structures in the pharynx region of the transformed animals, indicating expression in pharyngeal neurons, furthermore in the neuron-rich anal region of the worms and in locations consistent with expression in the nerve cords (data not shown). For the C. elegans slo-1 promoter reporter construct, GFP expression was observed in body wall muscle cells within the forepart of the body (data not shown). However, due to the restricted number of observations these investigations thus far do not allow to draw final conclusions and therefore need to be further pursued. The hypothesis of the functional involvement of SLO-1 in the mechanism of action of emodepside in parasites was further supported by a series of experiments with emodepside and penitrem A. Penitrem A is a tremorgenic mycotoxin known to completely suppress bovine BK channel currents at a concentration of 10 nM [27]. It has also been used as a BK channel inhibitor in a study on muscle fibres of the liver fluke Fasciola hepatica [62]. The concentration in those experiments was 10 µM, but the authors do not report, whether they tested other concentrations. In our experiments, we used penitrem A in a concentration of 1 µM and showed its ability to antagonise the paralysing effect of up to 10 µM emodepside on A. caninum larvae and young C. elegans adults. While lower concentrations of penitrem A (10 nM and 100 nM, data not shown) did not impair the effect of 10 µM emodepside, 1 µM penitrem A antagonised emodepside at all emodepside concentrations analyzed. The need for higher penitrem A concentrations than in experiments with cultured mammalian cells might be explained by a lower accessibility of the target in the intact nematode larvae, e.g. due to the cuticula – at least for the non-feeding A. caninum third-stage larvae. Currently there are no data available on whether penitrem A is indeed also a specific BK channel inhibitor in nematodes and on what penitrem A concentrations are needed for this inhibition. However, the present data show antagonistic effects of emodepside and penitrem A, indicating that both drugs target the same pathway requiring SLO-1. To conclude, the examination of the actual role of SLO-1 in the signalling of emodepside is still under way. The prevailing view is that emodepside directly or indirectly activates SLO-1 [20], [63]. In contrast to the effects of emodepside on pharyngeal pumping, the effects of emodepside on locomotion are not mediated by the previously described latrophilin-activating pathway [19]. The current model includes latrophilin and SLO-1 for the pharyngeal neurons and SLO-1 but not latrophilin for the body wall musculature [63]. The presented study aimed primarily to test the hypothesis that the mechanism of action of emodepside as far as currently known is conserved in nematodes. Our results are based on functional expression of A. caninum and C. oncophora slo-1 in C. elegans driven by different promoters and demonstrate the ability of the parasitic SLO-1 to act in the mode of action of emodepside. These results are further supported by the experiments with the BK channel inhibitor penitrem A antagonising emodepside. Therefore the current findings suggest that the mode of action is conserved across the three nematode species, providing an important example for functional analysis of the role of individual parasite genes as targets for anthelmintic drugs. Furthermore, these experiments emphasise the potency of C. elegans as an authentic functional model for expression of parasitic nematode genes – at least from clade V – and the subsequent physiological examination of drug/target interactions. Experiments of this type close the gap between research in model organisms and in parasitologically relevant target species. The results presented in this work open new perspectives on functional analysis of parasitic nematode genes in general and in particular allow further analysis of putative targets for emodepside and the elucidation of the mode of action in detail. Transgenic worms from the present study expressing C. elegans slo-1 driven by the C. elegans slo-1 promoter have already been used as a control in a parallel study regarding the expression of the human slo-1 orthologue kcnma1 in C. elegans (Crisford et al., submitted). Another possible application of the system is its use to analyse the impact of certain mutations on emodepside susceptibility, for instance single nucleotide polymorphisms (SNP), identified in resistant populations and suspected to contribute to resistance development. In the long-term, these methods might also enhance development of new anthelmintically active agents.

10.1371/journal.pgen.1005569

LINE-1 Mediated Insertion into Poc1a (Protein of Centriole 1 A) Causes Growth Insufficiency and Male Infertility in Mice

Skeletal dysplasias are a common, genetically heterogeneous cause of short stature that can result from disruptions in many cellular processes. We report the identification of the lesion responsible for skeletal dysplasia and male infertility in the spontaneous, recessive mouse mutant chagun. We determined that Poc1a, encoding protein of the centriole 1a, is disrupted by the insertion of a processed Cenpw cDNA, which is flanked by target site duplications, suggestive of a LINE-1 retrotransposon-mediated event. Mutant fibroblasts have impaired cilia formation and multipolar spindles. Male infertility is caused by defective spermatogenesis early in meiosis and progressive germ cell loss. Spermatogonial stem cell transplantation studies revealed that Poc1a is essential for normal function of both Sertoli cells and germ cells. The proliferative zone of the growth plate is small and disorganized because chondrocytes fail to re-align after cell division and undergo increased apoptosis. Poc1a and several other genes associated with centrosome function can affect the skeleton and lead to skeletal dysplasias and primordial dwarfisms. This mouse mutant reveals how centrosome dysfunction contributes to defects in skeletal growth and male infertility.

Severe short stature in humans has many causes including defects in skeletal and hormonal growth regulation. Primordial dwarfisms are congenital growth defects that involve mutations in genes for DNA repair, DNA replication, splicing of U12 introns, and centrosome dynamics. We discovered that the spontaneous, dwarf mouse mutant chagun is caused by loss-of-function of the gene Poc1a, which encodes protein of the centriole 1A. Mutants exhibit disproportionate dwarfism, and bones formed by endochondral and intramembranous ossification processes are affected. The epiphyseal growth plates of their long bones are disorganized. The chagun males are infertile due to Sertoli cell dysfunction and the failure of germ cells to complete meiosis. The chagun mouse is a model for human dwarfism and provides insight into the mechanism whereby this centriolar protein affects bone growth and spermatogenesis.

Normal adult stature in humans is achieved primarily through regulation of long bone growth, which occurs through endochondral ossification [1,2]. This process begins with the differentiation of mesenchymal stem cells into chondrocytes in regions of the body where skeletal elements will eventually reside. Hypertrophic differentiation of these chondrocytes directs the vascularization of the forming element, allowing osteoblasts to enter and commence mineralization of the cartilage-based template. Pools of cells at the epiphyses of the long bones retain their cartilage identity as a means to secure the progressive addition of bone matrix throughout the period of skeletal growth. These structures are the epiphyseal growth plates. Tight control of chondrocyte proliferation and terminal hypertrophic differentiation allows new bone tissue to replace terminally differentiated chondrocytes in a spatially and temporally regulated manner, ensuring the proper growth of the skeletal elements and the individual overall. Disruption of these processes can lead to skeletal dysplasias. The growth plate maintains a highly ordered architecture to carry out its function and is divided into three distinct zones: the resting, the proliferative, and the hypertrophic zone [1,2]. Perturbation of growth plate organization can result in profound growth defects in mice and humans [3–5]. Resting chondrocytes have a low rate of cell proliferation, and the plane of cell division is not controlled. In contrast, rapidly dividing chondrocytes in the proliferative zone undergo directed cytokinesis orthogonal to the main direction of bone growth, and then intercalate to form distinctive columns of disc-shaped chondrocytes [3,6]. Major orchestrators of growth plate architecture include the primary cilium [3], the WNT-planar cell polarity [6], TGFβ, BMP, and hedgehog signaling pathways [4]. Primordial dwarfisms are a subset of growth insufficiency disorders that are classified as forms of skeletal dysplasia [7,8]. Primordial dwarfism typically involves a proportionate reduction in longitudinal growth that commences early in fetal life [7]. Primordial dwarfisms result in profound reductions in height, and many are associated with head and/or facial dysmorphisms [1]. The primordial dwarfisms include Seckel Syndrome, Microcephalic Osteodysplastic Primordial Dwarfism I-III (MOPD I-III), and Meier-Gorlin Syndrome (MGS) [7]. These disorders interfere with processes that are intimately connected to the cell cycle rather than an endocrine disturbance. Common pathways that are disrupted in primordial dwarfism patients include: DNA repair [9–12], DNA replication [13–16], centrosome dynamics [12–15,17–23], and splicing of minor, or U12 introns, in mRNA [24,25]. We report discovery of the gene responsible for autosomal recessive skeletal dysplasia and male infertility in chagun mice [26]. The mutation is a LINE-1 mediated insertion of a processed Cenpw cDNA into Poc1a, which encodes protein of the centriole 1A. This disruption causes exon skipping, and the mutant protein lacks a highly conserved WD-40 domain necessary for normal spindle pole and cilia formation. POC1A is necessary for normal growth plate architecture, craniofacial development, and spermatogenesis. The mechanism underlying the growth defect is cell death and failure to regulate polarized cell division and intercalation at the growth plate. This animal model clarifies the molecular basis for the clinical features in human patients with POC1A mutations and predicts that male patients will be azoospermic [21,22]. The chagun mutation arose on the DBA/2J strain and was mapped to a 6 Mb region of Chr 9 using recombinant progeny from an F1 x F1 intercross with CAST/EiJ [26]. Initial studies revealed an autosomal recessive disproportionate growth disorder, likely caused by disorganization of the proliferative zone of the growth plate. We used dermestid beetle preparations to characterize the mutant bones (Fig 1). The chagun mutant skull has a slightly domed and disproportionate shape. The skull and mandible are foreshortened along the length, but the widths are similar to wild type. Incisor and molar teeth appear normal. We used micro-CT to assess bone architecture and mineralization (Fig 1). Representative cross sections and 3D volumetric renderings of 4-month-old male mice were selected based on medial bone mineral content and are shown with matched display settings. We found no differences in cortical mineralization, but the trabecular bone was thinner in both vertebrae and metaphyseal long bones of the mutants. The chagun femora are obviously short, with undulating cortical bone and lateral third trochanter irregularities. Disorganization at the distal femoral growth plate is apparent, likely leading to the noticeable lack of trabecular bone in the metaphysis. Metaphyseal periosteal shaping appears to be significantly impacted by the mutation, with abnormal anterior-posterior and medial-lateral shaping. Similar findings were observed in the tibia, with decreased tibial length, proximal tibial growth plate abnormalities, and excessive proximal tibial and mid-fibular flaring. Skeletal shortening extends to the axial skeleton, with pronounced loss of height in lumbar vertebrae. Genome-wide exome capture and high throughput sequencing of a chagun genomic DNA sample was conducted (http://www.broadinstitute.org/mouse-mutant-resequencing). No obvious deleterious mutations were uncovered in the mapped critical region using this approach. There were some synonymous changes and a single missense mutation that encoded an amino acid present in normal individuals of other species. Some variants were detected in intronic regions captured with the exons, but none appeared likely to disrupt splicing or create ectopic splice donors or acceptors (S1 Table). We extended our search using regional DNA capture to enrich for all non-repetitive genomic DNA in the most broadly defined 8.5 Mb chagun critical region (D9Mit183-D9Mit212). We captured genomic DNA from an obligate carrier (cha/+) and a known mutant (cha/cha), and obtained high throughput sequence with at least 20X coverage for approximately 90% of the bases sequenced from both samples. Manual sequence curation was carried out to detect insertions, deletions, inversions, etc. An insertion of an L1Md-GF-5 end non-LTR transposable element [27] was detected in an intron of the gene that encodes collagen type VI, alpha 6 (Col6a6). This element is not reported in 16 mouse reference genomes (http://www.sanger.ac.uk/resources/mouse/genomes/) [28]. PCR amplification and Sanger sequencing of this region in DNA samples from affected, unaffected, and DBA/2J mice indicates that it is present in all three samples. Thus, this transposable element does not cause the mutant phenotype. The difficulty inherent in sequencing and mapping repetitive elements likely led to omission of this element from the reference genomes. Manual sequence curation revealed a steep decrease in sequence coverage near exon 8 of Poc1a, which encodes protein of the centriole 1A (S1 Fig). This is more pronounced in the mutant than the heterozygote. In addition, many captured fragments from this region have mismatched paired end sequencing reads, where one read maps to chromosome 9 and the other to chromosomes 5, 10, or 12. PCR amplification of Poc1a exon 8 and flanking intronic regions produced a product that is ~500 bp larger in genomic DNA from chagun mice than in unaffected animals (Fig 2). Sanger sequencing of the genomic amplification products indicated that exon 8 is disrupted by the insertion of a nearly full-length transcript of centromere protein W (Cenpw) including a 61 nucleotide poly-A tail that is 24 bp downstream from the common polyadenylation sequence AUUAAA (S1 Fig). The total size of the insertion is 495 bp, and it includes a 5’ untranslated region (UTR) lacking only seven bp relative to the published transcription start site, all three exons, and a 3’UTR. This insertion is unique to chagun; it is not present in any of the 16 additional strains analyzed. The bona fide Cenpw gene is located on mouse chromosome 10, and processed Cenpw pseudogenes exist on chromosomes 5 and 12. The sequence of the Cenpw insert in Poc1a shares 100% identity with the exons of Cenpw on chromosome 10, and 90% and 84% identity with the pseudogenes on chromosomes 12 and 5, respectively. Thus, the inserted sequences of Cenpw in exon 8 of Poc1a are derived from a processed transcript of the bona fide Cenpw gene. The insertion does not delete any DNA from Poc1a exon 8. The exon is interrupted by the insertion of the Cenpw transcript in reverse orientation relative to Poc1a. The point of insertion is flanked by short (15 bp) stretches of identical Poc1a sequence (5’-CACCGTTGCCTTTTC-3’). This arrangement is consistent with target-site duplications characteristic of LINE-1 retrotransposon-mediated insertions (Reviewed in [29]), and with the consensus LINE-1 endonuclease ORF2 insertion site sequence (5’ TTTTC|A), just before the insertion of the Cenpw transcript [30–33]. RNA extracted from tibiae was prepared for RT-PCR analysis of Poc1a transcripts. The chagun cDNA produced a smaller amplification product than that detected in normal tibia (Fig 2). Sanger sequencing of the mutant RT-PCR product revealed that exon 8 of Poc1a is skipped precisely. No splicing into exon 8 was detected in chagun mutants (Fig 2). Quantitative RT-PCR was carried out using probes designed to amplify exons 2–3 at the 5’ end of the Poc1a transcript and exon 11–12 at the 3’ end of the transcript. The same level of transcripts were detected in RNA from wild type and chagun mutant tibiae using probes for both the 5’ or 3’ ends of the transcript (S1 Fig). This suggests that exon 8 is cleanly skipped and that the mutant transcript is similar in stability to the wild type. Skipping exon 8 preserves the reading frame of the Poc1a transcript. The mutant transcript is predicted to produce a POC1A protein that lacks 23 amino acids within the seventh and final 40 amino acid WD40 repeat (Fig 2). Western blots were carried out to detect POC1A protein in tibiae using a polyclonal antibody generated against full-length human POC1A, which is highly conserved (89% identity, UniProt) between human and mouse. A single band of similar intensity was detected in wild type and mutant tibia (Fig 2). Pan-tubulin immunoreactivity was used to normalize protein preparations. The molecular structure of mouse POC1A was modeled using the WD40 structure predictor algorithm [34]. Molecular modeling suggests that the deletion would destabilize the seven bladed propeller structure characteristic of WD40 repeat domain proteins because the first and seventh WD40 repeats normally interlock to form the propeller, and the mutant protein lacks 23 of 40 amino acids that would comprise the seventh propeller. Additionally, the WDSP algorithm predicts certain amino acids to be hotspots for protein-protein interactions based on their biochemical properties and their location on the top face of the propeller structure [35]. There is a predicted hotspot in WD6, Asn233, which normally forms hydrogen bonds with Ala274, a residue that is deleted in chagun mutants. The loss of this interaction likely disrupts the protein-protein interactions of Asn233. To validate the causal role of the mutation in Poc1a in the chagun phenotype, we undertook a BAC transgenic rescue experiment [36]. The selected BAC clone RP24-384G5 contains four genes in addition to Poc1a: Twf2, Tlr9, Alas1, and Dusp7. These genes do not cause problems when overexpressed, nor do their known loss-of-function phenotypes mimic any aspect of chagun mutants (S2 Table). Two of three independent BAC transgene insertion sites in founder mice provided completely penetrant correction of the chagun phenotype and caused no abnormalities in control transgene-positive animals. Transgenic chagun animals exhibit normal growth and testicular development (Fig 3). Body weight and length, as well as testicular weight, appearance, and histology were normal. This evidence supports our assertion that loss of Poc1a function due to the exon 8 insertion causes the chagun mutant phenotype. We obtained embryonic stem (ES) cells from the knockout mouse project in which exons 3–6 of the Poc1a locus were deleted and replaced by a gene trap internal ribosome entry site (IRES) LacZ and a neomycin resistance expression cassette (S2 Fig). We used these ES cells to generate mice with a null allele of Poc1a [37]. Heterozygous carriers for the LacZ knock-in were normal, and mice homozygous for this null allele phenocopied the growth insufficiency and hypogonadism of chagun mutants (Fig 3). We performed immunohistochemistry (IHC) using an antibody raised against a 50 amino acid peptide that contains the entire seventh WD40-repeat domain of POC1A. We detected POC1A in the proliferative zone of two-week old wild type (postnatal day 14; P14) tibial growth plates (Fig 4). The discoid chondrocytes in this zone have robust POC1A staining in the cytoplasm. Twenty-one amino acids of the peptide used to raise antibodies are intact in Poc1acha/cha mutants, and staining is observed in Poc1cha/cha mutant tibia. POC1A is also expressed in seminiferous tubules of the testis, and the immunostaining varies depending on stage of the seminiferous cycle. In wild type mice, POC1A protein is detectable in the cytoplasm of Sertoli cells and as single puncta in spermatozoa and spermatids, which likely represents labeling the centrosome of the spermatids as they begin to form flagella [38–40]. Little POC1A staining is observed in the testes of Poc1cha/cha mutants, suggesting that the mutant protein may be unstable or not retained in this tissue. Poc1a knockout mouse testes do not stain with this antibody, confirming the specificity of the antibody for POC1A (Fig 4 and S3 Fig). The expression of POC1A in wild type tibia is consistent with the observed defects in mutant chondrocyte proliferation and expression in wild type male germ cells and Sertoli cells suggests that one or both cell types could be involved in the infertility of male mutants. The centrosome is the microtubule-organizing center of the cell, and it is responsible for formation of the primary cilium and the mitotic spindle, and organization of the Golgi apparatus [41,42]. The centrosome is comprised of a mother and daughter centriole, and POC1A is a centriolar protein. Therefore, we hypothesized that the Poc1acha/cha mutation could affect any or all of these centrosome-dependent organelles. To test this prediction, mouse embryonic fibroblasts (MEFs) were isolated independently from three wild type and three mutant embryos and cultured for 24 hours on gelatin-coated coverslips. MEFs were either supplemented with serum to promote cell division and formation of mitotic spindles or serum-starved to facilitate visualization of primary cilia. After the culture period, cells were fixed and immunocytochemistry was conducted. Immunostaining with an antibody against acetylated tubulin, which labels the mitotic spindle and the primary cilium, revealed the effects of the mutation on these two cellular structures. MEFS cultured with serum revealed dividing wild type cells with equal bipolar spindles, as expected (Fig 5). The dividing Poc1cha/cha cells, however, frequently displayed evidence of centrosome amplification, with three or four spindle poles present in a single cell. Centrosome amplification likely results in aneuploidy and cell death. Immunostaining with a GM130 antibody revealed the presence of Golgi near the nuclei of wild type and Poc1acha/cha mutant MEFs. No obvious differences in Golgi organization were noted. The majority of wild type MEFs cultured without serum had a primary cilium detected by acetylated tubulin immunostaining. Significantly more MEFs from wild type embryos had detectable cilia than Poc1acha/cha mutants, 87% vs. 28% respectively. The lengths of the measureable cilia were similar in wild types and mutants, 3.06 μm vs. 2.67 μm, respectively. Loss of Poc1a function leads to defects in centrosome-mediated processes, including formation of a bipolar spindle and formation or maintenance of the primary cilium in MEFS. The growth plates of vertebrae and long bones of Poc1acha/cha mice are obviously disorganized [26]. Histology of tibiae from the neonatal period to P21 indicates that the growth plate becomes progressively more disorganized with age. At P0, the mutant growth plates look normal (Fig 6), but by P15, the coin stack structure of the chondrocytes in the proliferative zone begins to deteriorate in mutants, and some mutant chondrocytes have a rounded, rather than flat, appearance. By P21, Poc1acha/cha tibial growth plates are even more severely disorganized (S4 Fig) The subcellular position of the Golgi within the columnar chondrocytes of the proliferative zone can provide information about the polarity of the cell [3,43]. The Golgi is typically located to the right or left of the nucleus in chondrocytes, towards the edge of the coin stack column. To test whether the Golgi are positioned normally in Poc1acha/cha chondrocytes, immunohistochemistry was conducted on sections of the tibial growth plate utilizing primary antibodies against the Golgi marker GM130. The lateral subcellular positioning of the Golgi is evident in both wild type and mutant cells in the proliferative zones (Fig 6). The mutants have more nuclei that are more rounded, and the mutant cells appear mal-rotated relative to the cellular column. This indicates that the chondrocytes in Poc1acha/cha mice exhibit some polarized Golgi localization, but fail to maintain cell shape or ability to re-align after cell division. The primary cilia of chondrocytes in the columns of the proliferative zone are oriented in the direction of the column—pointing toward the resting or hypertrophic zone [3,44,45]. We conducted immunohistochemistry of tibia using a primary antibody against acetylated tubulin on sections from P15 wild type and Poc1acha/cha mutant mice to test whether primary cilia were properly oriented in the mutant growth plate. Wild type growth plates displayed normal orientation of the primary cilia within their columns. Primary cilia were detected in the Poc1acha/cha growth plate, and some of them are oriented properly (Fig 6). Formation of multipolar spindles could lead to premature chondrocyte cell death. Ki67 staining is similar in Poc1acha/cha mice, being largely expressed to the proliferative zone. TUNEL staining of P15 tibial sections demonstrated a substantial increase (10.3 fold, p = 0.002) in TUNEL-positive nuclei in the Poc1acha/cha growth plate (Fig 6). Together, these data indicate that the proliferating chondrocytes in Poc1acha/cha mice fail to maintain their cell shape, fail to intercalate properly after cell division, and undergo a significantly higher rate of cell death. Poc1acha/cha males are infertile. In adult mutant males, the latest stage of spermatogenesis that is detected is the pachytene stage, and hypogonadism ensues with age [26]. However, the time of initial germ cell disruption during postnatal development was not previously determined. To address this, we conducted histological analysis of the testis in normal and Poc1acha/cha mutants from birth to adulthood. Mutants were identified by Poc1a exon 8 genotyping and periodic acid-Schiff (PAS) histological staining was conducted on cross-sections of testes from males at postnatal day 1 (P1), P7, P14, P21, and 12 weeks (Fig 7). Immediately after birth (P1), the number of germ cell and Sertoli cell nuclei per tubule and the diameter of seminiferous cords in testis cross-sections from Poc1acha/cha males were indistinguishable from wild type counterparts, indicating that during embryonic development the proliferation and migration of germ cells to the genital ridge to form testis cords occurs normally. At P7, seminiferous cords from Poc1acha/cha males are still morphologically similar to wild type; however, the number of γH2AX positive cells is reduced in mutant males, indicating loss of spermatocytes at the earliest stages of meiosis (S5 Fig). By P14 the morphology of seminiferous tubule cross-sections from Poc1acha/cha mutants are dramatically different from controls. Progression of germ cell maturation to the pachytene spermatocyte stage is clearly evident in tubules of wild-type mice at P14. In contrast, only a few spermatocytes are detected in the mutant tubules. At P21, germ cell maturation is still arrested in testes of Poc1acha/cha mutants, and most of the germline appears to be lost in some of the tubules. The testis of sexually mature 12 week old wild type mice contain all stages of spermatogenesis, including undifferentiated and differentiating spermatogonia, primary and secondary spermatocytes, and round and elongating spermatids. In contrast, very few germ cells are present in testis cross-sections from 12 wk Poc1acha/cha mutants. Although some tubules contain a few leptotene or pachytene spermatocytes, most tubules appear to be devoid of all differentiating germ cells and contain spermatogonia only. None of the tubules in Poc1acha/cha adult mice contain secondary spermatocytes or post-meiotic round/elongating spermatids. Taken together, these observations demonstrate that the Poc1acha/cha testis phenotype is caused by a postnatal defect in spermatogenesis beginning at the pre-leptotene stage of meiosis, rather than impairment of testis cord formation during embryogenesis. Sertoli cells function to nourish the germ cells, phagocytose cytoplasmic remnants during spermatogenesis, and promote the translocation of the developing germ cells from the base of the tubule to the lumen. Impaired spermatogenesis could result from defects in Sertoli cells, germ cells, or both. To assess the relative contributions of Sertoli cells and germ cells, we used immunohistochemistry to detect these cell types, with primary antibodies against SOX9, PLZF (promyelocytic leukemia zinc finger protein) and c-KIT. SOX9 is a marker of Sertoli cells. PLZF marks the undifferentiated spermatogonial population that consists of spermatogonial stem cells (SSCs) and progenitor spermatogonia that are primed to undergo differentiation [46,47]. The expression of c-KIT indicates that spermatogonia have initiated differentiation [48]. At P14 the number of Sertoli cell nuclei per tubule cross-section is not different in wild type and Poc1acha/cha males (WT = 23.3±2.3 cells/tubule vs Poc1acha/cha = 26.6±4.0 cells/tubule, p = 0.52). At 3 months of age, the number of Sertoli cells per tubule is also not diminished in Poc1acha/cha testis. The apparent increase in Sertoli cell number per tubule in mutants (Fig 8A) is likely due to reduced tubule size and defective spermatogenesis in mutants that arbitrarily causes a greater concentration of Sertoli cells per mm of tubule length. The number of PLZF+ cells per testis section does not differ in wild type and mutant testes at 3 mo (Fig 8A), or at earlier ages (S6 Fig). The quantitation of undifferentiated spermatogonia (i.e. PLZF+ cells) in wild type vs Poc1acha/cha testis at each age was as follows: P7: 8.1 ±0.4 vs 4.8±1.2, p = 0.07. P14: 13.4±1.6 vs. 9.5±0.2, p = 0.07. P30: 5.9±0.9 vs 9.5±1.4, p = 0.1. The presence of c-KIT+ spermatogonia indicates that spermatogonial differentiation occurs in both wild type and Poc1acha/cha testes (Fig 8). The histone marker γH2AX is rapidly phosphorylated in response to double strand breaks, and it recruits DNA damage response factors. It is prominently expressed in all intermediate and B spermatogonia and in pre-leptotene to zygotene spermatocytes. While the proportion of tubules in Poc1acha/cha testes with γH2AX staining is similar to wild types at P7 (61%±14% vs 65%±8%, not significant), there is a major reduction (1.9 fold) in the number of γH2AX-positive cells within these tubules in the Poc1acha/cha mutant testes: 12.98% vs 6.73%, respectively (Fig 8, Panel D). The defect is much more severe at P14 (S5 Fig). This suggests a loss of germ cells early in meiosis at the pre-leptotene and zygotene stages, resulting in some spermatogonia at the pachytene stage and none at later stages. Condensation of chromosomes and formation of synaptonemal complexes takes place at these early meiotic stages. Next, we utilized spermatogonial stem cell (SSC) transplantation to determine whether the spermatogenesis defect was attributable to poor Sertoli cell function, intrinsic defects of the germ cells, or both [49]. In the first set of experiments, we transplanted Rosa-LacZ marked wild type spermatogonia into the seminiferous tubules of wild type recipients pre-treated with busulfan to eliminate the endogenous germline, or alternatively into adult Poc1acha/cha mutant males with impaired spermatogenesis (N = 4 recipients and 8 testes per genotype). Approximately 3 months after transplantation, X-gal staining revealed that the wild type donor SSCs generated colonies of donor-derived spermatogenesis in wild type recipient tubules, as expected. No colonization was detected in any of the Poc1a mutant testes (Fig 8B). The failure of Poc1a mutant testes to support the engraftment of wild-type SSCs is not attributable to the presence of residual mutant SSCs because donor SSCs can colonize normal testes with normal spermatogenesis occurring, albeit at a much lower rate [50]. These results are consistent with a Sertoli cell defect in the Poc1a mutants that could impair the homing ability of transplanted wild-type SSCs and/or re-establishment of colonies of continual spermatogenesis. Impaired homing seems unlikely to be the major cause because a primary spermatogonial population is established during neonatal development in Poc1acha/cha mutant testes. Thus, migration of germ cells from the lumen of the cords to the basement membrane during postnatal development occurs normally in Poc1acha/cha males. For the second set of experiments, we crossed the Rosa-LacZ marker into the Poc1a cha/+ stock to produce Poc1acha/cha males with LacZ marked germ cells. The marked, mutant spermatogonia were transferred into the seminiferous tubules of busulfan-treated wild type recipient nude mice. As a control, Rosa-LacZ marked wild-type spermatogonia were transferred to identical recipients at the same time. As expected, the wild-type SSCs generated robust colonies of spermatogenesis with uniform blue staining in testes of all recipients (N = 4 recipient mice and 8 testes). In contrast, only small patches of spermatogenesis with non-uniform blue staining were generated from Poc1acha/cha mutant SSCs, consistent with blunted expansion of colonizing SSCs and arrested spermatogenesis. Transplantation of mutant cells did not generate densely stained colonies of donor-derived spermatogenesis in any recipient transplanted testes (N = 4 recipients and 8 testes). Taken together, these findings suggest that the SSCs of Poc1acha/cha mutants are capable of initiating a colonization of a wild type microenvironment, but that intrinsic defects in the germ cells cause arrested spermatogenesis, even with wild type Sertoli cell support. These results are consistent with the idea that POC1A expression in both germ cells and Sertoli cells is important for normal testicular function. We report the discovery of the genetic basis of skeletal dysplasia and spermatogenic failure in chagun mutants. Insertion of a processed Cenpw transcript into exon 8 of the gene that encodes protein of the centriole 1A (Poc1a) causes elimination of a portion of one of the seven highly conserved WD40-repeat domains and abrogates function. We conclude this based on: 1) identification of an exonic insertion that causes skipping of exon 8 in Poc1a mRNA transcripts, 2) co-segregation of this Poc1a mutation with the chagun mutant phenotype, 3) the lack of unique insertions, deletions, or coding region mutations in other genes within the chagun critical interval, 4) successful BAC transgenic rescue of the mutant phenotype, 5) recapitulation of the mutant phenotype with a null, lacZ knock-in allele, and 6) expression of Poc1a in the affected tissues. Taken together, this provides compelling evidence that the insertion in Poc1a causes the chagun phenotype. Failure to detect the insertion by exome sequencing is probably attributable to the fact that all of the exon 8 genomic DNA sequence is still present in the mutants. The insertion of the Cenpw processed transcript increases the size of exon 8 from 69 bp to 564 bp, and increased exon size can cause skipping [51]. Skipping of exon 8 maintains the Poc1a open reading frame, and the predicted POC1A mutant protein lacks 23 amino acids in the most C-terminal, highly conserved WD40 repeat domain. POC1A has seven WD40 repeats that are expected to form a seven bladed, circular beta propeller structure. The lesion in the 7th repeat likely causes a failure of the blades to interlock. The mutant POC1A protein is readily detectable in tibia but not testis of Poc1acha/cha mice, suggesting tissue specific effects on protein stability. The Cenpw cDNA insertion appears to be a LINE-1 mediated event because it is flanked by a 15 bp target site duplication. In addition, the sequence 5’ TTTC|A in wild type exon 8 matches the ORF2 consensus (reviewed in [29]). Cleavage between the C and A at these sites permits target-primed reverse transcription by ORF2 reverse transcriptase [30,52,53], resulting in insertion of the transcript into exon 8 of Poc1a by the LINE-1 encoded proteins [54,55]. We are aware of only two other examples of LINE-1-mediated mutagenesis that involve non-LINE1 transcripts [56–58]. An insertion of a retrogene into an intron of FGF4 causes short stature in 19 dog breeds, and early dog breeders apparently selected for it. The insertion causes atypical expression of FGF4 in chondrocytes, rather than loss of function that we observed in Poc1acha/cha mice. Chronic granulomatous disease, an immunodeficiency disorder in humans, is caused by a LINE-1 mediated insertion of a partially processed transcript into an intron of the CYBB gene, which encodes a cytochrome that is essential for phagocytosis by leukocytes. The insertion causes loss of function; it disrupts splicing, resulting in incorporation of a novel exon and premature termination. Processed pseudogenes have also been detected in cancer genomes [59]. The difficulty in identifying the chagun mutation by exome sequencing suggests that the paucity of examples of LINE-1 mediated mutagenesis could, in part, due to alignment issues with the programs used to analyze exome sequencing data, resulting in ascertainment bias. Two mutations in POC1A (p.Arg81X and p.Leu171Pro) were recently reported in separate consanguineous pedigrees with severe short stature and craniofacial abnormalities [21,22]. Sarig, Sprecher, and colleagues labeled the syndrome a primordial dwarfism called SOFT Syndrome (short stature, onychodysplasia, facial dysmorphism, and hypotrichosis) [21]. Shaheen, Alkuraya and colleagues reported the p.Leu171Pro mutation, and they noted global developmental delay including cognitive impairment in some affected individuals, but they did not observe hair or nail defects [22]. We observed no fur or nail abnormalities in the chagun mutants or ES derived Poc1a null mutants. There was no discussion of hypogonadism or fertility of the human subjects in either report. The POC1A mutation p.Arg81X permits significant read through translation, which left open the possibility that it is a hypomorphic allele. The growth insufficiency, however, is severe in patients with either mutation. Mice homozygous for the ES-derived null allele of Poc1a have the same growth defect, skeletal dysplasia, and testicular features as Poc1acha/cha mice. This indicates that Poc1acha is a loss of function allele. The variable secondary features in the two families could be due to functional differences between the two alleles, contributions from mutations in other genes, or genetic modifiers. Both mutant mice exhibit severe growth abnormalities and craniofacial dysmorphism, and provide excellent models for the major features of the human syndrome. In metazoans, the centrosome regulates organization of microtubules and progression of the cell cycle [60]. Two centrioles, cylindrical, tubulin-rich structures that, together with additional peri-centriolar material, form a single centrosome in each cell. The centrosome is typically located centrally near the nucleus of the cell, but it moves to the leading edge of polarized, migrating cells. At the G1 to S transition, the centrosome begins to duplicate, and after the G2 to M transition, the mother and daughter centrosomes migrate to opposite poles of the cell and form the mitotic spindles. The centrosomes are not strictly required for spindle formation, but they are believed to enhance its efficacy and ensure the fidelity of cell division. While cells are in the quiescent state the centrosome migrates to the cell surface and forms the basal body of the primary cilium. The centrosome and Golgi apparatus are juxtaposed during interphase and are thought to interact functionally for directional protein transport. Proteome analysis in Chlamydomonas and Tetrahymena identified eighteen different proteins of the centriole, including Poc1, which is a core component of the centriole and basal body in all organisms with motile cilia [61]. Vertebrates have two genes, Poc1a and Poc1b, which are broadly expressed, exhibit 49% amino acid identity, and have overlapping function in cell culture [62]. Embryonic fibroblasts from Poc1acha/cha embryos exhibit a number of anomalies in centrosome-mediated processes. The majority of dividing mutant fibroblasts form aberrant mitotic spindles, and the primary cilia are infrequent and abnormally shaped. We did not notice any obvious defects in the organization of the Golgi. We observed binuclear Poc1acha/cha MEF cells, which are likely indicative of failed cytokinesis, similar to observations of Poc1 dysfunction in Tetrahymena [63]. Skin fibroblasts from human patients with POC1A dysfunction also had abnormalities in mitotic spindle polarity and the formation and length of the primary cilium [21], and although centrosome ultrastructure was normal, centrosome number increased and Golgi trafficking was impaired [22]. Knock down of Poc1a in cells leads to an inability to recruit markers of a mature centrosome [62]. Taken together, these data suggest that although POC1A deficient fibroblasts can form normal looking centriolar structures, the recruitment of proteins to the centriole is defective, causing abnormal centrosome function and increased number of centrosomes per cell. This leads to multipolar spindles and failed cytokinesis. The centrosome abnormalities also lead to defects in cilia formation and/or maintenance. Poc1acha/cha chondrocytes likely have defects in centrosomes, spindle poles and cytokinesis similar to those observed in Poc1acha/cha embryonic fibroblasts and POC1A patient fibroblasts [21,22]. Cell division in the growth plate is an intriguing process in which chondrocytes undergo polarized cell division orthogonal to the main direction of bone growth and spread back over one another to form columnar structures comprised of disc-shaped chondrocytes [6,64]. The growth plates of newborn Poc1acha/cha mutants have normal organization into zones and normal morphology of proliferating chondrocytes, but by puberty the proliferating chondrocytes are losing the typical coin-stack structure, and their shape is abnormally rounded rather than flat. As the proliferating cells become more disorganized, apoptosis is increased, and the final size of the proliferative zone is reduced. Although mutant chondrocytes can produce cilia and exhibit polarization of the Golgi, the function of the centrosomes and cilia are likely impaired. This view is supported by the similar chondrocyte disorganization and abnormally round cells in mice with disruption of Kif3a, a kinesin II motor complex protein required for intraflagellar transport and cilia formation [3]. Cilia defects are known to affect the Wnt/Planar cell polarity pathway, which is important for cellular shape, migration and organization [42,65,66]. The primary cilium is also important in the cellular response to mechanical cues that govern growth plate structure. For example, disruptions of the osteoblast primary cilium affect mechanoresponsiveness, mesenchymal cell differentiation [67], and the response to mechanical loading [68,69]. Thus, the centrosomal defects in Poc1a mutants could account for abnormal microtubule organization affecting chondrocyte morphology, aneuploidy leading to increased cell death, and poor cilia function influencing cell alignment at the growth plate. Testosterone levels, seminal vesicle development, and Leydig cell structure are all normal in Poc1acha/cha males [26], suggesting normal function of the hypothalamic-pituitary-gonadal axis. The germline is established normally in Poc1acha/cha males during embryogenesis and early neonatal life, as indicated by normal testicular morphology and PLZF-positive spermatogonia. Hypogonadism and infertility arise later in postnatal development when spermatogonia normally undergo a transition from mitotic to meiotic divisions. Defects in germ cells and Sertoli cells contribute to the infertility. Mutant SSCs transplanted into wild type hosts are capable of initial colonization, but they do not establish colonies typical of normal spermatogenesis. There are examples of spermatocytes with chromosome mispairing or DNA repair defects that are arrested at the pachytene checkpoint and undergo apoptosis [70]. In the current study, we extended the assessment of spermatogenic defects and discovered a reduction of meiotic cells at P7 in Poc1acha/cha males compared to wild type counterparts, which is when pre-leptotene spermatocytes first arise in the male germ cell lineage. Also, we found that the number of spermatocytes is greatly reduced in Poc1acha/cha males at P14 when pachytene stage cells first arise. Together, these findings indicate that defects in germ cell maturation begin to arise at the earliest stages of meiosis but become more pronounced as meiotic progression ensues, eventually leading to elimination of a majority of the population. We predict that Poc1a mutant spermatocytes are aneuploid, which triggers apoptosis at the pachytene checkpoint. Also, our findings suggest that Poc1acha/cha Sertoli cells do not support the engraftment of wild type SSCs or development into mature sperm, possibly due to defects in the microtubule organizing center and/or failure to secrete factor(s) essential for spermatogonial survival and/or germ cell maturation. The arrangement of microtubules in Sertoli cells is unique and highly specific to this cell type [71]. Defective centrosomes lacking POC1A could disrupt microtubule organization, leading to defective SSC niches [72] and/or an inability to support cells at later stages of spermatogenesis. The inability of wild-type SSCs to colonize Poc1acha/cha testes suggests that the mutation disrupts the ability of the testis microenvironment to support engraftment of normal SSCs in the niche at the basement membrane. Compromised trafficking of components required for the germ cells, or even the ability of the Sertoli cells to support the development of spermatocytes and spermatids could contribute to impaired spermatogenesis in Poc1acha/cha mutant testes. Immature or multiple centrosomes could impair the ability of Sertoli cells to support differentiation and development of spermatozoa from SSCs. Poc1a and Poc1b are both expressed in testis and bone (Fig 4 and S7 Fig), and have functional overlap in cell culture [62,63]. Mutations in human POC1B, however, cause a very different disorder. Homozygotes for the p.Arg106Pro mutation in POC1B have severe, syndromic retinal ciliopathy with defects in kidney and cerebellar function [73]. Individuals with p.Gln67del mutations also have recessive, non-syndromic cone rod dystrophy [74]. The basis for these tissue-specific effects is not clear. A loss of function of Poc1a causes skeletal dysplasia and male infertility in chagun mice. Insertion of a processed cDNA causes exon skipping and predicted deletion of 23 highly conserved amino acids necessary for the structural integrity of the protein. POC1A is expressed in the growth plate of long bones and the seminiferous tubules of the testis, the tissues with the most obvious cellular phenotypes. Several processes contribute to the growth defect. First, proliferating chondrocytes undergo enhanced cell death, likely due to multipolar spindle formation. Second, rapidly proliferating chondrocytes fail to maintain the characteristic flattened cell shape and fail to intercalate into a cellular column after cell division, consistent with defective cilia function. Male infertility is a consequence of the inability of Sertoli cells to support germ cell development, as well as germ cell defects in spermatogenesis. The Poc1acha/cha mouse has provided information on the molecular mechanism underlying clinical features of human patients, and highlights male fertility as a potential area of interest for clinicians examining adult male patients with mutations in Poc1a. The Poc1acha/cha and Poc1atm1(KOMP)Mpb mice are valuable tools for studying the molecular mechanisms of dwarfisms and for testing therapeutic interventions. The chagun mutation arose spontaneously on the DBA/2J strain in Dr. Linda D. Siracusa’s laboratory at Thomas Jefferson University (Philadelphia, Pennsylvania). The mice were transferred to the University of Michigan (Ann Arbor, Michigan). Recently they have been maintained on a hybrid background consisting of C57BL/6J and DBA/2J background strains, and also outcrossed to the FVB/NJ strain for routine maintenance. The BAC RP24-384G5 was purchased from Children’s Hospital of Oakland Research Institute (CHORI). Transgenic mice carrying this BAC were generated by the University of Michigan Transgenic Animal Model Core [36]. Embryonic stem cells containing the Poc1a tm1(KOMP)Mbp mutation, CSD45930, were purchased from the KOMP (University of California, Davis, CA). These stem cells, JM8A3.N1, were derived from the C57BL/6N-Atm1Brd strain. Two targeted clones were used: DEPD00572_5_B12 (Poc1a_B12), and DEPD00572_5_C09 (Poc1a_C9). These stem cells were injected into blastocysts and transferred to pseudopregnant surrogate mothers by the University of Michigan Transgenic Animal Model Core. Chimeras were mated to C57BL/6J females to obtain germline transmission. All mice were housed in a specific pathogen free facility with 12-h light, 12-h dark cycle in ventilated cages with unlimited access to tap water and Purina 5020 chow. All procedures using mice were approved by the University of Michigan Committee on Use and Care of Animals (UCUCA), and the Washington State University Institutional Animal Care and Use Committee (IACUC), and all experiments were conducted in accordance with the principles and procedures outlined in the National Institutes of Health Guidelines of the Care and Use of Experimental Animals. Euthanasia was conducted by CO2 inhalation, except for newborn mice that are unresponsive to this method. Decapitation was used for those animals. Surgery was conducted with approved anesthetics. Experienced veterinary care was provided. Two protocols were used to genotype for chagun. Before the chagun mutation was uncovered, the genotype at cha was inferred based on polymorphic flanking markers. Primers were designed to amplify regions of genomic DNA flanking the chagun critical interval that contained informative SNPs. These SNPs differ between the mutant (DBA/2J) and the wild-type (FVB/NJ) backgrounds: rs30174769, rs29637716. These amplification products were digested with Bsr1 endonuclease and run on a 2% agarose/tris-boric acid-EDTA gel to visualize the different bands that segregated differentially between the two background strains. After the mutation was uncovered, mutants, heterozygous, and wild-type animals were distinguished using primers designed to detect the presence of the insertion from flanking sequence (two potential products; a small wild-type allele, and the larger chagun allele with the insertion). Animals from the BAC transgenic rescue experiment were genotyped using the polymorphic SNP markers, and at least two additional sets of primers that amplify products specific to the BAC backbone. All primers anneal optimally at 60°C. The forward (Fwd) and reverse (Rev) primer sequences are: Animals were genotyped for the presence of the Poc1a tm1(KOMP)Mbp mutation by PCR with primers for LacZ and primers in Poc1a exons 6 and 7, which are not present in the Poc1a tm1(KOMP)Mbp mutant allele. Primers for LacZ amplification: Fwd: ATCCTCTGCATGGTCAGGTC Rev: CGTGGCCTGATTCATTCC. Amplification was at 94C for 3 min followed by 35 cycles of 94C 30sec, 58C 30 sec, 72C 30sec, followed by 72C 10min. Primers for detection of wt Poc1a: Fwd: TCTGCTTTGCGGTGTACGAA Rev: TTGGGTAGGGTGGGGTACAT. The conditions for this PCR are 92C for 2 min followed by 30 cycles of 92C 10sec, 57C 30sec, 72C 30sec. Two different capture approaches were taken. The Broad Institute conducted genome-wide exome capture and sequencing (Mutant Mouse Resequencing Project, The Broad Institute of MIT and Harvard, Cambridge, MA). Regional targeted enrichment was performed using custom capture probes targeting non-repetitive sequences within the chagun critical interval, and targeted enrichment libraries were generated (Roche/Nimblegen, Madison, Wisconsin). Genomic DNA samples from both a heterozygote (known carrier) and a chagun mutant were used to generate these targeted enrichment libraries, which were indexed separately to allow for multiplexed high-throughput sequencing (IlluminaHiSeq) by the University of Michigan DNA Sequencing Core. DNA Sequence reads were aligned to the B6 reference genome (mm9) using BWA [6]. Custom perl scripts were used to identify the location of potential insertional elements, using the paired-end read mapping data, and these were manually reviewed using Broad Institute’s Integrative Genomics Viewer to visualize and evaluate the DNA sequence data [28]. The region around Poc1a exon 8 was amplified by PCR of genomic DNA using the following primer sequences: Fwd: 5’-TGTCCCACTGCCACTGCCACTCA-3’, and Rev: 5’-GGAAGACTCGCCCCACAGGACTCA-3’. Optimal annealing temperature: 60C. PCR was conducted with the GoTaq DNA Polymerase and buffers provided by the distributor (Promega). Sanger sequencing was completed by the University of Michigan DNA Sequencing Core. Extraction of total RNA was completed with the RNAqueous 4-PCR Kit (Ambion) according to the manufacturer’s instructions. The total RNA was utilized to generate cDNA with Oligo dT Primers (Invitrogen/Life Technologies) and Superscript II Reverse Transcriptase (Invitrogen/Life Technologies) according to the manufacturer’s instructions. The cDNA was used as a template in standard PCR reactions using GoTaq DNA Polymerase (Promega) and the following primers to amplify two regions in the Poc1a cDNA: Optimal annealing temperature: 55°C. Sequencing was completed by the University of Michigan DNA Sequencing Core. TaqMan Universal PCR Master Mix (Applied Biosystems/Life Technologies) was used according to the manufacturer’s instructions. The following TaqMan probes were included to test expression levels of Poc1a: Exon 2–3 Assay ID: Mm01235877_m1. Exon 10–11 Assay ID: Mm01235875_m1. Reactions were loaded into MicroAmp Optical 96-well reaction plates (Applied Biosystems/Life Technologies) and run using an Applied Biosystems 7500 Real-Time PCR System. Isolation of protein from postnatal day 3 tibiae and western blot analyses were carried out as previously described [75]. The blot was incubated with a 1:500 dilution of a mouse anti-human POC1A primary antibody (Abcam, ab67698) overnight at 4°C. The blot was incubated with a 1:5000 dilution of a goat anti-mouse IgG secondary antibody (Jackson ImmunoResearch Laboratories, Inc. #115-035-003) for one hour at room temperature. The blot was stripped and incubated with a 1:5000 dilution of a rat anti-yeast tubulin antibody (Abcam, ab6160) overnight at 4°C. The blot was incubated with a 1:10,000 dilution of a goat anti-Rat IgG secondary antibody (Jackson ImmunoResearch Laboratories, # 112-035-102) for one hour at room temperature. All antibodies were diluted in a blocking solution made up of 1% weight:volume Bovine Serum Albumin;Tris-buffered saline with Tween-20. Adult skulls and were completely stripped of flesh by incubation in a dermastid beetle colony (http://www.lsa.umich.edu/ummz/mammals/dermestarium/default.asp.) Briefly, mice were euthanized, and bony parts were trimmed of excess flesh. These specimens were fed to the beetles. After the remaining soft tissues were removed, the bones were frozen at ~-20 C for 72 hrs, and the dead beetles removed. To assess shape and mineralization of bones, mice were euthanized, bones were dissected and skin and flesh trimmed. Specimens were imaged in water using a cone beam microCT system (eXplore Locus SP; E Healthcare Pre-Clinical Imaging, London, ON, Canada). Scan parameters were 0.5 degree rotational increment, 4 frames averaged, 80 kVp/80 μA X-ray, and 0.508 mm Al filter plus beam flattener to reduce beam hardening artifacts [76]. Volumes were reconstructed at 18 μm isotropic voxel size and calibrated for grayscale value by a manufacturer-provided phantom of air, water, and hydroxyapatite-mimicking material. Tibiae were dissected from 15 day old animals, fixed overnight in 4% paraformaldehyde (PFA), rinsed in PBS, and decalcified in 14% EDTA solution (weight:volume) for approximately 7 days, changing the solution each day. Testes were removed from mice of the listed ages and fixed overnight in Bouin’s Fixative Solution (Sigma). Both the testes and tibiae were then dehydrated through an ethanol series and embedded in paraffin. Sections were stained with hematoxylin and eosin (tibiae) or periodic acid-Schiff’s reagent (testes) according to standard protocols. Immunohistochemistry was conducted after removing paraffin from the samples by incubation of slides in xylenes and rehydration in an ethanol series to 1X PBS. Afterward, the testis sections were boiled in a 100 mM solution of citric acid (pH 6.0) for 10 minutes to expose epitopes, and cooled. Tibia sections were fixed for 10 minutes in 4% PFA, washed in PBS, and were either incubated with proteinase K in a buffer containing 50 mM Tris base, 1 mM EDTA and 0.5% Triton-X100 (acetylated tubulin), or placed in citric acid (pH 6.0) heated to 65°C for 1 hour (GM130, POC1A). Both testis and tibia sections were incubated for 20 minutes in solution of 3% hydrogen peroxide diluted in methanol to quench endogenous peroxidase activity. The slides were blocked in a solution included in the Tyramide Signal Amplification (TSA) Kit (Perkin-Elmer #SAT701001EA) or Mouse-On-Mouse (M.O.M) Blocking Reagent (Vector #BMK-2202) for one hour at room temperature. This was followed by an overnight incubation with primary antibodies in either TSA Block or Vector Mouse-On-Mouse Diluant at 4C. The rabbit anti-rat POC1A antibody was raised against rat POC1A aa 242–291, which is homologous to mouse POC1A aa 280–329. The mouse aa 269–308 comprise the WD7 repeat. The primary antibodies include: rabbit anti-rat POC1A (Abcam, ab135361, diluted 1:100), rabbit anti-human PLZF (Santa Cruz Biotechnology Inc., sc-22839 diluted 1:50), rabbit anti-SOX9 (Millipore, # AB5535, diluted 1:200), mouse anti-acetylated tubulin (Sigma #T6793, diluted 1:500), mouse anti-GM130 (BD Biosciences #610822, diluted 1:200), rabbit anti-KI67 (Novocastra NCL-Ki67p, diluted 1:250), rabbit anti-γH2AX (Ser139) (20E3) (Cell Signaling Technology #9718, diluted 1:500), and rabbit anti-c-KIT antibody (Cell Signaling Technology, # 3074S, diluted 1:400). Sections were washed in PBS, and incubated with the following secondary antibodies at room temperature: goat anti-rabbit biotin-conjugated secondary (Jackson ImmunoResearch Laboratories Inc., #111-067-003, POC1A, PLZF, SOX9, Ki67, γH2AX, and c-KIT for 1 hour), or the biotinylated anti-mouse secondary included in the M.O.M Kit from Vector Laboratories according to the manufacturer’s instructions. The subsequent steps were carried out according to the instructions provided in the TSA Fluorescein Tyramide Kit (TSA-FITC Kit) by Perkin-Elmer. Sections were counterstained with DAPI to reveal nuclei, cover slipped and photographed with a Leica Leitz DMRB/E compound microscope. Quantitation was done on three animals per genotype and age on 5–10 sections per individual and presented ± std. dev. Mouse embryonic fibroblasts (MEFs) were isolated from individual embryos at embryonic day 13.5 by the University of Michigan Transgenic Animal Model Core, and stocks were genotyped for Poc1acha/cha and frozen. MEFs were thawed, grown on gelatin-coated coverslips for two days, then serum treated or serum starved for 24 hours and fixed for 20 minutes in 4% paraformaldehyde at room temperature. The coverslips were washed in PBS, permeabilized in 0.1% SDS for 10 minutes at room temperature, and were incubated with the same primary antibodies for acetylated tubulin and GM130 listed above. The coverslips treated with acetylated tubulin primary antibody were then incubated in an Alexa 488 conjugated anti- mouse secondary (Life Technologies, # A-21141). The coverslips incubated with the GM130 primary were incubated with the M.O.M. anti-mouse Biotinylated IgG included in the M.O.M. Kit (Vector Laboratories). The GM130-incubated coverslips were then treated in the same way as the tissue sections according to the TSA-FITC Kit (Perkin-Elmer). All coverslips were counterstained with DAPI and photographed using an Olympus FluoView Laser Scanning Confocal Microscope (Microscopy and Image Analysis Laboratory, University of Michigan, Ann Arbor, MI). Sections through P15 tibiae were de-waxed in xylene, rehydrated through an ethanol series to 1 X PBS, and then pretreated by incubating the slides in citric acid (pH 6.0) at 65°C for 1 hour. Slides were washed in PBS and treated according to the manufacturer’s instructions thereafter (Roche In Situ Cell Death Detection Kit, Fluorescein, #11684795910). Quantification was done on three separate sections from two different mice of each genotype and presented as average per section ± standard error. The two-tailed T test was applied to assess significance. Rosa-Lac Z mice (Jackson Laboratories, Bar Harbor, ME, #002073) were crossed to Poc1acha/cha homozygous females to generate reporter positive heterozygous mice for intercrossing and generation of reporter positive wild type and mutant males for the experiments requiring labeled mutant germ cells. Spermatogonial stem cell transplantation was performed as described previously [77]. Briefly, single cell suspensions from donor testes were generated by two-step enzymatic digestion and spermatogonia were enriched by selection with a 30% continuous Percoll gradient. Adult wild type recipient mice were prepared by busulfan treatment (60 mg/kg of body weight) at least 6 weeks before transplantation to deplete endogenous germ cells as described previously [77]. Adult Poc1a mutant recipient mice were not prepared with busulfan treatment because the endogenous germline was already depleted. For experiments involving Poc1a mutant and wild type counterparts as SSC donors, NCr nude mice (Taconic) were used as recipients to avoid immunological incompatibility. For all transplantations, donor cells were re-suspended in mouse serum-free injection media [78] at 1X106 cells/ml and approximately 10 μl of cell suspension was microinjected into the seminiferous tubules of each recipient testis. The recipient testes were examined for donor-derived colonies of spermatogenesis via X-Gal staining ~3 months after transplantation. The molecular structure of mouse POC1A was predicted using the WD40 structure predictor algorithm (WDSP) [34] and models generated using PyMOL (The PyMOL Molecular Graphics System, Version 1.7.2.2 Schrödinger, LLC).

10.1371/journal.pntd.0004778

Improvement of POC-CCA Interpretation by Using Lyophilization of Urine from Patients with Schistosoma mansoni Low Worm Burden: Towards an Elimination of Doubts about the Concept of Trace

Accurate diagnostic techniques for schistosomiasis are essential for prevalence determination and identification of positive patients. A point-of-care test for detecting schistosome circulating cathodic antigen (POC-CCA) has been evaluated for its accuracy in different endemic regions. This reagent strip/dipstick based assay has showed high sensitivity for individuals with high or moderate worm burden, but the interpretation of light infections is less clear, especially for trace readings. We introduced a urine lyophilization step to the POC-CCA assay to improve its sensitivity and clarify the interpretation of traces. We evaluated POC-CCA sensitivity and specificity within individuals with low parasite burdens in a Brazilian endemic area where a high number of traces were detected. Patients that were positive for other helminths were also evaluated for cross reactions. In all cases, a combined parasitological diagnosis using Kato-Katz (24 slides) and Saline Gradient (1 g of feces) were used as reference. At baseline, diagnosis by POC-CCA (1–2 cassettes) showed 6% sensitivity, inaccurately predicting a low prevalence of Schistosoma mansoni infections (2 POC-CCA positives/32 egg positives). After urine lyophilization, the sensitivity was increased significantly (p < 0.05). Prevalence rates changed from 2% to 32% (27 POC-CCA positives/32 egg positives), equivalent to parasitological techniques. Most of the trace readings changed to positive after lyophilization while some negatives turned into traces. Cross reaction analysis confirmed the specificity of POC-CCA. Trace readings cannot be primarily defined as positive or negative cases. It is critical to verify case-by-case by concentrating urine 10 fold by lyophilization for the diagnosis. Following lyophilization, persistent trace readings should be read as negatives. No trained technician is needed and cost is restricted to the cost of a lyophilizer and the electricity to run it.

Schistosomiasis mansoni is a relevant disease affecting millions of individuals in different countries, in particular countries in Africa, and Brazil. Diagnosis performed by Kato-Katz technique for the detection of eggs in stool and a point-of-care test for circulating cathodic antigen detection in urine (POC-CCA) has been evaluated. Both methods have decreased sensitivity when diagnosing patients with low parasite burdens, which can lead to infected individuals not receiving treatment. Here, we focused on interpretation of POC-CCA results in persons with low parasite burdens. We noted a high number (49%) of indeterminate results, including false negatives and trace readings. A urine concentration step was included to improve the test’s sensitivity. Important differences on sensitivity and prevalence rates were noted when comparing diagnosis by POC-CCA before and after urine concentration. Notably, indeterminate results were easily defined after introduction of this step. Cross reaction analysis confirmed the specificity of POC-CCA, with exceptions noted for individuals with hookworm infection. In conclusion, trace readings cannot be primarily defined as positive or negative cases. It is imperative to analyze each case individually by concentrating urine prior to the introduction of treatment, instead of relying on a point-of-care test with indeterminate results.

World Health Organization (WHO) guidelines for control and elimination of schistosomiasis require pre-treatment evaluations of the prevalence of Schistosoma infections to inform decisions on how often to treat within endemic areas [1]. The WHO has articulated goals to control the disease by 2020 [2]. Accurate diagnostic techniques are essential for accurate determination of prevalence [3], evaluation of mass drug administration programs [4–8], elimination of the parasite [9–11], and/or drug-resistance and pharmacovigilance [12,13]. To address some of the concerns with the Kato-Katz reference technique, such as the need for evaluation of multiple slides to improve sensitivity, extensive research has been devoted to alternative methods with enhanced sensitivity and specificity for detection of S. mansoni infections [14]. Sensitivity and specificity of a urine-based point-of-care test (POC-CCA) has been evaluated in different endemic settings to detect schistosome circulating cathodic antigen (CCA) [15–20]. The test’s rapid turnaround and ease of use eliminate the need for multiple sample collections and specialized technicians. In addition, bulk production and purchasing of cassettes, particularly in the context of drug administration programs, have real potential for cost savings [21]. Different studies concluded that reagent strip/dipstick based tests have a good performance in detecting CCA in urine from individuals actively infected with S. mansoni [22–24]. Results from these studies consistently show higher S. mansoni prevalence scores by POC-CCA test in comparison to when single, double, quadruple or sextuple Kato-Katz thick smears were used [15, 16, 20, 24]. It was well established that the detection of S. mansoni increases with the increasing number of Kato-Katz smears examined and this pattern was consistently maintained in epidemiological studies [24–29]. However, controversies are found when discussing POC-CCA sensitivity in low endemicity sites, showing a consistent performance only in patients with moderate or high parasite burden [15, 16, 20–23]. It is unclear whether persons with positive POC-CCA readings who are Kato-Katz negative are truly infected or if they have false positive POC-CCA results. Although several analyses have been performed to explore POC-CCA performance [30], no data have been published concerning cross reaction of the test with helminths or other parasites. In addition, as a qualitative method based on an individual interpretation, doubts have been raised on how to differentiate the “trace” readings between low infection, cross reaction or even no active infection. Instead of being consistent about the meaning of the trace result, authors have chosen to perform a two-way analysis and consider traces as sometimes positive and sometimes negative. This produces vast discrepancies in prevalence intensities [15–18, 20–24, 30]. Thus in this paper we try to clarify the implications of a trace result. A better understanding of the interpretation of trace results is imperative for POC-CCA application in schistosomiasis control programs worldwide, especially in low endemicity areas that need particularly accurate diagnoses. Otherwise, praziquantel may be given to healthy individuals in error if a person has been incorrectly diagnosed. Adjustments to the assay’s implementation or interpretation of its ability to diagnose individuals with light infections are still needed. We introduced a single step in an attempt to improve the sensitivity and interpretation of POC-CCA trace result. Moreover, we show data from a Brazilian endemic region using the POC-CCA as a diagnostic tool. This improvement was then evaluated within individuals from a low endemicity area where a number of trace results were noted. The work includes initial diagnosis of patients with low parasite leads, that are commonly noted in endemic settings, and a comparison with a combined reference of 24 Kato-Katz slides plus 2 analyses (1 g of feces) by the Saline Gradient technique. The potential implications are discussed. This study was approved by the Ethical Research Committee of the Rene Rachou Research Center (CEPSH/CPqRR 03/2008) for human studies. All participants received an explanation of the study objectives. In addition, written informed consent was obtained before admission to the project. Parents/guardians provided written consent on behalf of all child participants. After parents/guardians had signed the informed consent, children received an explanation about the procedure, in a clearly explained language, and had the right to express their opinion. Procedures were performed in the presence of parents/guardians. Samples were coded and results were treated confidentially. Participants that were positive for parasitological tests were clinically examined by a physician and treated with praziquantel (60 mg/Kg for children and 40 mg/kg for adults) and albendazole (400 mg), in single oral dose, as recommended by the Brazilian Health Ministry. This study was conducted in Estreito de Miralta, a schistosomiasis-endemic region, next to the city of Montes Claros, Minas Gerais, southeastern Brazil, approximately 500 km from the state capital (Belo Horizonte). This endemic setting has a population of 163 individuals that had not received treatment for schistosomiasis within the last 2 years and had a low migration index. A schistosomiasis prevalence of 10.34% had been previously reported by the Montes Claros Zoonosis Control Centre in 2008. Positive patients were treated with praziquantel. The present study was conducted in 2013. Positive individuals were identified and treated, as recommended by the Brazilian Ministry of Health. Those patients submitted new fecal samples 30 days post-treatment for parasitological diagnosis and were retreated if needed. All the 84 individuals that provided urine samples (46 females and 38 males, 1–86 years old) were included in this study [31]. One sample of stool per individual of all the 163 residents was provided for Kato-Katz thick smear examination [31], performed with a total of 24 slides, a total of 1 g of feces examined per individual (24 x 41.7 mg of feces). Results were expressed as eggs per gram (epg) of feces, calculated by the number of S. mansoni eggs on the 24 slides. Fecal samples were also analyzed by Saline Gradient test (with two portions of 500 mg, total of 1 g of feces), as previously described [32]. Briefly, the separating column holding a filter was pre-wet with 3% saline solution. The separating column was filled with a fecal suspension prepared by diluting 500 mg stool sample in 3 ml of 0.9% saline solution. The saline flow was adjusted to 10 drops/min. After the slow and continuous flow of the 3% saline solution, low-density fractions were discharged and sediment was retained on the bottom of the latter column. Eggs, which had high density remained on the surface. The fractions containing eggs was moved to glass slides and examined under a bright field microscope. In order to detect other helminths, both parasitological methods were used as diagnostic tools. Results are expressed as epg. Each participant was asked to provide one midstream urine sample. Urine samples of the 84 individuals who provided urine were lyophilized to concentrate antigens. Briefly, the urine samples were aliquoted into vials (5 ml/vial), frozen at -20°C for 2 h and then overnight at -70°C and subjected to freeze-drying in a lyophilizer (Alpha 2–4 LD plus, Martin Christ Gefriertrocknungsanlagen GmbH, Osterode am Harz, GE) at 0.023 mbar and -55°C for 24 hours. Lyophilized samples were resuspended with water in a final volume of 0.5 ml, resulting in 10 times concentrated urine samples. POC-CCA tests were performed in accordance to the manufacturer’s instructions (Rapid Medical Diagnostics), before and after lyophilization. When a trace reading was obtained, a second cassette was used for confirmation. Briefly, one drop of urine was placed in the cassette’s well. Once it was absorbed, a drop of the kit buffer was placed in to the same well. Results were read after 20 min of test development. The tests were read as invalid when the control band did not appear or when the tests were left to develop for more than 25 min. Results were scored as “0” if the result was negative (i.e., the control line developed, but no test line appeared); trace if a very light test line appeared, “1+” if a test line appeared, but its color was less intense than that of the control line; “2+” if the test and control lines were equally intense in color; and “3+” if the test line’s color had a higher intensity than the control line’s color. Data collected from the evaluations were entered into an Excel data base and analyzed by Minitab statistical software (Minitab Inc, United States of America). The reference was defined as any positive slide performed for each individual stool sample by Kato-Katz or Saline Gradient technique. The sensitivity and specificity were determined with OpenEpi software (OpenEpi, Brazil) [33]. The agreement between the parasitological methods and POC-CCA were assessed by Kappa (k) statistics calculated by GraphPad (GraphPad Software, Inc., USA): k < 0.01 no agreement; k = 0.01–0.20 ‘poor’; k = 0.20–0.40 ‘fair’; k = 0.40–0.60 ‘moderate’; k = 0.60–0.80 ‘substantial’; k = 0.80–1.00 ‘almost perfect’ [34]. A total of 84 individuals participated in this study providing stool and urine samples. Using the combined results of Kato-Katz and Saline Gradient (both 1 g of feces/individual), no egg was detected in 52 individuals. Within those negative cases, POC-CCA was also negative for 42 individuals and 10 egg negative individuals had trace results. Together, 18 individuals presented eggs in stool for either parasitological test (7 by Kato-Katz, 7 by Saline Gradient, and 4 by both methods). Within those 18 individuals, based on POC-CCA test, 3 were negative (2 to 38 epg), 13 presented trace (1 to 55 epg) and only 2 were positive (8 and 16 epg). When two cassettes were used, results were reproduced for 75% of the urines. When cassette performances varied, one result was negative and the other was a trace for the same urine sample, but never a positive result. Table 1 shows the individual descriptive data for the three diagnostic tests. It is important to note that from the 49 individuals that were negative for POC-CCA, 3 presented eggs in only 2 Kato-Katz slides. The estimated prevalence in Estreito de Miralta by each of the three tests is shown in Table 2. For POC-CCA, analysis was first done by the direct application of the urine sample on the cassette and a second analysis was performed after 10 fold concentration by lyophilization. Kato-Katz and Saline Gradient had the same prevalence prediction of 30%. This prevalence rate was much higher than the one predicted by POC-CCA (1–2 cassettes) using unconcentrated urine of 2%. This 2% rate turned into a prevalence of 32% after the urine samples were concentrated, achieving a similar estimated prevalence as either parasitological technique. Fig 1 shows how readings obtained for the same individuals before and after urine lyophilization from negative to positive and from trace to positive, respectively. The sensitivity and specificity of the POC-CCA before and after lyophilization of urine samples were estimated with 95% exact CIs and are shown in Table 3. Respectively before and after lyophilization, POC-CCA presented 6% and 56% sensitivity and, 100% and 83% specificity. The concentration step improved POC-CCA performance, especially when trace results represented negative cases (Tables 4 and 5). Before lyophilization, 49 individuals were detected as negative, 33 presented trace and only 2 were positive for S. mansoni infection. Then, 13 initially negative individuals turned into traces and 11 into positives, after lyophilization. Of the 13 new traces, 10 individuals had no S. mansoni eggs in their stool, but 3 individuals were egg positive with 3, 8 and 52 epg of feces. By contrast, 4 individuals of the 11 concentrated urine positives presented 1, 2, 3 and 55 epg and the POC-CCA reading intensities were 1+, 1+, 3+ and 3+, respectively. However, 7 of the post-concentration positives were from egg negative patients; 5 of them had 1+ results, and one each had 2+ and 3+ results. Supporting initial data that considered trace as negative for infection, 15 traces turned into positives after lyophilization. Among those, 13 cases presented eggs in stool (3 patients with 1 epg, 1 with 2 epg, 1 with 7 epg, 1 with 8 epg, 2 with 9 epg, 2 with 10 epg, 1 with 15 epg, 1 with 38 epg and 1 with 55 epg). The two exceptions that had no S. mansoni eggs were positive for hookworms eggs. The Kappa index was used to compare parasitological data and POC-CCA urine assay and to better understand the implications of trace results. ‘Poor’ agreement was obtained when comparing parasitological and POC-CCA assays on unconcentrated urine (0.002 and 0.076, respectively for trace as positive and negative). The agreement changed to ‘moderate’, with a Kappa Index of 0.401, after urine samples were concentrated, but only when considering trace as negative for schistosomiasis. No change in the agreement between parasitological and POC-CCA data was noted when considering trace as positive for concentrated urine (0.125). Some individuals presented no S. mansoni eggs in stool, even after an extensive search, but had eggs of other helminths including hookworms, Hymenolepis nana, Enterobius vermicularis and Ascaris lumbricoides. No study participants were co-infected. Because it is common to find individuals with helminth infections other than schistosomiasis in endemic areas, we evaluated the POC-CCA on those individuals to determine whether infection with these other worms is associated with a false positive result. As shown in Table 6, among 7 individuals who were negative for schistosomiasis but positive for hookworms, 3 had a negative POC-CCA result and 4 had a trace result. Three S. mansoni-negative individuals were positive for H. nana. One was negative by POC-CCA, and 2 had trace reactions. All 4 individuals who were negative for schistosomiasis but positive for E. vermicularis and the one who was positive for A. lumbricoides had trace results. In order to see if traces would turn into positives after urine concentration, we evaluated the lyophilized urine samples for POC-CCA. In 12 of 15 patients, the POC-CCA results were the same before and after lyophilization. Three exceptions were found, one POC-CCA negative changed to trace for a person with hookworms infection, one pre-concentration trace result changed to 1+ for a patient with hookworms and pre-concentration trace result changed to a negative result for a person with E. vermicularis infection. The POC-CCA test is a promising technique that uses a nitrocellulose strip coated with monoclonal antibody to detect schistosome CCA antigen in urine samples. The antigen binds to the labelled monoclonal antibody immobilized on the nitrocellulose strip when urine from infected individuals flows through the strip. A band becomes visible with the binding of labelled monoclonal antibody [35]. Thus far, stool microscopy is the recommended diagnostic ‘gold’ standard, but individuals with low parasite burdens (i.e. < 100 epg of stool) are often missed [3, 6, 14–18, 21, 22, 25–29] and is indispensable to have technical expertise in microscopic recognition of intestinal parasite eggs. In the last few years, the POC-CCA has been extensively tested in areas endemic for schistosomiasis on the African continent [15–24, 30, 35]. These studies have shown the efficiency of a single or a double POC-CCA analyses in comparison to one or two Kato-Katz thick smears as the diagnostic standard. In all the cases, same relation was seen—for high parasite burden, higher positivity of POC-CCA is obtained. Although an increasing number of studies evaluating POC-CCA has been reported in endemic areas of Africa, there is so far no study evaluating its performance in Brazilian affected areas. Jointly, only 52 countries of the 78 countries considered endemic for schistosomiasis have populations requiring preventive chemotherapy, according to WHO [36]. It is a consensus that endemic areas in Africa and Brazil have different profiles regarding prevalence and morbidity. In this regard, we have assessed the accuracy of POC-CCA in a Brazilian endemic area where the parasite burden is low (1–80 epg). Our data showed that a number of trace readings was obtained (33/84 individuals) among whom 17 were negative and 16 were positive using parasitological techniques. Correspondingly, the sensitivity of the POC-CCA test used to evaluate prevalence was poor, as the prevalence rate was estimated as 2% when 1–2 cassette tests were performed. Kato-Katz is often criticized for its declining sensitivity when egg count intensities decrease, but the same situation was noted with POC-CCA under these conditions. POC-CCA seems to be appropriate for the diagnosis of S. mansoni when the prevalence is above 25% and no recent control efforts have been implemented [16]. Our findings show that patients from areas of low endemicity are difficult to detect. So, we propose to add a urine concentration step to the POC-CCA methodology to improve its sensitivity in low endemicity areas of Brazil. With this new step, the prevalence rate changed from 2% to 32%, achieving a comparable rate of Kato-Katz and Saline Gradient, performed on 1 g of feces (30% on both cases). The efficient identification of infected populations warrants effective chemotherapy, allows the development of new efforts toward elimination, including control interventions, assessment of drug efficacy, and patient management [4, 16, 37–38]. We must emphasize the importance of an accurate diagnosis at the individual and population level. By considering traces as positive, treatment may be performed inappropriately. On the other hand, infected patients could be deprived of receiving praziquantel treatment when traces are considered negative. Treatment-based control programs worldwide have been successful in reducing infection intensity and the number of persons with severe schistosomiasis. Conversely, transmission remains active in several endemic areas, and subtle but persistent morbidities are often found in persons with low-level reinfections, this is routinely seen, in Brazil. Accurate case-finding is indispensable for the effective execution of control programs [39]. Each trace reading must be individually analyzed since it may report a positive (low epg) or a negative situation, or even a cross reaction case. For urine tests that are scored as trace, we propose a lyophilization step of the urine to concentrate the sample. The introduction of this step in the POC methodology showed a clear diagnostic result. We show here that after lyophilization of urine, all remaining traces were negative. In addition, 15 traces turned into positive cases, 13 of which had a very low number of eggs in stool (1 to 55 epg) and the two others, although negative for S. mansoni eggs, were positive for hookworms. The Kappa index changed from ‘poor’ to ‘moderate’ agreement when considering concentrated trace results as negative when compared to parasitological data. When analyzing sensitivity, was 6% when we followed the manufacturer’s instructions. After the lyophilization, sensitivity increased to 56%. Considering the different profile of Brazilian endemic areas, it is relevant to compare the sensitivity rates achieved by Kato-Katz parasitological assay with the one achieved by POC-CCA. It is noticeable that increasing positive rates are obtained as slides are augmented in number, moving from 38.9, 43.5, 49.1, 50.0, 52.8 and 53.7, respectively for 1, 2, 3, 4, 5 and 6 slides [25]. Definitely, 6% is an unacceptable sensitivity rate for a reference diagnostic method, but a rate of 56% is comparable to the performance of several Kato-Katz slides after the concentration of urine. Differential diagnosis is also important since co-infection between helminths is commonly seen. In those cases, treatment may require different drugs. Only parasitological assays are capable of revealing differential egg identification, but the search for eggs in stool needs a trained and experienced technician. We tested POC-CCA cross reaction by analyzing urine samples of positive patients for other helminths (hookworms, H. nana, E. vermicularis, A. lumbricoides). Trace readings were seen for the POC-CCA assay (in individuals positive for hookworms, H. nana, E. vermicularis and A. lumbricoides eggs). If these traces were considered positives, as most of the authors do in Africa, we would had 73% of individuals incorrectly receiving praziquantel, instead of the correct drug (Albendazol), which would result in untreated patients with persistent infection and morbidity. The POC-CCA test is particularly well-suited to accurately demonstrate moderate to heavy S. mansoni infections and can be considered as a useful method for diagnosis in peripheral health centers and schistosomiasis control programs [16], but it does not present accurate results for low infections, as presented here unless the lyophilization step is included. A new potential diagnostic method called UCP-LF CAA has been tested for its accuracy as a urine-based up-converting phosphor-lateral flow circulating anodic antigen assay. The UCP-LF CAA assay showed high sensitivity for the diagnosis of S. haematobium in low-endemicity settings. According to the authors, the availability of scanners to analyze the UCP-LF CAA strip is a major step toward POC applications in poor resourced sites to accurately identify low (30 pg CAA/ml serum; equivalent to about 10 worm pairs) to heavy Schistosoma infections [40, 41]. The improvement of diagnostic methods with high sensitivity and specificity, and of simple execution and low cost will be vital for the accomplishment of the goals recently established by WHO [2]. These goals address the transmission control of schistosomiasis worldwide. Except for Africa, the transmission interruption should be accomplished by the end of 2020. For countries in the African continent, this goal should be achieved by 2025. The improvement of the POC-CCA test described in this study reinforces the possibility of introducing this methodology within the WHO schistosomiasis control proposal [2], not for all populations due to the obvious logistical difficulties, but as a tool to obtain additional data would that allow accurate interpretation of POC-CCA results in areas of low prevalence.

10.1371/journal.pbio.0050300

Distinct Mammalian Precursors Are Committed to Generate Neurons with Defined Dendritic Projection Patterns

The mechanisms that regulate how dendrites target different neurons to establish connections with specific cell types remain largely unknown. In particular, the formation of cell-type–specific connectivity during postnatal neurogenesis could be either determined by the local environment of the mature neuronal circuit or by cell-autonomous properties of the immature neurons, already determined by their precursors. Using retroviral fate mapping, we studied the lamina-specific dendritic targeting of one neuronal type as defined by its morphology and intrinsic somatic electrical properties in neonatal and adult neurogenesis. Fate mapping revealed the existence of two separate populations of neuronal precursors that gave rise to the same neuronal type with two distinct patterns of dendritic targeting—innervating either a deep or superficial lamina, where they connect to different types of principal neurons. Furthermore, heterochronic and heterotopic transplantation demonstrated that these precursors were largely restricted to generate neurons with a predetermined pattern of dendritic targeting that was independent of the host environment. Our results demonstrate that, at least in the neonatal and adult mammalian brain, the pattern of dendritic targeting of a given neuron is a cell-autonomous property of their precursors.

The mammalian brain contains a large number of different classes of neurons that are connected in a specific manner. A long-standing question is how such stereotyped connections emerge during the assembly of the brain. Here, we investigated whether neonatal and adult brain stem cells give rise to neurons whose connections can be influenced by the partners that they encounter while maturing, or alternatively, whether these connections are predetermined from the moment that a neuron is born. We observed the existence of distinct populations of precursor cells committed to generating neurons with a specific pattern of connections. Furthermore, the pattern of connections formed by these neurons was largely independent of the environment in which the neurons matured. These results have important implications for the formation of neuronal circuits, as they indicate that the connections of a new neuron can be determined in their precursors. In particular, these observations suggest that for neuronal replacement therapies to be successful, it will be necessary to understand the genetic programs that control how stem cells are prespecified to produce neurons with a stereotypic pattern of connections.

Dendrites are the major source of synaptic input for neurons. Thus, the specific computation that a neuron can accomplish is largely determined by the synaptic partners that contact its dendrites. In many regions of the central nervous system, including the cortex, spinal cord, retina, and olfactory bulb [1–4], neurons that share a common morphology and similar microenvironment have dendrites that target synaptic partners in different laminae. Although significant advances have been made in understanding the mechanisms that control axonal pathfinding during development [5,6], relatively less is known about the regulation of dendritic connectivity [7–9]. In recent years, some of the cellular mechanisms involved in dendritic growth into specific laminae have been characterized [4,7,9–11]. In addition, some cell-adhesion molecules involved in the formation of cell type–specific dendritic connectivity have been identified [8]. However, it remains unknown whether lamina-specific targeting is a cell-autonomous property of immature neurons, or alternatively, determined by the cellular environment in which the neurons differentiate. How lamina-specific dendritic targeting is specified is a particularly interesting question for neurons generated in the neonatal or adult period, because these new neurons have to integrate into a functioning, mature neuronal circuit. In this study, we examined the regulation of differential dendritic targeting of one neuronal type, the granule cell (GC) neuron of the olfactory bulb. GCs are axonless inhibitory interneurons that are continuously incorporated into the olfactory bulb throughout life [12,13]. GCs form dendro-dendritic synapses with the two types of projection neurons of the bulb, the mitral and the tufted cells (Figure 1A). GCs have distinct patterns of dendritic targeting—innervating either a deep or superficial lamina, where they connect to either mitral or tufted cells, respectively [14–16]. How is the lamina-specific dendritic targeting regulated in GCs? Throughout postnatal life, GCs are generated from neuronal precursors that proliferate in the subventricular zone (SVZ) and give rise to neuroblasts that migrate through the rostral migratory stream (RMS) into the olfactory bulb [12,13]. One possible scenario is that local cues within the olfactory bulb regulate the lamina-specific dendritic targeting of immature GCs at the time of their differentiation. Another possibility is that immature GCs are already committed to specific patterns of dendritic arborization at the moment of their birth in the SVZ, before they reach the bulb. To investigate these possibilities, we performed retroviral fate-mapping and transplantation experiments to test whether different populations of precursors give rise to GCs with lamina-specific dendritic arborizations. We discovered that the SVZ contained distinct populations of neuronal precursors committed to generate GCs with dendritic targeting to specific laminae. Furthermore, these precursors were largely restricted in their developmental potential with respect to dendritic targeting even when challenged with a SVZ microenvironment that normally generated GCs with dendrites that targeted the other lamina. Our results demonstrate that, in the neonatal and adult mammalian brain, the pattern of dendritic targeting of a given neuron can be a cell-intrinsic property that is already determined at the time of its birth. These findings have important implications both for assembly of neuronal circuits, and for the potential uses of adult neuronal stem cells in cell replacement therapies. In the neonatal brain, precursors in the SVZ give rise to GCs that integrate into the olfactory bulb. Most GCs have an apical dendrite that branches either in the deep or the superficial lamina of the external plexiform layer (EPL) (Figure 1A). To investigate whether precursors along the entire length of the neonatal SVZ have a similar developmental potential to give rise to GCs with a specific pattern of dendritic targeting, we used retroviral fate mapping to label precursors located in either one region of the anterior or posterior SVZ (aSVZ and pSVZ, respectively) of neonatal rats (Figures 1B and S1D). Oncoretroviruses have a half-life of only 6 h [17] and infect only actively dividing cells. Since the transient amplifying cell population is the most abundant actively dividing cell type in the SVZ, and the direct precursor to immature GCs, oncoretroviral infection is very effective for birth dating a single cohort of immature neurons [18]. Indeed, after infecting SVZ precursors with oncoretroviral vectors, we detected a single wave of labeled precursors that reached the olfactory bulb together. At 14 and 21 days postinfection (d.p.i.), only 4.8% (n = 1,218) and 1.3% (n = 780), respectively, of the total number of retrovirally labeled cells were still found in the RMS of the olfactory bulb. An oncoretroviral vector expressing green fluorescent protein (GFP) was stereotactically injected into the aSVZ or the pSVZ, and the morphology of GFP-positive (GFP+) GCs in the olfactory bulb was assessed 28 d.p.i., when they had acquired a mature neuronal morphology (Figure 2). We observed that the apical dendrite of GCs generated in the aSVZ of neonatal animals branched predominantly in the superficial lamina of the EPL, whereas the branches of the apical dendrite of GCs generated in the pSVZ of neonatal animals were mostly confined to the deep lamina of the EPL (Figure 2). This result was found to be independent of the strain or sex of the animals (see Materials and Methods and Figure S1A). Furthermore, we confirmed that the lamina-specific targeting and position of the initial dendritic branch point of new GCs was observed at all stages of their maturation (Figures 1C and 2). To quantify these findings, we measured the position of the initial branch point of the apical dendrite of GFP+ GCs from each precursor population (n = 250 for each time point after injection and each site of injection). To determine the position of the initial dendritic branch point, the width of the EPL was assigned percentages, with 0% being the mitral cell layer (MCL), and 100% being the border between the EPL and glomerular layer (GL) (Figure 1B). The EPL was then divided into 10% steps, and the position of the initial branch point of the apical dendrite was assessed using this scale. The cumulative distribution of the initial branch-point positions revealed that the dendrites of GCs generated from neonatal aSVZ precursors branched superficially, with a median initial branch point approximately halfway through the EPL (50.1% of EPL, median for all time points, n = 1,000; Figure 1C). In contrast, neonatal pSVZ precursors gave rise to GCs with apical dendrites that branched in the deep EPL, with a median branch point close to the MCL (2% of EPL, median for all time points, n = 500; Figure 1C). The distribution of the initial dendritic branching point was significantly different for neonatal-generated GCs from the aSVZ and pSVZ (p < 0.0001; n = 1,000 and 500, respectively). It is important to note that retroviral injection into the aSVZ is also likely to label some of the precursors that originate in other parts of the SVZ (e.g., pSVZ), but that proliferate in the aSVZ while they migrate through it on their way towards the olfactory bulb. However, after injecting into the aSVZ, the number of transit-proliferating cells that originated from the pSVZ was less than 5% (n = 500). As shown in Figure 1C, the cumulative distribution of the initial branching point for the pSVZ cells was very steep close to the MCL, whereas that of the aSVZ cells was nearly flat in this same region. In summary, fate-mapping experiments demonstrate the existence of at least two distinct precursor populations in the neonatal SVZ, each committed to generate GCs with specific patterns of dendritic targeting in the olfactory bulb. Because new GCs continue to be added into the olfactory bulb throughout life, we also investigated whether distinct GC precursor populations persisted into adulthood. Similar to our previous experiments, we observed that oncoretroviral infection of the adult aSVZ led to the efficient labeling of a single cohort of immature GCs in the olfactory bulb. At 14 and 21 d after retroviral labeling of precursors in adult rats, only 8.7% (n = 922) and 0.8% (n = 615), respectively, of the total number of cells labeled were still found migrating in the RMS of the olfactory bulb. Retroviral fate mapping revealed that GCs generated from precursors located in the aSVZ of adult animals had apical dendrites that branched predominantly in the deep lamina of the EPL (Figure 2). Quantification of this result indicated that the dendrites of GCs born from adult aSVZ precursors had a median branch-point position close to the MCL (4.5% of EPL, median for all time points, n = 1,000; Figure 1C), similar to the dendritic branching pattern of GCs born from neonatal pSVZ precursors. The distribution of the initial dendritic branching point was significantly different for neonatal- and adult-generated GCs from the aSVZ (p < 0.0001; n = 1,000, respectively). Again, this result was found to be independent of the strain or sex of the animals (see Materials and Methods and Figure S1A). Furthermore, as discussed for the neonatal animals, we confirmed that the lamina-specific targeting and position of the initial dendritic branch point of new GCs was observed at different stages of their maturation (Figures 1C and 2C). Interestingly, the normalized soma position of GCs from adult aSVZ was between that of GCs from neonatal aSVZ and pSVZ (Figure 1D) when compared at the same age (postnatal day [P]69–70). We also investigated the fate of actively dividing precursors in the pSVZ and in the sector of the RMS rostral to the SVZ in adult animals. The dividing precursors in these two regions both gave rise mainly to GCs with deep dendritic targeting even though both regions also contained some GCs with superficial dendritic targeting (see Figure S1B). This observation suggested that superficially branching GCs are still generated in the adult. However, using the retroviral labeling technique described in this study, we could not detect a SVZ region in the adult animal that exclusively contained actively dividing precursors committed to the generation of superficially branching GCs. Two recent studies [19,20] observed that generation of superficial GCs peaks in the neonatal period and decreases thereafter, consistent with our findings that precursors labeled in the different regions of the adult RMS and SVZ mainly gave rise to deep-targeting GCs. Taken together, these findings indicate that distinct populations of precursors in the SVZ gave rise to GCs that target either the deep or superficial EPL. In our initial experiments, we quantified the position of the initial branch point of the apical dendrite as a surrogate measure of lamina-specific dendritic targeting. To obtain a more comprehensive view of the dendritic targeting of new GCs generated from different SVZ precursor populations, we reconstructed the dendritic arbors of retrovirally labeled GFP+ GCs, selecting cells that displayed complete dendrites without apparent truncation due to tissue sectioning (Figure 2; n = 10 GCs for each time point and condition). Most GCs generated from neonatal pSVZ and adult aSVZ precursors had fine dendritic branches that ended in the deep lamina of the EPL, whereas GCs generated from neonatal aSVZ had fine dendritic branches that were located in the superficial EPL (Figure 2). GC reconstructions indicated that their dendritic arbors were largely confined to specific laminae, thereby suggesting that these GCs establish synaptic contacts in specific laminae. An alternative possibility is that GC synapses are not uniformly distributed throughout the dendritic arbor, and that the lamina-specific elaboration of the terminal fine dendritic branches did not reflect lamina-specific innervation. To further investigate these possibilities, we quantified the distribution of spine protrusions studding the branches of the apical dendritic arbor of GCs in the EPL using single-cell GC reconstructions (n = 10 GCs for each time point and condition). Dendritic spines are the major sites of excitatory synaptic transmission in the mammalian brain [21] and are thought to be morphological correlates of synapses. In GCs of the olfactory bulb, dendritic spines are the primary sites of both the inputs and outputs of dendro-dendritic synapses to and from mitral and tufted cells [16]. Few spines were found along the primary apical dendrite prior to the initial dendritic branching point. We found that neonatal pSVZ and adult aSVZ precursors generated GCs with spine protrusions confined to the deep lamina of the EPL (Figure 3), consistent with their pattern of dendritic arborization. Furthermore, neonatal aSVZ generated GCs with spine protrusions confined to the superficial lamina of the EPL (Figure 3). To further investigate the distribution of synaptic contacts in GC dendritic arbors, we also labeled the postsynaptic sites of glutamatergic synaptic inputs into GCs by expressing a genetic marker, PSD-95 fused to GFP, in GCs using an oncoretroviral vector. PSD-95 is a major scaffolding component of the postsynaptic density at excitatory synapses [22]. When GFP-tagged PSD95 is expressed in neurons, it clusters at the postsynaptic densities of glutamatergic synapses [23–25]. We found that the distribution of postsynaptic sites in GCs generated from precursors in the neonatal aSVZ and adult aSVZ, as labeled using PSD-95:GFP, was very similar to that described above for spine protrusions (see Figure S2 and Text S1). Thus, two independent methods indicate that distinct neuronal precursors in the postnatal SVZ generate GCs with lamina-specific patterns of synaptic innervation. GCs with deep and superficial branching dendrites labeled by retroviral infections shared similar cell morphology (Figure 2). We explored whether GCs with deep or superficial dendritic targeting may also share intrinsic somatic electrical properties that provide a useful criterion towards neuronal type classification [26–28]. Towards this aim, we performed targeted whole-cell recordings in acute slice preparations from GFP+ neurons with either superficial or deep dendritic targeting (21–23 d.p.i.) that were labeled in the aSVZ or pSVZ, respectively, in neonatal animals (Figure 4). Both deep and superficial neurons had similar delayed firing patterns (Figure 4A) and afterdepolarizations (Figure 4B). In addition, cells with either deep or superficial dendrites had IA currents that were abolished by exposure to 10 mM of 4-aminopyridine (unpublished data) and similar membrane properties (Figure 4C). The membrane capacitance of deep branching cells was larger than that of the superficial cells, most likely due to their differences in membrane surface. In summary, these data confirm that the new neurons with deep and superficial dendritic targeting not only shared a common morphology, but both also had similar intrinsic somatic electrical properties typical for GCs [29–32]. Thus, these observations suggest that distinct precursors in the SVZ are committed to giving rise to a single neuronal type with two alternative patterns of dendritic targeting. The factors regulating the lamina-specific targeting of GCs in the olfactory bulb are currently unknown. Our results raised the possibility that different populations of SVZ precursors may be committed to generate GCs with a particular pattern of dendritic targeting. Alternatively, local cues in the SVZ or olfactory bulb may control the developmental program of GC precursors or migrating GCs with respect to dendritic targeting. To investigate these possibilities, we performed heterochronic and heterotopic transplantations of different SVZ regions to examine whether their progeny adopt a different fate when challenged with new environments. In transplantation experiments, we isolated explants from three different sources, the neonatal aSVZ or pSVZ and adult aSVZ of GFP+ transgenic donor rats [33], and stereotactically injected them into the neonatal aSVZ or pSVZ and adult aSVZ of wild-type host rats (Figure 5A). The dendritic targeting of GFP+ GCs in the olfactory bulb was assessed 35 d post-transplantation to allow for their full maturation (Figure 5) (n = 200–731 neurons from 6–16 host hemispheres, per experiment). In order to validate that precursors in the SVZ retain their endogenous ability to generate GCs with lamina-specific dendritic targeting, we first performed isochronic, isotopic transplantation experiments in which the aSVZ from neonatal donors was grafted into the aSVZ of neonatal hosts. GCs derived from these transplanted precursors extended dendrites that targeted the superficial lamina of the EPL (Figure 5B and 5C), confirming the results of our retroviral fate-mapping experiments (Figure 1D and 1E). Similarly, isochronic, isotopic transplantation of either pSVZ of neonatal donors into the pSVZ of neonatal hosts or of the aSVZ of adult donors into the aSVZ of adult hosts resulted in GFP+ GCs whose dendrites targeted the deep lamina of the EPL (Figure 5B and 5C). These data demonstrate that the transplantation procedure itself did not perturb the endogenous developmental potential of GC precursors with respect to lamina-specific dendritic targeting. We then tested whether heterochronic and heterotopic transplantation of distinct SVZ regions can give rise to GCs with different dendritic targeting when exposed to a different environment. After heterochronic transplantation of precursors from adult aSVZ donors to neonatal aSVZ hosts as well as heterotopic transplantation from neonatal pSVZ to neonatal aSVZ, GFP+ GCs largely maintained their deep initial branching point (Figure 5C). After transplantation of precursors from neonatal aSVZ to adult aSVZ or to neonatal pSVZ, the GCs largely maintained their fate and had a superficial initial dendritic branching point (Figure 5C). We only observed a small increase in the number of GFP+ GCs that had a deeper initial dendritic branching point compared to isochronic and isotopic transplantation from the neonatal aSVZ donors. To quantify these observations, we counted the number of neurons whose initial branching point occurred below or above the midpoint of the EPL. We then calculated the ratio of cells that branched below the EPL midpoint threshold to the total number of GFP+ GCs. We measured this ratio for GCs derived from the same donor grafted heterochronically or isochronically, and then calculated the change in the ratio of GCs that initially branched below the EPL midpoint, and expressed it as a percentage. When neonatal aSVZ was transplanted into adult aSVZ, we observed a small increase (17.1%, Mann-Whitney test: p < 0.05%) of GCs that branched below the EPL midpoint, when compared to isochronic transplantation (neonatal aSVZ to neonatal aSVZ) from the same donor. Heterotopic transplantation from neonatal aSVZ to neonatal pSVZ resulted in 15.8% change (p = 0.13; not statistically significant) of GCs that branched below the EPL midpoint (Figure 5C). When we transplanted neonatal pSVZ into neonatal aSVZ, we observed a very small change (4%, not statistically significant) of GCs that branched above the EPL midpoint (Figure 5C). The small change observed in the population of GC precursors from the neonatal aSVZ after transplantation could be due to some partial phenotypic plasticity of these neonatal progenitors, or alternatively, to the transplantation procedure used in these experiments (see Discussion). Finally, heterochronic transplantation of adult aSVZ into neonatal aSVZ host did not induce a change (0.4%, not statistically significant) in GCs that branched below the EPL midpoint, when compared to isochronic transplantation (adult aSVZ to adult aSVZ) from the same donor (Figure 5C). In summary, even though some small changes can occur when the precursors are challenged with a new environment, the vast majority of new GCs derived from the different SVZ regions maintained their pattern of dendritic targeting. In order to obtain a more complete picture of the lamina-specific targeting of the apical dendritic arbor of transplant-derived GCs, we reconstructed the morphology of representative GFP+ GCs, as described above (Figure 6, n = 10 cells per condition). Neuronal reconstructions revealed that precursors that generate GCs with deep dendritic targeting in their native environment still gave rise to GCs with targeting of the deep lamina after heterochronic or heterotopic transplantation (Figure 6). The same observation applied to precursors that generate GCs with superficial dendritic targeting when challenged with a different proliferative environment (Figure 6). Finally, to confirm that the lamina-specific dendritic targeting reflects a lamina-specific distribution of synapses in these GCs, we determined the distribution of spines within the dendritic arbors of transplant-derived GCs. Similar to our previous findings, the distribution of dendritic spines reflected the lamina-specific dendritic targeting of transplanted GCs (Figure 7). Thus, our transplantation experiments indicate that precursors in the neonatal and adult animals appeared to be committed to generate GCs with a specific pattern of dendrite branching that was not modified by exposure to a brain environment that normally generated GCs with opposite dendritic targeting. In this study, we investigated lamina-specific dendritic targeting of neurons generated in neonatal and adult rats. In particular, we tested whether local factors determine the dendritic targeting or, alternatively, whether a cell-intrinsic program is conferred onto the neuron from its precursor. To investigate these possibilities, we performed retroviral fate mapping and reconstructions of GCs generated in postnatal life in different regions of the SVZ. Our results indicate that at least two separate populations of precursors exist in the SVZ and give rise to GCs that target either the deep or the superficial lamina. Therefore, distinct precursors can produce one type of neuron, as defined by its morphology and intrinsic electric properties, but exhibiting different dendritic targeting. We performed heterochronic and heterotopic transplantations of different SVZ regions and observed that the majority of new GCs maintained their fate of targeting a specific lamina even when their precursors were grafted into SVZ environment that normally generated GCs with dendritic targeting of the opposite principal neuron lamina. Our results demonstrate that, in the mammalian brain, the pattern of dendritic targeting of a given neuron can be a cell-intrinsic property determined at the time of its birth. These findings have important implications, both for understanding the assembly of brain circuits during development and for the potential uses of adult neuronal stem cells in cell replacement therapies. Our findings indicate that the connectivity of one type of neuron as defined by its morphology and intrinsic somatic electrical properties in the same brain region, the GCs of the olfactory bulb, can in fact be determined by the particular population of neuronal precursor from which they derive. This observation suggests that neuronal precursors in the mammalian brain may be committed to produce neurons that are tailored to perform specific functions in a given neuronal microcircuit from as early as the time of their birth. Our finding that lamina-specific dendritic targeting can be an intrinsic property of an immature neuron determined by the identity of its precursor has important implications for the logic of neuronal circuit assembly. In particular, GCs in the olfactory bulb that target the superficial lamina of the EPL are believed to establish synapses with tufted cells [16,34], whereas GCs that target the deep lamina are connected to mitral cells, and these two microcircuits are believed to serve different functions. The tufted-GC circuit is thought to be an intrabulbar association microcircuit [34] that may be important for low-threshold perception of odorants [35]. In contrast, the mitral-GC circuit is thought to mediate lateral inhibition and to participate in odor discrimination [36]. Thus, GCs that participate in specific microcircuits may be committed to their function already at the time of their birth from distinct populations of neuronal precursors. To validate whether the neurons we labeled indeed constitute the same neuronal type with different dendritic targeting as suggested by their similar morphology, we measured their intrinsic somatic electrical properties, a useful feature for classification of neurons [26–28]. Indeed, labeled neurons with either deep or superficial dendritic targeting had similar intrinsic somatic electrical properties. This observation further suggests that one neuronal type (GC in the olfactory bulb) can exhibit two alternative patterns of dendritic targeting. This observation does not, however, preclude that minor differences may exist between GCs with deep and superficial dendritic targeting, such as differential expression of neurotransmitter receptor subunits [37]. In summary, our results suggest that distinct precursors can generate “tailor-made” neurons with different dendritic targeting connected to specific microcircuits. In addition, within one neuronal type, cells with alternative patterns of dendritic targeting may have subtle functional differences, such as differential expression of neurotransmitter receptors or ion channels, specific for their function in separate microcircuits. Our findings also provide important insights into the developmental processes by which dendritic patterning is established in the mammalian brain. Particularly in comparison to axonal targeting, the mechanisms that regulate dendritic connectivity and allow neurons to establish proper contacts with their synaptic partners are not well understood [3]. Existing models for dendritic targeting can be divided into two major camps: outgrowth followed by pruning, or directed growth. For instance, in the mammalian retina, the dendritic arbor of retinal ganglion cells initially ramifies broadly, but as development proceeds, part of the dendritic branches are eliminated, such that the dendrites are ultimately segregated into two different laminae [7]. Additionally, a large body of work suggests that the refinement and stabilization of dendritic arbors may also be dependent on experience, a mechanism that would allow the maturing brain to adapt to a changing environment in postnatal life [4,10,11]. In other cases, the growth of dendrites can be targeted to specific laminae or layers in a directed manner. This mode of directed dendritic targeting has recently been demonstrated in the Drosophila olfactory system [38] and in the vertebrate retina [9]. For instance, in vivo imaging of zebrafish retinal development revealed that the dendrites of distinct classes of neurons directly grow to and innervate a specific lamina during their development [9]. How are such programs of dendritic development specified and implemented? Our experiments indicate that, in the mammalian olfactory bulb, the lamina-specific dendritic targeting of GC neurons is an intrinsic property determined by the precursor from which it arises. Our findings are compatible with both modes of dendrite growth described above. In one scenario, new neurons may directly extend their dendrites into the specific EPL lamina where they will form synaptic contacts. Alternatively, dendrites may initially grow in an exuberant manner through both laminae, but they will only form contacts with one type of principal neuron in either lamina, as determined by their precursors, and prune the rest of their dendritic arbors. Finally, after these lamina-specific dendritic contacts have been established, neuronal activity-dependent mechanisms then may play a role in the fine sculpting of GC dendritic arborization. The mechanisms that regulate the generation of neuronal diversity in the vertebrate nervous system have been investigated extensively. Various studies have shown that both the spatial and temporal origins of precursors determine the neurotransmitter phenotype, firing properties, calcium binding protein expression, and position of the cell body in different layers [28,39–43]. In particular for interneurons, distinct precursors defined by their expression of transcription factors give rise to specific types of interneurons for different brain regions [42]. Such specialization of precursors to produce different cell types persists throughout life in the SVZ for periglomerular and GC neurons [44–47]. While this work was under review, a study was published [48] demonstrating that the SVZ of postnatal animals has a mosaic organization, with different zones containing precursors committed to generate different types of periglomerular and granule neurons, as revealed by the presence of a set of immunocytochemical markers and the position of their cell bodies in the olfactory bulb. Our study advances previous observations [44,48] by demonstrating that the location of the dendrites and synapses of granule neurons is an intrinsic property of the cell, and by showing that there exist distinct precursors committed to generate neurons with dendritic targeting to specific laminae. Furthermore, we demonstrate that dendritic targeting is determined in the precursor cells in the lateral ventricle, before the progeny from these precursor cells have reached their target in the olfactory bulb. The hypothesis of the protomap, as originally proposed for the developing cortex, stated that the progenitors in the brain ventricles already contain the information [49,50] that specifies the identity of the neuronal cell types of the progeny that they will give rise to, their final destination in the different cortical layers, and the features specific to the different functional areas of the cortex. Our observations extend the protomap hypothesis by showing that the pattern of dendrite arborization can also be a feature already determined in the brain ventricles, before the progeny of the neuronal stem cells have reached their target. Furthermore, heterochronic as well as heterotopic transplantations of precursors confirmed that the fate of the dendritic targeting of a new neuron was maintained for the large majority of donor-derived GCs independent of the host environment in which their precursors had been grafted. These observations suggest that the connectivity of a neuron can be a cell-autonomous characteristic, determined by an intrinsic program in neonatal and adult neuronal stem cells. Interestingly, we observed a small population of new GCs (<17%) derived from the neonatal aSVZ that displayed dendritic targeting to the opposite lamina both after heterochronic (into adult aSVZ) and heterotopic (into neonatal pSVZ) transplantation. In contrast, GCs derived after heterochronic and heterotopic transplantation of neonatal pSVZ and adult aSVZ donor tissue did not change their fate of dendritic targeting when challenged with a new SVZ environment. Several reasons could account for the small change of dendritic targeting after heterochronic or heterotopic transplantation from neonatal aSVZ donor tissue. First, a small population of aSVZ precursors could be reprogrammed to generate GCs with deep dendritic targeting after transplantation. Second, cells from the pSVZ, which migrate through the aSVZ, may be induced to proliferate and expand when transplanted, and this phenomenon could increase the number of GCs with deep dendritic targeting after transplanting the aSVZ. Third, a previously quiescent stem cell present in the neonatal aSVZ may be activated when exposed to a SVZ environment that generates GCs with deep dendritic targeting. Our experiments cannot currently distinguish between these and other possibilities, since the aSVZ tissue that is transplanted contains neuronal precursors at different stages of commitment, including rarely dividing stem cells, transient amplifying cells, and immature migrating neurons. In addition, by transplanting explants of tissues into a new SVZ environment, it is possible that donor cells surrounding the neuronal precursors could preserve the status of the donor niche, thus preventing the full reprogramming of the grafted progenitors. Nevertheless, our findings indicate that the precursors in the SVZ are committed to generate GCs with a prespecified pattern of dendrite targeting before they reach their target. In recent years, there has been a surge in interest in the possibility of using different types of stem cells for cell replacement therapies aimed at correcting neurological disorders caused by neuronal loss, such as stroke and Parkinson, Huntington and Alzheimer diseases [51,52]. Our observations indicate that distinct neuronal stem cells are committed to generate not only a single neuronal type, but also cells with a prespecified pattern of dendritic targeting. Understanding the program by which neuronal stem cells specify how a neuron will target its dendrites towards a given synaptic partner could help to achieve neuronal replacement with cell type–specific connectivity. Further, the potential uses of endogenous adult neuronal stem cells/precursors for neuronal repair could be hindered by their lack of phenotypic plasticity as revealed by this and other recent studies [44–47]. Thus, the determination of cell type–specific dendritic connectivity by separate neuronal precursors may have important implications, both for the potential uses of adult neuronal stem cells in cell replacement therapies and for understanding the assembly of brain circuits during development. We used an oncoretroviral vector derived from the Moloney sarcoma virus expressing GFP under the control of the Rous sarcoma virus promoter (MolRG). Recombinant virus was prepared and stored as described [33]. The viral titer was 106–107 infectious units/μl. Animal care and procedures were approved by the local animal welfare committee. Neonatal (P5) and adult (>P56) Sprague-Dawley, Wistar Kyoto, and Lewis rats of either sex were anesthetized by hypothermia (neonatal rats) and with ketamine/xylazine (adult rats). In initial experiments, we injected P3 to P8 animals in the aSVZ and pSVZ. Between P3 to P8, we observed a superficial and deep branching population of GCs for aSVZ and pSVZ, respectively. For consistency, further experiments were performed at P5 for neonatal rats. Stereotactic injections were performed with a glass capillary with a tip diameter of 3–5 μm, and a volume of 0.1–0.5 μl of viral vector stock was injected. The following stereotactic coordinates (relative to bregma in millimeters) were used for neonatal animals: aSVZ: anterior 0.9, lateral ±2.1, ventral 2.1; pSVZ: posterior 0.6, lateral ±2.7, ventral 2.6; and for adult rats: aSVZ: anterior 1.2, lateral ±1.6 ventral 3.1; aRMS: anterior 2.8 lateral ±1.1, ventral 5.4; pRMS anterior 2.3, lateral ±1.4, ventral 4.5; pSVZ: posterior 2.7, lateral ±4.5, ventral 3.4. For neonatal sites of viral infection, see also Figure S1D. After surgery, animals were monitored for 24 h. Quantification of morphology of GCs was only performed at 14 d.p.i. and later time points in order to avoid including immature neurons that could still be migrating or had not yet acquired a mature morphology. After 14 d.p.i., most GCs from different origins (adult or newborn, aSVZ or pSVZ) had acquired a mature neuronal morphology (see Results). Infecting dividing precursors in the RMS within the adult olfactory bulb did not give rise to GFP+ GCs (n = 6 hemispheres injected; unpublished data), most likely due to the low level of proliferation in his region [20,53]. FUGW+ transgenic rats [33] were bred on a Sprague-Dawley background. The aSVZ or pSVZ of neonatal and adult GFP+ rats was dissected (same regions as in Figure S1D), cut in small pieces, and then stereotactically transplanted into the aSVZ or pSVZ of either neonatal or adult wild-type Sprague-Dawley rats with the stereotactic coordinates described above. Rats were killed with ketamine/xylazine at the indicated time points for retroviral fate mapping or 35 d post-transplantation and perfused intracardially with 3% paraformaldehyde. After 24 h post-fixation in 3% paraformaldehyde at 4 °C, brains were cut into 50-μm coronal sections on a vibratome. Tissue sections were incubated with rabbit polyclonal anti-GFP antibody (1:3,000; Chemicon) in blocking solution containing phosphate buffered saline (PBS), bovine serum albumin, and 0.3% TritonX100 overnight at 4 °C, washed several times with PBS, and incubated with secondary anti-rabbit Alexa488 or Alexa555 conjugated antibody (1:750; Molecular Probes) for 2 h at room temperature. Tissue sections were washed in PBS and counterstained with Hoescht 33258 (Molecular Probes). For stereological analysis and neuronal reconstructions, we used a Neurolucida system coupled to an inverted Olympus fluorescent microscope with a motorized X-Y-Z stage. For stereological analysis of the position of the soma and the initial branching point of the apical dendrite, we first determined the position of the soma and then traced the apical dendrite to its first branching point. All neurons of a tissue section that were not truncated before their initial dendritic branching were counted. For each time point, 250 neurons were traced from nine to 26 different olfactory bulb sections (n = 250 neurons for each time point from four to eight injected hemispheres from more than three animals) depending on the density of GFP+ GCs. For transplantation experiments, 200–731 neurons were traced. In initial experiments, we determined the distribution of the soma and the initial branching of the apical dendrite in serial sections throughout the anterior-posterior axes of the bulb. As we did not find any regional differences for the position of the soma and of the initial dendritic branching (unpublished data), we used sections from the central parts of the bulb for further analysis because they gave the highest yield of GFP+ GCs. In addition, we did not observe any differences in the soma distribution and the position of the initial dendritic branching point for animals of either sex, therefore data from both sexes were pooled. Host hemispheres differed in their density of GFP+ GCs. The distribution of deep and superficial GCs did not, however, differ regardless of the cell density in the same transplantation condition. For each tissue section, we traced the borders of the different layers based on the nuclear counterstaining with Hoechst 33258 (see also Figure 2). Based on these borders, we divided the granule cell layer (GCL) in percentages: 0% being the border between the RMS and GCL, and 100% being the MCL. For our analysis, dividing the internal plexiform layer (IPL) and the GCL did not prove useful because many of the GC somata were located in the level of the IPL or around the MCL. Based on these borders, we divided the EPL in percentages: 0% being the MCL and 100% being the border between the EPL and the GL. For further analysis, the GCL and the EPL were divided in 10% steps, and the position of the somata and of the initial branching of the apical dendrite was plotted as a cumulative distribution. We calculated the differences in percentages of the ratio of GCs that initially branched below 50% (40%) of EPL. The lower threshold (40%) of EPL gave similar results (±3.4%) to the 50% threshold. GCs for single-cell reconstructions were selected based on the typical position of the soma and on the initial branching of the apical dendrite as found for the respective population. Ten GCs for each time point or transplantation condition without apparent truncation due to tissue sectioning were then reconstructed. Of these GCs, we marked spine protrusions manually with a 40× lens while continuously adapting the z-axis. All spine-like protrusions were counted that emerged from the dendrites and had a thickness and morphology that would make them appear as spine- or filapodia-like structures. The distribution of spines of the apical dendrite for each GC was attributed to percentage ranges in the EPL (as defined above, here 20% steps) and then the distribution of spines in the EPL for ten GCs was averaged for each time point or transplantation condition. Statistical significance (p < 0.05) was determined with a nonparametric Mann-Whitney test for unpaired samples. Rat pups were bilaterally injected with 1 μl of oncoretroviral vector expressing GFP in aSVZ and pSVZ in neonatal rats. At 21 to 23 d.p.i., animals were anesthetized with isofluorane, and brains were rapidly removed. The 350-μm horizontal olfactory bulb slices were cut with a Leica vibratome in cutting solution containing (in mM): 212 sucrose, 3 KCl, 1.25 NaH2PO4, 26 NaHCO3, 7 MgCl2, 0.5 CaCl2, 10 glucose, 310 mOsm, (pH 7.3). Slices were recovered for 30 min at 32 °C with recording solution containing (in mM): 125 NaCl, 2.5 KCl, 1.25 NaH2PO4, 26 NaHCO3, 1 MgCl2, 2 CaCl2, 20 glucose, 310 mOsm, (pH 7.3) and continuously bubbled with carbogen. After recovery, slices were kept at room temperature. Targeted whole-cell recordings were performed on GFP+ nontruncated CGs with a MultiClamp700B amlpifier (Axon Instruments) and pipette solution containing (in mM): 2 NaCl, 4 KCI, 130 K-gluconate, 10 HEPES, 0.2 EGTA, 4 ATP-Mg, 0.3 GFP-Tris, 14 phosphocreatine, 0.02Alexa555 hydrazide, 292 mOsm, and pH 7.25. Pipette resistance was 6-9 MW. Access resistance was 12-30 MW, which was not compensated and regularly monitored during recordings. Liquid junction potential was not corrected. Data were acquired and analyzed with the pClamp9 software (Axon Instruments). Neurons were considered to hav I current if in a voltage ramp (10 mV steps for 400ms), the peak-to-plateau ratio was > 2. After recordings, the tissue was fixed and the neurons were reconstructed. Recordings from the correct GCs were conffirmed by coocalization of Alexa555 flourescence with GFP+ GCS.

10.1371/journal.pgen.1007759

Balancing selection at a premature stop mutation in the myostatin gene underlies a recessive leg weakness syndrome in pigs

Balancing selection provides a plausible explanation for the maintenance of deleterious alleles at moderate frequency in livestock, including lethal recessives exhibiting heterozygous advantage in carriers. In the current study, a leg weakness syndrome causing mortality of piglets in a commercial line showed monogenic recessive inheritance, and a region on chromosome 15 associated with the syndrome was identified by homozygosity mapping. Whole genome resequencing of cases and controls identified a mutation causing a premature stop codon within exon 3 of the porcine Myostatin (MSTN) gene, similar to those causing a double-muscling phenotype observed in several mammalian species. The MSTN mutation was in Hardy-Weinberg equilibrium in the population at birth, but significantly distorted amongst animals still in the herd at 110 kg, due to an absence of homozygous mutant genotypes. In heterozygous form, the MSTN mutation was associated with a major increase in muscle depth and decrease in fat depth, suggesting that the deleterious allele was maintained at moderate frequency due to heterozygous advantage (allele frequency, q = 0.22). Knockout of the porcine MSTN by gene editing has previously been linked to problems of low piglet survival and lameness. This MSTN mutation is an example of putative balancing selection in livestock, providing a plausible explanation for the lack of disrupting MSTN mutations in pigs despite many generations of selection for lean growth.

Lameness is an important problem in livestock production for both animal welfare and economic reasons. A severe piglet lameness syndrome was observed in a commercial pig population. The incidence of the condition was low (6.3%), but was higher in affected families (~25%), which suggested a genetic basis and a recessive mode of inheritance. We discovered a region on Chromosome 15 where cases shared the same alleles that were different to healthy piglets. In this region, we discovered a mutation that causes a premature stop codon in the myostatin gene. Myostatin causes ‘double-muscle’ phenotype in several mammalian species. Piglets with two copies of this mutant allele suffer the lameness syndrome and do not survive post 40 kg live weight. However, those that carry a single copy have higher muscle depth and lower fat depth compared to wild type. We suggest that despite the negative consequences of the mutant allele in homozygous form, the mutation was maintained in the herd due to positive selection for this allele in heterozygous form. This is an interesting example of so-called ‘balancing selection’ and may explain why naturally occurring myostatin mutations have not previously been reported in pigs despite centuries of selection for lean growth.

Leg weakness is a heterogeneous condition causing lameness in pigs, and has negative impacts on both animal welfare and productivity [1, 2]. Significant heritability estimates have been reported for leg weakness traits [reviewed in 3], with moderate to high estimates in certain pig breeds, e.g. h2 = 0.45 in Landrace [4]. Several quantitative trait loci (QTL) have been identified for these traits, albeit they are generally not consistent across studies and breeds [5–8], which may be partly due to the heterogeneity of this condition. Interestingly, significant genetic correlations between leg weakness and other production traits (such as growth and muscle depth) have been detected [4]. Further, in a divergent selection experiment in Duroc lines, selection for high leg weakness was associated with a significant increase in muscle length and weight [9]. Taken together, these results suggest a degree of antagonistic genetic relationship between leg weakness and muscle growth traits in pigs, potentially explaining increases in the syndrome observed with intense selection for lean growth in recent decades. Deleterious alleles can be maintained at relatively high frequency in commercial livestock populations due to heterozygous advantage for traits under selection [10]. Examples of such balancing selection in cattle include a frame-shift mutation in the mannose receptor C type 2 gene (MRC2) responsible for crooked tail syndrome and also associated with increased muscle mass in Belgian Blue [11], and a large deletion with antagonistic effects on fertility and milk production traits in Nordic Red breeds [12]. In pigs, balancing selection at c.C1843T mutation in the ryanodine receptor 1 (RYR1) gene [13, 14] is likely to have caused an increase in incidence of porcine stress syndrome (also known as malignant hyperthermia) in the 1970s and 1980s, due to the association of the causative missense mutation with reduced backfat–a trait under selection. The ryanodine receptor 1 protein RYR1 acts as a calcium release channel in skeletal muscle. The c.C1843T mutation interferes with the proper function of the calcium release channel rendering homozygotes susceptible to stress induced malignant hyperthermia and death. Involuntary muscle contraction by the leaky channel may contribute to reduced fat levels in carriers and homozygotes [15]. More recently, a study showed an unique example of allelic pleiotropy in which one allele (deletion) is responsible for both increased growth and late foetal mortality by affecting two different genes [16]. We investigated genetic parameters and mode of inheritance for a leg weakness syndrome causing piglet mortality in a commercial Large White terminal sire line historically marketed by JSR Genetics under the product name “Yorker”. A monogenic recessive inheritance was observed, and homozygosity mapping was used to identify a genomic region on Sus scrofa chromosome (SSC) 15 associated with the trait. A mutation causing a premature stop codon in exon 3 of the Myostatin (MSTN) gene (similar to mutations causing the ‘double-muscling’ phenotype in cattle [17] was the outstanding functional candidate in the region. Comparison of MSTN genotype frequencies at birth and 110 kg supported this hypothesis, and carriers were shown to have significantly higher muscle mass and reduced fat depth than wild type homozygotes. Therefore, we propose that the MSTN mutant allele is highly deleterious in this population in homozygous form, but was maintained at moderate frequency due to heterozygous advantage. The overall prevalence of leg weakness in the commercial cohort was 6.3% (Table 1). When only affected litters were considered, the mean proportion of affected piglets was 23% ± 0.7. This within-litter prevalence is consistent with the expectation under the hypothesis of a single recessive locus (i.e. 25%). Complex Bayesian segregation analysis [18] suggested that almost all the variation was explained by a single locus with almost no environmental variation. The estimate of the additive effect was 0.50 ± 0.001 and dominance effect was -0.50 ± 0.001, which is in precise agreement with a recessive locus model. Estimates of heritability for the leg weakness syndrome (analysed as a binary trait on the underlying liability scale) was high (0.57 ± 0.10 in the sire and dam model) with low (0.17 ± 0.02 and 0.11 ± 0.02) but significant effects observed for permanent environmental effects due to the dam and litter, respectively (Table 2, S1 Table). Homozygosity mapping was used to map the underlying recessive variant, and the longest shared homozygous segment was a region of ~ 8.3 Mbp on SSC15. This unique region, contained 55 informative single nucleotide polymorphisms (SNPs) on the Illumina PorcineSNP60 SNP chip [19]. These 55 SNPs were homozygous in affected animals, but contained SNPs that were heterozygous or homozygous for the alternative allele in the unaffected animals, albeit there was one unaffected animal also sharing the homozygous segment. The segment started with SNP ALGA0110636 (rs81338938) at position 86,745,668 in the current Sscrofa11.1 reference genome assembly (Genbank assembly accession GCA_000003025.6) and finished with SNP H3GA0044732 (rs80936849) at position 95,062,143 (Fig 1A and 1B). The MSTN gene was located within this homozygous segment, from position 94,620,269–94,628,630. The 8.3 Mbp segment was assumed to represent an identical-by-descent (IBD) region likely to contain the underlying causative mutation, and became the focus of further analyses to discover and characterise this mutation. The SNP genotype used in the homozygosity mapping at SSC15 is given in S3 Table. To identify candidates for the causative variant, whole genome sequence data from ten cases, six presumed heterozygous carrier dams, and 22 controls were analysed. A total of 40 SNPs identified within the homozygous segment fitted the pattern of a potential causative variant assuming a recessive mode of inheritance. Functional annotation of these SNPs revealed that 19 were intergenic, 19 were intronic, 1 was in a pseudogene, and 1 caused a premature stop codon. There were also 10 InDels identified, 3 of which were intergenic and 7 of which were intronic (S4 Table). The outstanding functional candidate was a mutation in the third exon of the MSTN locus that results in the replacement of a codon for glutamic acid with a stop codon in exon 3 at position 274 (c.820G>T; p.E274*) (Fig 2A). The mutation is located in a region that is highly conserved across multiple species, and is predicted to result in truncation of the protein (Fig 2B). Functional annotation of all other variants detected in the IBD region did not reveal any other obvious causative candidates. This stop gain mutation was not present on the Ensembl variation database, accessed 25th July 2018. The whole genome sequencing raw reads are available in the NCBI, BioProject accession PRJNA506339. The MSTN c.820G>T mutation showed no statistically significant deviation from Hardy Weinberg equilibrium (HWE) in the 486 piglets sampled at birth (q = 0.22, α = 0.019, χ2 = 0.18, P > 0.05). Random mating of the dioecious population would be expected to result in a value of α that is slightly negative [20], and the value observed does not differ significantly from this value. However, the mutation deviated significantly from HWE at 40 kg (q = 0.17, α = - 0.180, χ2 = 12.2, P > 0.001) and at 110 kg (q = 0.17, α = - 0.210, χ2 = 11.45, P > 0.001). This was due to the loss of homozygous mutant piglets, with all but one dying (or being euthanized) shortly after birth, and the remaining piglet being euthanized due to poor health before it reached 110 kg. For detailed information, see S2 Table. The large change in α in a negative direction is quantitative evidence of the selective disappearance of homozygote genotypes, as opposed to disappearance as a result of selection against the allele itself. There were no significant changes in the relative genotype frequencies (GG, GT) over the total period or any sub-period from birth to the end of the test, confirming all changes in q and a are due to the selective loss of homozygotes, and that any other mortality or culling was at random with respect to MSTN genotype. The association of the porcine MSTN c.820G>T mutation with performance traits was assessed on 384 pigs which had completed a commercial performance test. Given the loss of the homozygous mutant animals the effect of the MSTN c.820G>T mutation was only estimated by the difference between the heterozygotes and the wild type pigs. The genotype means and differences are shown in Table 3, with the most notable of these being a major increase in muscle depth and a reduction in fat depth in the carriers (p < 0.001), with no evidence of a difference in live weight at 110 kg. Approximately 31% of the genetic variation in muscle depth and 18% of the genetic variation in fat depth was explained by this single variant (Table 3). The heterozygous animals had on average 5 mm increased muscle depth, and 1.7 mm decreased backfat depth when compared with wild type homozygous animals. Histological comparison of biceps femoris muscle between a piglet carrying the TT genotype (putative double MSTN knockout) and a heterozygous piglet revealed had significantly larger myofibre sizes in the homozygote (13.4 μm2 vs 11.5 μm2, P < 0.001), suggesting comparative myofibre hypertrophy. However, there was no difference in myofibre number between the two animals, which does not suggest hyperplasia (S1 Fig). A piglet leg weakness syndrome identified in a commercial line of Large White pigs showed moderate to high heritability, consistent with the upper range of estimates reported in the literature [2, 21, 22]. The within-litter incidence of the syndrome and the complex Bayesian segregation analysis both pointed to a monogenic recessive condition. Homozygosity mapping revealed a 8.3 Mbp segment on SSC15 as a putative IBD region likely to contain the causative variant. A single control animal appeared to also carry the putative IBD segment in this region, although it is possible that this animal was an affected animal misclassified as a control. The leg weakness syndrome is not lethal per se, but was debilitating under farm conditions and affected animals were typically crushed or failed to suckle effectively and were therefore typically euthanized. The outstanding functional candidate variant identified by whole genome sequencing of cases and controls caused a premature stop codon within exon 3 of the MSTN gene (Fig 2), and this variant was in HWE at birth but significantly distorted by 110 kg due to an absence of homozygous mutant animals. Knockout of MSTN by gene editing in previous studies has been associated with poor health and mortality of knockout piglets [23–26], including observations of a piglet leg weakness syndrome with an inability to stand or walk [24], strikingly similar to the syndrome described in the current study. MSTN is a member of the transforming growth factor beta (TGF-β) superfamily, which is highly conserved across species, and is typically expressed in developing and mature skeletal muscle as a key regulator of muscle growth [27]. The MSTN gene has been a gene of interest to animal breeders for over twenty years since the discovery of loss-of-function mutations in the cattle MSTN gene, which cause muscle hypertrophy leading to double muscling phenotypes [17, 27, 28]. Interestingly, these loss-of-function mutations causing double muscling are frequently due to premature stop mutations in the highly conserved exon 3 of MSTN [28], (Fig 2B). Further, in the case of Marchigiana beef cattle, the MSTN variant causing double muscling results is due to replacement of Glutamic Acid with a stop codon (the same change observed in the current study) in the equivalent position in the protein as observed in the current study [29]. This observation provides additional indirect evidence that the porcine MSTN mutation described herein is likely to cause the increase in muscle depth and decrease in fat depth associated with carrier animals. A histological analysis of muscle showed larger myofibre sizes in a homozygous TT (putative double MSTN knockout) than a heterozygous piglet suggesting comparative myofibre hypertrophy. A similar phenotype was observed in MSTN gene knockout piglets from a genome editing study [24]. However, unlike in another study of MSTN knockout piglets [25], where a greater number of myofibre nuclei were observed in mutants compared with the WT group, the density of myofibres in our study was comparable between the two animals which does not suggest hyperplasia (S1 Fig). In livestock breeding schemes, MSTN-inactivating mutations have been retained due to selection for increased lean growth associated with meat production, particularly for double-muscled cattle. However, despite many generations of selection for lean growth in pigs no such mutations have been reported previously. Associations between polymorphisms within the porcine MSTN gene and production traits have been shown in small-scale studies [30–32]. Further, a genome-wide association study in a sire line Large White population (related to the animals in the current study) detected SNPs on SSC15 significantly associated with rib fat between 81.1 and 87.8 Mbp [33]. While these SNPs are between 7 and 13 Mbp closer to the centromere than MSTN, two of these SNPs overlap with the region on homozygosity indicative of an IBD region in the current study. Therefore, it is plausible that this association may be due to linkage disequilibrium (LD) with the MSTN variant, or that other variants impacting performance traits exist on SSC15. The selection index applied in this line explicitly benefited animals with positive muscle depth and negative fat depth estimated breeding values, with analysis of the index showing that heterozygotes had a slight but significant advantage (p<0.05). Therefore, it is likely that the MSTN c.820G>T mutation was maintained at moderate frequency in this line despite its deleterious impact on piglet mortality due to heterozygous advantage for muscle and fat traits. Similar examples of balancing selection have observed to explain the maintenance of deleterious alleles at moderate frequency in commercial cattle [10, 12] and pig [16] populations. The results described herein have major implications for the targeted ablation of MSTN via gene editing to increase lean growth in pigs, and provide a plausible explanation of why MSTN loss-of-function mutations have not been previously reported in pigs despite decades of selection for lean growth. All samples were collected on a commercial nucleus farm as part of standard husbandry and management procedures in the nucleus herd, which complied with conventional UK red tractor farm assurance standards (https://assurance.redtractor.org.uk/) where sick or injured livestock that do not respond to treatment are promptly and humanely euthanized by a trained and competent stockperson. A piglet leg weakness syndrome was characterised in a Large White terminal sire line, historically marketed by JSR Genetics under the product name “Yorker”, reared in a nucleus herd, under standard conditions but with additional data recording. Leg weakness trait observations were collected on 19,006 piglets from 2007 to 2010, during which time a high incidence of the syndrome was detected. In addition, DNA was sampled on a further 119 piglets in 2011 and 486 piglets from the same population in 2012, of these 384 also had weight and carcass phenotypes available. The cohort born in 2007 to 2010 will be referred to as the commercial cohort and those in 2011 and 2012 as the survey cohort. Pedigree was available for all animals and spanned 9 generations. Details of the data and pedigree structure are presented in Table 1. The leg weakness in the phenotyped animals was visually classified as normal or affected (0 / 1, respectively). The leg defect is characterised by the piglet not being able to straighten its legs to stand, this being most apparent for the front legs, and being slow to suckle. Detailed post mortems were conducted on two affected individuals but the results were ineffective in providing additional diagnostic aids. The Online Mendelian Inheritance in Animals database (OMIA: http://omia.org/OMIA000585/9823/) was searched for previous reports of leg weakness in pigs but the syndrome observed here did not appear. These problems frequently resulted in death from either starvation or being crushed by the sow. Where the outcome resulted in poor welfare the piglets were euthanized in accordance with the Ethics Statement. Additional farrowing data were collected on the females in the commercial cohort including numbers born alive, dead or mummified, parity and year of birth. Body weights, ultrasound muscle and fat depths were obtained for individuals in the survey cohort that were retained in the herd until commercial slaughter age (when average weight is 110 kg). The ultrasonic measurements taken were the average depth of the m. longissimus dorsi and overlaying subcutaneous fat layer across the last four ribs. Initial studies to establish the genetic basis of the leg weakness syndrome were undertaken by fitting linear mixed models to the commercial cohort. The binary record of the syndrome was modelled on the observed 0 / 1 scale and the full model fitted was: y=Xβ+Z1u+Z2v+Z3w+e (1) where y is the vector of leg weakness phenotypes; β, a vector of fixed effects for month of observation (43 df), sex (1 df), parity (7 df), numbers in litter born dead (1 df) or alive (1 df), with design matrix X; u, additive polygenic effects, assumed to be distributed MVN(0, Aσ2a) with design matrix Z1; v, random litter effects assumed to be distributed MVN(0, Iσ2v), with design matrix Z2; w, maternal environment effects across litters assumed to be distributed MVN(0, Iσ2w) with design matrix Z3; and e residuals assumed to be distributed MVN(0, Iσ2e). Variations in this model were fitted replacing the individual polygenic effects with sire and/or dam effects, assumed to be distributed MVN(0, Aσ2s) and MVN(0, Aσ2d) respectively. All models were fitted using the ASREML software [34]. Likelihood ratio tests were used to assess the random effects. In addition, an analogous threshold mode with an underlying continuous liability was fitted with a logit link function and sire and dam effects associated with the pedigree, but not with individual polygenic effects, following recommendation of Gilmour et al.[34]. For the full model, the phenotypic variance was calculated as σ2p = σ2u + σ2v+ σ2w + σ2e. Where sire and dam models were used, σ2u was replaced by σ2s + σ2d. Heritability (h2) was calculated as σ2u/σ2p or 2(σ2s + σ2d)/σ2p depending on the model. The proportion of variance explained by the litter and maternal environmental effects were estimated as σ2v/σ2p and σ2w/σ2p, respectively. Heritabilities on the observed scale (0/1) was transformed to an underlying liability scale following Dempster and Lerner [35] using the observed prevalence of the syndrome in the commercial population which was 6.3% Inspection of the data suggested that the syndrome may be due to a single gene with the predisposing deleterious allele showing a recessive mode of inheritance, and this hypothesis was tested using chi square tests and segregation analyses. An initial test of a monogenic recessive mode of inheritance was carried out by pooling all affected litters, estimating the probability of being affected conditional on being born in an affected litter, and using chi-squared to test the null hypothesis that the probability of being affected was 0.25. A weakness of this approach is that some litters by chance will have no affected offspring, so a more complex segregation model was fitted. This model included all known phenotypes and pedigree data, and assumed a monogenic inheritance with environmental variation fitted by Gibbs sampling [18]. Ten affected animals from different litters and 10 unaffected full-sib controls from the commercial cohort were genotyped using the Illumina PorcineSNP60 SNP chip [19]. Only those SNPs that mapped to known positions on autosomal chromosomes and were not fixed nor completely heterozygous were retained. This left 38,570 segregating autosomal SNPs for the use in homozygosity mapping. Homozygous regions were assessed by alignment with the Sscrofa11.1 reference genome assembly sequence (Genbank assembly accession GCA_000003025.6). The genomes of the ten cases used for homozygosity mapping and six separate dams with affected offspring, assumed heterozygotes, were whole genome shotgun sequenced on an Illumina HiSeq 2500 platform as 125 bp paired-end reads. The dams were individually sequenced with a 10x genome coverage. The piglets were barcoded and individually sequenced at 3x coverage to achieve 30x coverage for the pool. The full sequencing output resulted in ~1.3 billion paired-end reads with an average of 48 million paired-end reads/sample for the piglets and 157 million paired-end reads/sample for the dams. Quality filtering and removal of residual adaptor sequences was conducted on read pairs using Trimmomatic v.0.32 [36]. Only reads where both pairs had a length greater than 32 bp post-filtering were retained, leaving a total of ~1.2 bn paired-end reads. Whole genome resequencing was followed by alignment to the Sscrofa11.1 assembly; using the Burrows-Wheeler Aligner with default parameters [37]. The average alignment rate of properly paired reads was of 92%. PCR duplicates were marked using Picard Tools (http://broadinstitute.github.io/picard). Variant calling was performed using the Genome Analysis Toolkit (GATK) HaplotypeCaller after read recalibration [38]. The parameter setting for the hard filters that were applied to the raw genotypes were: QualByDepth < 2.0, FisherStrand > 60.0, RMSMappingQuality < 40.0, MappingQualitySumTest < -12.5, ReadPosRankSumTest < -8.0. Candidate loci were identified from the sequences of the 16 animals from this study plus 22 additional Sus scrofa control sequences obtained from a public database [39], comprising 7 domesticated breeds (Duroc, Hampshire, Jiangquhai, Landrace, Large White, Meishan and Pietrain) and wild boar. The following criteria were used to identify candidate SNPs: It is worth noting that the limited sequencing depth means that both alleles will not be detected for all bases in all individuals. This limitation is particularly relevant to the reliable detection of heterozygous SNPs. A ‘kompetitive allele specific PCR’ (KASP) assay was designed by LGC Genomics (Teddington, UK) to enable genotyping of the mutation in the MSTN stop codon in large numbers of animals. The survey cohort of 486 piglets sampled at birth were genotyped by LGC. Of these, 265 remained as candidates for the final selection at 110 kg with a complete record of their performance test, together with another 119 pigs phenotyped at slaughter age (total n = 374). In both the 486 piglets and the subsequent subsets surviving to 40 and 110 kg, the frequency of the MSTN mutation (q) was calculated by counting, and the departure from Hardy Weinberg equilibrium the genotypes was estimated as α = 1-Hobs/Hexp [20] where Hobs is the observed heterozygosity and Hexp is the expected heterozygosity calculated as 2q(1-q). The significance of departure from true random mating genotype frequencies (α = 0) was tested using a chi-squared test. Associations between the MSTN c.820G>T locus and variation in performance traits in commercial testing conditions were examined in the survey cohort (n = 384). The performance test was started at a target weights of 40 kg, at an average age of 85 days, and continued for 54 (s.d. 12) days. The performance tests were performed over two distinct periods. The traits available were live weights and ages at the start and the end of the test, ultrasonic muscle and fat depths measured at the end of the test, days from birth to 40 kg and days from 40 to 110 kg. Univariate mixed models were fitted to these data in ASReml-R4 using the following model: y=1μ+X1β+X2b+Zu+e (2) where y is the vector of phenotypes; μ, a fitted mean, and 1, a vector of 1’s; β, a vector of fixed nuisance effects with design matrix X1; b, a scalar fixed effect for the effect of SNP genotype with design matrix X2, this has only 1 df due to the absence of homozygotes completing the test; u, additive polygenic effects assumed to be distributed MVN(0, Aσ2a), with design matrix Z; and e, residuals assumed to be distributed MVN(0, Iσ2e). For all traits, the sex of the piglet (1 df), parity of dam (4 df) and period of testing (1 df) were fitted as nuisance factors, together with cubic smoothing splines for the start date of the test fitted separately within each period [40]. The age at the time of measurement was fitted as covariate (1 df) for all traits other than days to 40 kg and days from 40 to 110 kg. The significance of fixed effects was assessed using Wald tests [ASReml Manual]. Samples of biceps femoris were taken from one myostatin homozygous knockout piglet (TT) and one heterozygous piglet (TG), and fixed in 10% neutral buffered formalin. After 24h of fixation, the tissues were processed to paraffin embedded blocks, and 4 μm sections were taken and stained with haematoxylin & eosin. One Tiff image was taken from each animal in an area where myofibres had been sectioned transversally using a Zeiss Axiocam 105 colour camera using the same light settings, and measurements were calibrated using a microscope graticle. For image analysis, images were loaded in FIJI (https://fiji.sc/), and for each image blood vessels were manually removed. The green channel was chosen, and the threshold adjusted to identify myofibres. This was followed by two binary functions (‘fill holes’ and ‘watershed’ consecutively), and the resulting mask was measured using the particle analyser, filtering particles below 5μm2 in size. The myofibre size was compared between the two samples using a Mann-Whitney test to assess hypertrophy. For assessment of hyperplasia, the total number of myofibres was expressed as the number of myofibres per 100 μm2.

10.1371/journal.ppat.1002161

Glycosaminoglycans and Sialylated Glycans Sequentially Facilitate Merkel Cell Polyomavirus Infectious Entry

Merkel cell polyomavirus (MCV or MCPyV) appears to be a causal factor in the development of Merkel cell carcinoma, a rare but highly lethal form of skin cancer. Although recent reports indicate that MCV virions are commonly shed from apparently healthy human skin, the precise cellular tropism of the virus in healthy subjects remains unclear. To begin to explore this question, we set out to identify the cellular receptors or co-receptors required for the infectious entry of MCV. Although several previously studied polyomavirus species have been shown to bind to cell surface sialic acid residues associated with glycolipids or glycoproteins, we found that sialylated glycans are not required for initial attachment of MCV virions to cultured human cell lines. Instead, glycosaminoglycans (GAGs), such as heparan sulfate (HS) and chondroitin sulfate (CS), serve as initial attachment receptors during the MCV infectious entry process. Using cell lines deficient in GAG biosynthesis, we found that N-sulfated and/or 6-O-sulfated forms of HS mediate infectious entry of MCV reporter vectors, while CS appears to be dispensable. Intriguingly, although cell lines deficient in sialylated glycans readily bind MCV capsids, the cells are highly resistant to MCV reporter vector-mediated gene transduction. This suggests that sialylated glycans play a post-attachment role in the infectious entry process. Results observed using MCV reporter vectors were confirmed using a novel system for infectious propagation of native MCV virions. Taken together, the findings suggest a model in which MCV infectious entry occurs via initial cell binding mediated primarily by HS, followed by secondary interactions with a sialylated entry co-factor. The study should facilitate the development of inhibitors of MCV infection and help shed light on the infectious entry pathways and cellular tropism of the virus.

Strong evidence suggests that Merkel cell polyomavirus (MCV or MCPyV) is a causative factor in the development of a large proportion of cancers arising from epidermal Merkel cells. While Merkel cell carcinoma is rare, it appears that infection with MCV is common, and many healthy people chronically shed MCV virions from the surface of their skin. In an effort to better understand the factors controlling MCV tissue tropism, we sought to characterize the cellular receptors that mediate MCV attachment to cultured cells. Several previously-examined polyomaviruses utilize sialic acid-containing glycolipids and glycoproteins to mediate cell binding and infectious entry. Our results show that, in contrast to other polyomaviruses, MCV does not require sialic acid-bearing glycans for attachment to cells, but instead uses a different group of carbohydrates called glycosaminoglycans for the initial attachment step of the infectious entry process. Interestingly, although sialic acid-bearing glycans are dispensable for initial attachment to cells, data using cells deficient in sialylated glycans suggest that sialic acids may form an essential element of a possible co-receptor that is engaged after the initial attachment of MCV to the cell via glycosaminoglycans.

The viral family Polyomaviridae consists of a diverse group of non-enveloped DNA viruses that infect humans as well as a range of other vertebrates. The family name is derived from the observation that murine polyomavirus causes tumors in various tissues in experimentally infected animals. The apparently broad tissue tropism of murine polyomavirus is consistent with the widespread distribution of its primary infectious entry receptors, a group of sialic acid-bearing glycolipids known as gangliosides [1]. Other well-studied polyomaviruses, such as the human polyomavirus BKV and its close relative, simian virus-40 (SV40), also employ gangliosides for infectious entry into cells (reviewed in [2]). Another BKV relative, JCV, has recently been shown to bind a specific sialylated pentasaccharide, known as LSTc, that decorates either proteins or gangliosides on a restricted range of cell types [3]. This is consistent with the much narrower cellular tropism of JCV [4], [5]. Although it has been suggested that initial attachment to sialic acid residues may be a universal infectious entry step for all polyomaviruses, the infectious entry pathways used by most members of the family have not yet been extensively investigated. Members of other non-enveloped virus families, such as the Parvoviridae, have been found to use a wide range of cellular receptors. For example, the primary cellular attachment receptors for adeno-associated viruses (AAVs) of the parvovirus genus Dependovirus range from gangliosides (bovine AAV; [6]), to protein-linked sialic acids (AAV4 and 5; [7]) or a very different type of carbohydrate side-chain, heparan sulfate (AAV2; [8]) (reviewed in [9]). Heparan sulfate (HS) is a type of glycosaminoglycan (GAG) that rarely contains sialic acid [10], but is instead characterized by specific patterns of N- and O-linked sulfate modifications [11]. AAV6 can bind to both sialylated polysaccharides and to HS on cells, and both interactions appear to modulate transduction into various tissues [12], [13]. In light of the Dependovirus precedent, the hypothesis that all polyomaviruses use sialic acid residues for initial attachment to cells should be viewed with caution. Seven of the nine polyomavirus species known to infect humans were discovered within the past four years [14], [15], [16], [17], [18], [19]. Perhaps the most intriguing of these new discoveries is a human polyomavirus species named Merkel cell polyomavirus (MCV or MCPyV). MCV is believed to play a causal role in Merkel cell carcinoma (MCC), a highly lethal form of skin cancer (reviewed in [20]). An emerging view is that, unlike BKV and JCV, which commonly infect the urinary epithelium, MCV establishes a chronic productive infection in the skin of most adults [17], [21], [22]. It remains unclear which of the dozen or so different cell types that can be found in the skin are the primary source of shed MCV virions. In an initial effort to better understand the cellular tropism of MCV, we set out to determine which receptors mediate initial attachment of the virion to the cell surface. A previous report by Erickson, Garcea and Tsai showed that recombinant MCV VP1 capsid protein subunits produced in bacteria can bind sialylated components of cell extracts, including the ganglioside GT1b [23]. Erickson and colleagues' data support speculation that MCV might follow an entry pathway similar to that of BKV, which has been shown to require GT1b or related complex gangliosides for infectious entry [24]. To further investigate this hypothesis, we employed MCV- and BKV-based reporter vectors (also known as pseudoviruses) as models for infectious entry into cultured cell lines [25]. To confirm the reporter vector-based results, we developed a system for titering the infectivity of native MCV virions. Our results support a model in which MCV uses GAGs, likely in the form of HS, as initial attachment receptors. The initial GAG-mediated binding appears to be followed by interactions between the MCV virion and sialylated host cell factors. The use of GAGs, such as HS, as attachment receptors for MCV infectious entry is strikingly reminiscent of a different family of non-enveloped viruses, the Papillomaviridae, which are exclusively tropic for keratinocytes, a cell type that forms the epidermal layers of the skin and mucosa. The results suggest a possible example of convergent adaptation to exploitation of the epidermis as an infectious niche. Hemagglutination assays (HA) are a classic method for investigating the interaction of virions with the cell surface. Hemagglutination is typically mediated by interactions between virion surface proteins and sialylated glycans displayed on red blood cells (RBCs). Erickson and colleagues have previously shown that recombinant MCV VP1 capsid protein subunits produced in bacteria can hemagglutinate sheep RBCs [23]. Using purified, fully-assembled MCV capsids produced in human cells (see below) [25], [26] we confirmed Erickson and colleagues' sheep RBC HA results (Figure 1). In contrast to MCV, BKV capsids did not display HA activity against sheep RBCs. Many virus types show differential abilities to hemagglutinate RBCs from different animal species. This is thought to reflect differences in the display of different forms of sialylated glycans or other binding targets on the surface of RBCs from different animals [27]. BKV HA assays typically employ human RBCs [28]. Consistent with previous results, recombinant BKV capsids induced robust HA of human RBCs. In contrast, MCV capsids showed a surprising lack of HA activity against human RBCs (Figure 1). The results show that MCV and BKV engage mutually distinct attachment factors on RBCs. When the MCV major and minor capsid proteins (VP1 and VP2, respectively) are co-expressed in human embryonic kidney-derived 293TT cells, they can spontaneously co-assemble around transfected reporter plasmids [26]. This results in the formation of reporter vector particles (also known as pseudovirions) that physically resemble native polyomavirus virions [26] and are capable of delivering encapsidated reporter plasmids to fresh target cells [25]. Similar systems have been developed for production of reporter vectors based on other polyomaviruses [5], [29] and for papillomaviruses [30]. Using MCV, BKV and human papillomavirus type 16 (HPV16) reporter vectors, we sought to identify candidate receptor or co-receptor molecules that are used by MCV for infectious entry into cultured cells. In a comparative analysis of MCV and BKV reporter vector transduction efficiency in over sixty different cell lines from various human tumors, we determined that the human lung epithelial cell line A549 was among the most MCV-transducible lines in the panel (unpublished results). A549 cells were chosen for initial experiments because they also have the convenient feature of being readily transducible with BKV and HPV16 reporter vectors, allowing comparisons to these better-studied virus types. To examine the binding of MCV or BKV capsids to cultured cells, we conjugated recombinant capsids to Alexa Fluor 488 to allow monitoring of cell binding by flow cytometry. The fluorochrome-conjugated capsids exhibited HA titers similar to unconjugated capsids (Figure S1A), suggesting that the dye conjugation process did not cause dramatic alterations in the cell binding properties of the capsids. Similarly, the dye conjugation procedure did not significantly affect the transducing potential of MCV reporter vectors (Figure S1B). In an initial series of experiments, we examined the binding of fluorochrome-conjugated MCV and BKV capsids to A549 cells. As shown in Figure 2 (and Figure S2), the binding of MCV to A549 cells was not significantly affected by pre-treatment of the cells with a broad-spectrum neuraminidase from Arthrobacter ureafaciens that is capable of hydrolyzing most forms of sialic acid linkage [31]. BKV capsid binding to A549 cells was, as expected, sensitive to neuraminidase. The transduction of a GFP reporter plasmid into A549 cells via MCV or BKV vectors in the presence or absence of neuraminidase mirrored the binding results (Figure 2 and Figure S2). Keratinocytes and melanocytes are the two most abundant cell types in the epidermal layer of the skin. Based on the speculative assumption that MCV might productively infect one of these cell types in vivo, we examined a variety of melanocyte and keratinocyte-derived cell lines for transducibility with MCV, BKV and HPV16 reporter vectors. The human melanoma-derived line SK-MEL-2, as well as primary adult human epidermal keratinocytes (HEKa cells) were found to be readily transducible with both MCV and BKV reporter vectors. Neuraminidase treatment of both SK-MEL-2 cells and HEKa cells resulted in inhibition of BKV transduction, but had little effect on MCV transduction (Figure S3), consistent with results observed using A549 cells. It is known that some sialic acids, such as the single sialic acid residue on the ganglioside GM1 (which serves as a receptor for SV40), are resistant to digestion with neuraminidase [32]. To address the possibility that MCV attachment to cells is mediated by a sialylated glycan that is resistant to neuraminidase, we used a cell line deficient in biosynthesis of sialylated glycans. The line, known as Lec2, is a Chinese hamster ovary (CHO)-based mutant that lacks a functional gene for SLC35A1, a CMP-sialic acid transporter required for sialylation of glycoprotein and glycolipid ectodomains in the lumen of the Golgi [33]. A control line, Lec2-mslc, was engineered to stably express a wild-type SLC35A1 allele. As seen in Figure 3, restoration of the SLC35A1 gene resulted in a 12-fold increase in BKV capsid binding, confirming that the introduced gene restored the production of sialylated glycans. In contrast to BKV, there was no effect on HPV16 and only a slight improvement in MCV binding to the Lec2-mslc line. The results indicate that MCV capsid attachment to this cell line is largely independent of sialylated glycans. All CHO-based cell lines are deficient in complex gangliosides, such as GT1b [34], [35]. Thus, restoration of the SLC35A1 sialic acid carrier to Lec2 cells would only be expected to restore sialylation of proteoglycans and simple gangliosides. Consistent with their lack of complex gangliosides, parental CHO-K1 cells (data not shown) and the CHO-based Lec2 cells with or without the restored SLC35A1 gene are highly resistant to BKV transduction (Figure 3). Despite the fact that MCV capsids readily bind to Lec2 cells, the line was surprisingly resistant to transduction by MCV reporter vectors (Figure 3). Reintroduction of the functional SLC35A1 allele rendered the line permissive for MCV transduction. A simple model that could explain the results would be that, while sialylated factors are not required for the initial attachment of MCV to the cell surface, the virus appears to require sialylation of a cellular factor for an entry step that occurs after stable attachment to the cell. The ability of MCV to transduce CHO-K1 and Lec2-mslc cells suggests that complex gangliosides are not necessary for MCV transduction. Indeed, while pre-treatment of cells with exogenous GT1b rescued BKV transduction of Lec2 cells, exogenous GT1b had little or no effect on MCV transduction (Figure S4). One way to reconcile the Lec2 line transduction results with the results observed for neuraminidase-treated A549 (Figure 2) would be to imagine that MCV attachment to a non-sialylated cellular factor allows the capsid to loiter on the cell surface until a hypothetical sialylated entry co-factor is regenerated after removal of the neuraminidase. Our past experience studying HPV binding and entry through interactions with HS led us to test the ability of purified protein-free GAGs, including heparin and chondroitin, to inhibit MCV infection. We found that, heparin can indeed block MCV transduction of A549 cells in a dose-dependent manner, with a 50% effective dose (EC50) of 4.2 µg/ml (Figure 4). Interestingly, moderate doses of heparin (∼1 µg/ml) appeared to increase the infectivity of MCV by up to two-fold in some experimental replicates. In contrast to heparin, chondroitin-A/C preparation was a much more effective inhibitor of MCV transduction (EC50 = 135 ng/ml) and did not appear to enhance MCV infectivity. Consistent with previous reports [36], [37], the transducivity of an HPV16 reporter vector was blocked more effectively by soluble heparin (EC50 = 1.2 µg/ml), while chondroitin-A/C only weakly inhibited HPV transduction. Comparable results were observed for MCV using the kidney-derived neuroblastoid line 293TT [30], [38] (heparin EC50≈12 µg/ml, chondroitin-A/C EC50≈0.3 µg/ml). The melanoma-derived line SK-MEL-2 and HEKa also showed similar GAG inhibition profiles for MCV reporter vectors (SK-MEL-2 heparin EC50≈3 µg/ml, chondroitin-A/C EC50≈0.1 µg/ml; HEKa heparin EC50≈2 µg/ml, chondroitin-A/C EC50≈0.05 µg/ml). As expected, BKV transduction of A549 cells was unaffected by either of the GAG compounds (Figure 4). Other soluble GAGs, such as dermatan sulfate and chondroitin-A, were found to be poor inhibitors of the transduction of all three reporter vectors on A549 cells (data not shown). Since chondroitin-A alone lacked inhibitory efficacy, it is likely that the chondroitin-C (chondroitin-6-sulfate) in the chondroitin-A/C preparation used here was primarily responsible for mediating inhibition of MCV transduction. Heparin has been shown to inhibit HPV entry by preventing binding of the virus to HS on the cell surface or extracellular matrix [37], [39]. To examine the mechanism through which heparin and chondroitin-C inhibit MCV entry into A549 cells, we measured the binding of Alexa Fluor-labeled MCV, HPV16 or BKV capsids to A549 cells in the presence of increasing concentrations of these GAGs. HPV and MCV binding to A549 cells was inhibited in a dose-dependent manner by both heparin and chondroitin (Figure S5), suggesting that these GAGs inhibit transduction, at least in part, by preventing cell attachment. As expected, heparin and chondroitin had little effect on BKV binding. Treatment of cell cultures with sodium chlorate inhibits the addition of sulfate groups to GAGs [40], reviewed in [41]. Although chlorate treatment can be toxic to some cell lines (for example, 293TT and HEKa cells do not appear to tolerate 50 mM chlorate), culture of A549 cells in 50 mM chlorate for several weeks did not appear to have noticeable effects on cell morphology or growth rate (data not shown). A549 cells maintained in 50 mM chlorate were extremely resistant to MCV transduction as well as binding (Figure 5). Chlorate-treated A549 cells were likewise resistant to HPV transduction. In contrast, BKV transduction of A549 cells was enhanced by chlorate treatment, confirming that the chlorate-treated cultures were healthy enough to support expression of reporter plasmids delivered via polyomavirus-based vectors. Similar chlorate treatment results were obtained with the melanoma cell line SK-MEL-2 (Figure S6). The data show that sulfate modifications, likely in the form of GAGs, are essential targets of MCV attachment and infectious entry. To examine the specificity of MCV interaction with different GAG types and to clarify the role of various GAG forms in infectious entry, cell surface HS and/or chondroitin sulfate (CS) were enzymatically removed using heparinase (HSase) and chondroitinase (CSase) enzymes. Enzyme activity and specificity was verified by immunofluorescent staining and flow cytometric analysis of cell surface HS and CS following treatment of A549 cells (Figure S7). Given the superior inhibitory effects of chondroitin-A/C relative to heparin, we expected that CSase treatment would have a greater impact on MCV binding and transduction than HSase treatment. Surprisingly, CSase treatment alone had little effect on MCV binding or transduction, while HSase caused a modest decrease in MCV binding and transduction (Figure 6). Combination HSase/CSase treatments synergistically inhibited MCV binding and transduction. The response of HPV to these treatments was very similar to MCV, while BKV was unaffected. Combination HSase/CSase treatments were also necessary to effectively inhibit MCV transduction of SK-MEL-2 cells and HEKa cells, confirming the importance of cell surface GAGs for MCV entry into skin-derived cell types (Figure S8). Neither CSase nor HSase alone significantly inhibited or enhanced MCV transduction of SK-MEL-2 or HEKa cells. A traditional approach to investigation of the role of particular GAG modifications in viral entry has been to compare the infectability of cell lines carrying mutations in the genes responsible for various steps in GAG biosynthesis [11], [42], [43], [44], [45]. These cell lines range from having no GAGs to simply lacking sulfate or other modifications at specific positions. We found that pgsA-745 cells, which are deficient in both HS and CS, did not bind MCV capsids efficiently and were transduced very poorly in comparison to the parental CHO-K1 line (Figure 7). Similarly, pgsD-677 cells, which lack HS but produce more CS than the parental cells, were highly resistant to MCV transduction. This suggests that HS, and not CS, is of primary functional relevance for MCV-mediated transduction of this cell type. Two other CHO mutant cell lines are deficient in the biosynthesis of specifically sulfated types of HS. Heparan sulfate modifications occur sequentially and, as a result, disruption of early modification events inhibits downstream modifications as well [11], [46]. Normally, the first step in HS modification involves N-deacetylation and N-sulfation of N-acetylglucosamine (GlcNAc) residues in the HS core chain. A subsequent modification step involves epimerization of glucuronic acid residues to iduronic acid. After these modifications, the HS sequentially becomes an appropriate substrate for 2-O-, 6-O- and 3-O-sulfotransferases. Thus, pgsE-606 cells, which lack GlcNAc N-deacetylase/N-sulfotransferase activity [42], produce HS that is deficient in all forms of modification. Another mutant cell line, pgsF-17, is deficient in 2-O-sulfotransferase function and thus expresses HS that carries N-sulfate and iduronic acid modifications, but lacks 2-O- and 3-O-sulfate modifications [45]. In addition to N-sulfated HS, pgsF-17 cells also produce HS carrying 6-O-sulfate modifications. MCV reporter vectors readily bound and transduced pgsF-17 cells but not pgsE-606 cells, (Figure 7) indicating that HS epimerization, N-sulfation and/or 6-O-sulfation are required to support MCV-mediated transduction, while HS 2-O- and 3-O-sulfation are dispensable. In comparison to HPV16, we found that the GAG type preferences of the two reporter vectors differ somewhat, as pgsF-17 cells show reduced HPV transduction, while the MCV reporter vector readily transduced this line. This result is consistent with previous reports indicating that 2-O-sulfate groups on HS are required for efficient transduction of cultured cells with HPV16 vectors [47]. Because many cell surface proteins display GAG-binding motifs, most cell types have substantial capacity to bind free GAGs non-covalently [48]. Non-covalently associated GAG chains, including exogenously-provided heparin, can participate in a wide variety of biological functions. For example, free heparin can serve as a functional “bridge” between vascular endothelial growth factor 164 and its co-receptor neuropilin 1 [49], [50]. Consistent with this type of bridging effect, we found that provision of exogenous heparin increased the transducibility of GAG-deficient pgsA-745 cells in a dose-dependent manner. At an apparent optimal concentration of heparin in the media of around 20 µg/ml, pgsA-745 cells became 10-fold more transducible than untreated parental CHO-K1 cells (Figure 8). MCV-mediated transduction of CHO-K1 cells was also enhanced by exogenous heparin, but the most effective dose was lower, presumably reflecting a reduced need for exogenous heparin due to the presence of native GAGs. Transduction of pgsD-677 and pgsE-606 cells was similarly enhanced by exogenously-supplied heparin, confirming that a heparin-like GAG is the primary missing factor required for MCV-mediated transduction of these HS modification mutant CHO lines (data not shown). Moderate doses of chondroitin-A/C also increased MCV transduction of GAG-deficient cells slightly, but the effect was very small in comparison to the effect of heparin (data not shown). In contrast to the GAG mutant CHO cell lines, MCV transduction of Lec2 cells was not rescued by exogenously-supplied heparin, suggesting that the block to MCV transduction in Lec2 cells occurs downstream of HS binding (data not shown). The rescue of MCV transduction of pgsA-745 cells by exogenous heparin correlated with an improvement in capsid binding to the cell (Figure 8). Surprisingly, neither pre-incubation of pgsA-745 cells with heparin nor pre-incubation of reporter vector stocks with heparin showed dramatic effects on MCV binding or transduction (data not shown). A possible explanation for this finding might be that one or more interactions in a hypothetical termolecular complex between heparin, MCV and cell surface binding targets may be of low overall affinity and relatively transitory. To test the idea that limitation of the ability of MCV to loiter on the cell surface might curtail access to a hypothetical sialylated co-receptor, we performed MCV binding and transduction assays on pgsA-745 cells supplied with exogenous heparin and treated with or without neuraminidase. Although neuraminidase treatment again had no effect on MCV binding in these experiments, the treatment modestly suppressed MCV transduction (Figure S9). The results are consistent with the idea that limitation of the ability of MCV to loiter on the cell surface reduces the engagement of a sialylated entry co-factor that regenerates after neuraminidase treatment. The results shown above suggest that attachment to cell surface HS is a critical step in MCV vector-mediated reporter gene transduction. However, the strong inhibition of MCV entry by soluble chondroitin-A/C raises questions surrounding the precise interaction of MCV with different GAG types. In an effort to better understand the physical interaction between MCV and GAGs, an ELISA-style binding assay was developed using a commercially available GAG-rich basement membrane extract (BME) derived from murine Engelbreth-Holm-Swarm tumor to coat the surface of 96-well protein-binding plates. Since the binding of VP1-specific antibodies might be affected by GAG-capsid interactions, we elected to detect bound reporter vector particles using Quant-iT PicoGreen stain [51] to render encapsidated DNA carried within the particles fluorescent. Increasing concentrations of HPV16 or MCV capsids in the BME-coated wells correlated with an increase in fluorescence (Figure 9A). BKV capsids bound the BME-coated wells very poorly (data not shown), suggesting the BME displays few binding sites for BKV. To determine whether capsid binding to the BME was the result of interactions with GAGs, BME-coated wells were pre-treated with increasing doses of HSase or CSase. Only HSase treatment of the BME resulted in major dose-dependent decreases in binding by MCV and HPV, and the highest concentration of HSase resulted in nearly complete abrogation of binding (Figure 9B), indicating that both viruses predominantly bind HS displayed on BME. The slope of the MCV capsid dose-response curve for binding to BME (Figure 9A) is relatively shallow, with a Hill coefficient of 0.64±0.14. A simple explanation for the occurrence of Hill slopes of less than one is that the assay is simultaneously measuring multiple binding interactions with differing affinities. This explanation is consistent with the fact that native GAGs are heterogenous and carry complex modifications that can dramatically alter their affinity for GAG-binding proteins. To circumvent this problem, we measured the ability of the more homogenous preparations of heparin and chrondroitin-A/C to interfere with the binding of MCV to BME. Interestingly, although high doses of chondroitin-A/C were able to entirely block the binding of MCV capsids to the BME, apparently saturating doses of heparin reduced MCV binding by only about 75% (Figure 9C). A model for these observations would be that the BME displays two distinct targets for MCV binding and the capsid carries two distinct glycan-binding motifs. Under this model, chondroitin-A/C is capable of blocking both of the glycan-binding motifs on the capsid surface, while heparin is capable of blocking only one binding motif. Systems for culturing MCV have not yet been developed. We have previously speculated that the relative inactivity of recombinant MCV genomes transfected into cultured cells may reflect regulation of the viral life cycle in a manner reminiscent of the extensive regulatory controls on the papillomavirus life cycle [17], [52]. Although we have previously shown that the genomic DNA of MCV primary isolates can drive the production of low levels of native virions after transfection into 293TT cells, the yield of native virions was relatively poor [17]. We found that virion yield can be improved substantially if the cloned genome is co-transfected together with expression plasmids encoding MCV small and large T antigen cDNAs (data not shown). To monitor the infectivity of native MCV virions, we generated a 293TT-based line, named 293-4T, which stably expresses the MCV small and large T antigen proteins. The stable line supports the replication of MCV genomes delivered by infection with native MCV virions, allowing monitoring of the infection using quantitative PCR (qPCR). The extent of MCV replication observed over several days varied between experiments, ranging from 7.5 fold to 70 fold. In separate experiments, we found that purified native MCV virions can be propagated in 293-4T cells (Figure S10). Using the native virion/293-4T infection system and enzymatic removal of cell surface GAGs, we confirmed that GAGs are required for MCV infection. qPCR analysis of 293-4T cells harvested immediately after viral inoculation and washing of cells treated with or without HSase/CSase revealed a 90–93% reduction in the number of cell-bound virions in the HSase/CSase treatment condition (Figure 10). The failure of the virus to bind efficiently to the HSase/CSase treated cells was reflected by a comparable decrease (76–83%) in the number of replicated viral genomes observed after 5–6 days of cell growth. To control for cell health after enzyme treatment, parallel experiments were performed to measure the number of genome copies for native BKV virions. As expected, native BKV infection was either unaffected by HSase/CSase treatment or, in one of the three replicates, modestly enhanced by the enzyme treatment. Native MCV virions were also used for a panel of additional confirmatory experiments (data not shown) on cell types that do not appear to support the replication of MCV genomes delivered by native virions. For these experiments we made the simplifying assumption that failure to bind the cell would result in failure to infect the cell. Native virion binding was measured using qPCR of viral genomes stably associated with cells. In an initial control experiment, we found that monoclonal antibodies specific for assembled MCV capsids [53] blocked the binding of MCV virions to A549 cells. Native MCV virions also failed to bind A549 cells in the presence of chondroitin-A/C. Treatment of A549 cells with sodium chlorate likewise prevented the binding of native MCV virions. We also found that native MCV virions readily bind both Lec2 and Lec2-mslc, confirming that native MCV does not require sialylated carbohydrates for attachment to cells. Taken together, the data show that native MCV virions exhibit binding and infectivity characteristics similar to MCV reporter vectors. Interaction with cell surface receptors is an essential first step in the process of viral infectious entry. Here we present multiple lines of evidence demonstrating that the initial attachment of MCV to cultured cells is mediated primarily by GAGs. Like HPV16, MCV binding and infectious entry can be antagonized by soluble GAGs and the attachment and infectivity of both viruses depends on the presence of cell surface GAGs. Although MCV capsids can bind to both CS and HS, experiments using CHO-based mutant cell lines indicate that N-sulfated and/or 6-O-sulfated forms of HS are specifically required for infectious entry. The handful of other polyomaviruses whose infectious entry pathways have been carefully studied all appear to utilize sialic acid-containing receptors for the initial cell attachment step of the infectious entry process [24], [54], [55], [56], [57]. Our studies show that sialylated glycans are not required for initial attachment of MCV to cultured cell lines. Further work is needed to determine whether MCV is unusual in this regard or rather provides an example of a common trait among the two dozen or so polyomavirus species that have not yet been subjected to extensive scrutiny. MCV appears to require a sialylated glycan for a post-attachment step in the infectious entry process. It remains uncertain whether this apparent requirement for sialylated glycans is due to indirect or direct effects. For example, failure to sialylate a cellular factor might impair a biological function or subcellular localization required to support MCV entry. In this scenario, MCV might not directly bind the sialylated glycan. A more intriguing possibility is that MCV directly interacts with a sialylated glycan during the infectious entry process. Although we found no clear evidence for direct interactions between MCV capsids and sialic acid residues on cultured human cell lines, Erickson and colleagues have previously shown that MCV VP1 capsomers can bind neuraminidase-sensitive factors in concentrated extracts of sheep RBCs [23]. Erickson and colleagues also demonstrated that MCV VP1 can bind GT1b when the ganglioside is presented at high concentrations in a cell-free flotation system. This suggests a possible scenario in which an unknown sialylated factor that resembles the glycan headgroup of GT1b serves as a co-receptor that the virion directly engages after initial attachment to the cell via HS. This would be analogous to the infectious entry of HIV, which generally requires the direct engagement of a chemokine co-receptor after initial attachment to a primary attachment receptor, CD4. It is tempting to speculate that the hypothetical sialylated co-receptor required for MCV entry might be a ganglioside. However, the fact that CHO-based lines, which are deficient in complex gangliosides [34], [35], are readily transducible by MCV reporter vectors would argue against this hypothesis. Further work is needed to determine which sialylated glycans, if any, MCV binds during infectious entry into human cells. Although our results using CHO cell lines indicate that HS is a more important factor than CS for MCV infectious entry, soluble heparin proved to be a less effective inhibitor of entry than chondroitin-C on all tested cell lines, including CHO (Figure 4). The results of the competitive inhibition experiments on basement membrane extracts (Figure 9) suggest a possible explanation for the apparently greater efficacy of chondroitin-C for inhibiting MCV infection. These analyses indicate that although MCV has higher affinity for heparin, chondroitin-C may be a better infection inhibitor because it blocks a secondary glycan binding site on the MCV virion surface that the highly homogenous heparin preparation cannot saturate. This model could explain the observation that chondroitin-A/C is a more effective inhibitor of MCV transduction. Although heparin doses >10 µg/ml effectively inhibited MCV transduction of several human cell lines, lower doses of heparin showed variable enhancement of MCV transduction of these lines (Figure 4 and data not shown). For CHO-based lines, heparin only enhanced infectivity, even at 20 µg/ml doses (Figure 8). The variable ability of heparin to either inhibit or enhance infectivity on various cell types is reminiscent of models for antibody-dependent neutralization or enhancement of the infectivity of flaviviruses (reviewed in [58]). In this model, antibodies that can neutralize flaviviruses when bound at high occupancy can also enhance infection when bound at low occupancy. It is thought that this effect reflects the ability of some antibodies to serve as a bridge between the partially occluded virion and antibody-Fc receptors expressed on the surface of some cell types. Analogously, heparin might serve as a bridge in a termolecular complex between heparin-binding proteins on the cell surface and heparin binding motifs on the surface of the MCV capsid. A similar model has recently been proposed for the infectious entry of human T-cell leukemia virus-1 (HTLV-1) [59]. It is also conceivable that, rather than forming a physical bridge between the MCV capsid and cell surface GAG-binding factors, heparin might induce a reversible change in the capsid structure that, in turn, permits direct binding of the capsid to a cellular co-receptor moiety. This would be reminiscent of conformational changes that are thought to occur in HPV capsids during infectious entry (reviewed in [60]). In either event, it is clear that the effectiveness of GAG inhibition of MCV reporter vectors can vary dramatically between cell lines. Resolution of this issue will require more detailed knowledge of the cellular factors that support the post-attachment steps of MCV infectious entry. Polyomaviruses have a long and complex history as suspected agents of human cancer [61]. The data implicating MCV as a cause of cancer in epidermal Merkel cells appears to be the strongest case yet described for a polyomavirus. That MCV particles can be isolated from human skin surfaces and cause tumors is reminiscent of certain aspects of HPV biology. Our data clearly show that both of these viruses require initial attachment to specific forms of HS, followed by transfer to poorly understood co-receptors for infectious entry to occur. Whether this is coincidence is difficult to determine, but it will be interesting to learn if these unrelated viruses share other aspects of their biology. A549 cells and SK-MEL-2 cells were obtained from the Developmental Therapeutics Program (NCI/NIH) and maintained in RPMI medium (Invitrogen) supplemented with 5% FBS (Sigma) and Glutamax-I (Invitrogen). HEKa (human epidermal keratinocytes, adult) were purchased from Invitrogen and maintained in Medium 254 supplemented with HKGS. CHO-K1 cells, pgsA-745, pgsD-677, pgsE-606, Pro5, and Lec2 cells were obtained from ATCC and maintained in DMEM (Invitrogen) with 10% FBS, Glutamax-I and MEM non-essential amino acids (D10 medium). pgsF-17 cells (a kind gift from Jeff Esko [45]) were maintained in D10 medium. Medium for the Lec2-mslc cells was supplemented with blasticidin S (5 µg/ml; Invitrogen). 293TT cells were maintained in D10 supplemented with hygromycin (250 µg/ml; Roche) and 293-4T were maintained in D10 supplemented with zeocin (100 µg/ml; Invitrogen) and blasticidin S (5 µg/ml; Invitrogen). Plasmids reported in this study will be made available through Addgene.org. The pMslc plasmid used to restore expression of SLC35A1 (accession number NM_006416) in Lec2 cells was created by transferring the human cDNA clone of SLC35A1 (OriGene, restriction enzymes XbaI and NcoI) into the expression cassette of pMONO-blasti-msc (InvivoGen, restriction enzymes AvrII and NcoI). 293-4T cells were created through two stable transfection steps. In the first step, 293TT cells were transfected with pMtB, an expression plasmid carrying the small T antigen ORF of MCV isolate R17a (GenBank accession number HM011555, [17]) in the expression cassette of pMONO-blasti-msc. Stable blasticidin-resistant clones were isolated by limiting dilution and analyzed for small T antigen expression by immunofluorescence microscopy and western blot using polyclonal serum raised against bacterially-produced MCV small t antigen fused to a maltose binding protein affinity tag (unpublished data). Stable expression of MCV small t antigen appears to be relatively toxic to 293TT cells and few clones maintained expression of the protein. One clone that stably expressed MCV small t antigen was super-transfected with a construct named pADL*, encoding MCV Large T antigen. The construct was generated by first silently mutating the splice donor and acceptor sites for the 57 kT isoform of MCV Large T antigen in the context of expression plasmid pCDNAclt206antigen1 (p2582), which was a generous gift from the Chang/Moore lab [62]. The Large T antigen ORF was also modified to remove the V5 epitope tag and proline residue 156 was mutated to serine to match the wild-type MCV consensus at that site. The modified T antigen gene was transferred into pMONO-zeo-mcs (InvivoGen) by restriction enzyme-based cloning. The polyclonal pADL* population was selected with both zeocin and blasticidin and the resulting stable line was named 293-4T. Nucleotide maps of plasmids used in this work and detailed protocols are available on our laboratory website <http://home.ccr.cancer.gov/Lco/>. MCV reporter vector stocks were produced using previously reported methods [25], [63]. Briefly, 293TT cells [30] were transfected with plasmids pwM2m [53] and ph2m [25] expressing codon-modified versions of the VP1 and VP2 genes of MCV strain 339. HPV16 reporter vectors were produced using the L1/L2 expression plasmid p16sheLL [36]. Production of BKV reporter vectors used a mixture of four novel plasmids, pwB2b pwB3b, ph2b and ph3b, which carry codon-modified versions of the capsid proteins of BKV genotype IV isolate A-66H (accession number AB369093, [64]). The capsid protein expression plasmids were co-transfected with a mixture of two reporter plasmids, pYafw [30] and pEGFP-N1 (Clontech) which express GFP from recombinant EF1α or CMV immediate early promoters, respectively. Forty-eight hours after transfection, the cells were harvested and lysed in Dulbecco's phosphate buffered saline (DPBS, Invitrogen) supplemented with 9.5 mM MgCl2, 25 mM ammonium sulfate (starting from a 1 M stock solution adjusted to pH 9), antibiotic-antimycotic (Invitrogen), 0.5% Triton X-100 (Pierce) and 0.1% RNase A/T1 cocktail (Ambion). The cell lysate was incubated at 37°C overnight with the goal of promoting capsid maturation [65]. Lysates containing mature capsids were clarified by centrifugation for 10 min at 5000×g twice. The clarified supernatant was loaded onto a 27–33–39% iodixanol (Optiprep, Sigma) step gradient prepared in DPBS with a total of 0.8 M NaCl. The gradients were ultracentrifuged 3.5 hours in an SW55 rotor at 50,000 rpm (234,000×g). Gradient fractions were screened for the presence of encapsidated DNA using Quant-iT Picogreen dsDNA Reagent (Invitrogen). The VP1 concentration of Optiprep-purified reporter vectors was determined by comparison to bovine serum albumin standards in SYPRO Ruby (Invitrogen)-stained SDS-PAGE gels. The MCV reporter vector stock contained 8.6 ng of VP1/µl, the BKV vector stock contained 4.3 ng of VP1/µl, and the HPV vector stock contained 2.9 ng of L1/µl. In various experiments examining reporter vector-mediated transduction, 0.2–0.4 µl of MCV stock, 0.3–0.6 µl of BKV stock, and 0.03–0.15 µl HPV stock was used per 96 well plate well. These concentrations generally produced between 5 and 25% GFP positivity in cell populations at the time of flow cytometric analysis. Recombinant capsids were produced as above, except that Benzonase (Sigma) and Plasmid Safe (Epicentre) nucleases were added to the lysis buffer at 0.1% each, with the goal of liberating capsids carrying fragments of cellular DNA [63]. Hemagglutination and basement membrane extract experiments used unlabeled capsids, while cell-binding studies used capsids covalently conjugated to Alexa Fluor 488 using previously-reported methods [53]. For production of Alexa Fluor 488 labeled capsids, a reporter plasmid encoding Gaussia luciferase (phGluc; [25]) was included in the initial transfection mixture. All conjugated capsid stocks were between 150 and 275 ng/µl and binding experiments used 0.2–0.4 µl of stock per 5×104 cells suspended in a volume of 100 µl. This generally achieved 10–30 fold fluorescence over background in flow cytometric analyses. Sheep blood in sodium citrate was purchased from Lampire Biological Products. Human type O+ blood was collected by finger prick immediately prior to use. Red blood cells (RBCs) were washed and suspended in PBS without calcium or magnesium (Invitrogen) at a final concentration of 0.5% (v/v). The suspension was chilled on ice in round-bottom 96-well plates then mixed with various doses of purified capsids and allowed to settle overnight at 4°C. A549 cells were plated at 7,500 cells/well in 50 µl of culture medium in a 96 well plate the day prior to infection. Stock solutions of porcine heparin (Sigma H4784), porcine dermatan sulfate (chondroitin sulfate B, Sigma C3788), bovine chondroitin sulfate-A (Sigma C9819), or shark chondroitin sulfate-A/C (Sigma C4384) were dissolved at 10 mg/ml in PBS (Invitrogen). The GAGs were serially diluted in media to 3× the indicated concentration and 50 µl was added to cells. Reporter vector stock was then added to the cells+GAG mixture in a volume of 50 µl. To minimize plate edge effects, the outer wells of the plate were not used for the assay and were instead filled with culture medium. Approximately 72 hrs post-infection, cells were incubated with trypsin to detach them from the plate and transferred to an untreated 96 well plate and suspended in wash medium (WM; DPBS with 1% FBS, antibiotic-antimycotic, and 10 mM HEPES, pH 8) and analyzed by flow cytometery for GFP reporter gene expression in a FACS Canto II with HTS (BD Biosciences). To calculate 50% effective inhibitory concentrations (EC50), Prism software (GraphPad) was used to fit a variable slope sigmoidal dose-response curve to values representing the percentage of GFP positive cells relative to untreated infected cells. Error bars represent the standard deviation for at least three independent experiments. Cells were dislodged using PBS supplemented with 10 mM EDTA, and then pipetted with an equal volume of WM. Fifty thousand cells were added to wells of an untreated 96 well plate and washed once with WM. Cells were then washed once with a dilution series of GAG in WM. Next, the same dilution series of heparin or chondroitin A/C in WM containing Alexa Fluor conjugated capsids was added to cells, such that each well contained about 60 ng of VP1 in the indicated concentration of GAG. These plates were incubated at 4°C for one hour, and then cells were washed 3 times in WM before measurement of their fluorescence by flow cytometry. A549 cells were cultured in D10 supplemented with 50 mM sodium chlorate (Sigma) for 2–6 days, then pre-plated overnight at 9,000 cells/well in 96 well plates. The next morning, half the plate was changed into medium without chlorate to allow regeneration of sulfate modifications. The other half of the plate was changed into fresh chlorate-containing media. Six to eight hours later, reporter virus was added in medium with or without sodium chlorate to maintain the concentration of chlorate present on the cells. Forty-eight hours later, the cells were fed by addition of 100 µl of media without chlorate. After a total of about 72 hours, cells were harvested for analysis of GFP expression by flow cytometry. For experiments examining the effects of neuraminidase treatment on reporter vector transduction, cells pre-plated in 96 well plates were washed and incubated with 50 µl of DPBS containing 70 mU of neuraminidase from Arthrobacter ureafaciens (NorthStar Bioproducts) per 5×105 cells for 1 hour at 37°C. The cultures were then inoculated with reporter vector stock and incubated for an additional 2 hours at 37°C. The cultures were then washed once and fed with 100 µl of culture medium. In some replicates, culture medium was added directly to the neuraminidase-containing PBS in the culture well. Removing or washing away the neuraminidase/reporter vector mixture did not appear to alter the experimental outcome. After three days, the cells were harvested and analyzed for GFP expression by flow cytometry. For binding studies, conjugated capsids were added to 5×104 neuraminidase-treated (or mock-treated) cells in suspension in an untreated 96 well plate and incubated for 1 hour at 37°C. The cells were then washed three times prior to analysis of fluorescence by flow cytometry. Heparinase I (50 units, Sigma) and heparinase III (5 units, Sigma) were solubilized in 100 µl each of resuspension buffer containing 20 mM Tris, pH 7.5, 50 mM NaCl, 4 mM CaCl2 and 0.01% BSA. The two enzymes were then combined. Chondroitinase ABC (2 units, Sigma) was solubilized in 200 µl of resuspension buffer. A549 cells plated the day prior at 7,500 cells/well in a 96 well plate, were washed once with digestion buffer (20 mM HEPES, pH 7.5, 150 mM NaCl, 4 mM CaCl2 and 0.1% BSA), and then treated with 2.5 µl of heparinase I/III stock, 2.5 µl of chondroitinase stock (or both) in 50 µl of digestion buffer. Cells were incubated in digestion buffer with or without enzyme for 2 hours at 37°C. Various doses of reporter vector stock were then added to the wells in 50 µl of OptiMEM-I (Invitrogen) and incubated for an additional 1 hour at 37°C. The cells were washed twice with culture medium, and then incubated in 150 µl/well culture medium for three days. Cells were then analyzed for GFP expression by flow cytometry, as above. For binding analyses, 2×105 cells dislodged using PBS and 10 mM EDTA, were treated with 3.5 µl each enzyme in digestion buffer for 1.5 hours at 37°C. Alexa Fluor conjugated capsids diluted in Opti-MEM were then added to the cells and incubated for an additional one hour at 37°C prior to washing and flow cytometric analysis. CHO-K1, pgsA-745, pgsD-677, pgsE-606, and pgsF-17 cells were plated at a density of 10,000 cells/well in 50 µl culture medium in a 96 well plate. Binding and infectivity studies to analyze the effect of exogenous GAGs were performed as above, except that cells were pre-plated and infected the same day in order to avoid changes in cell number resulting from slightly differing rates of growth. Cultrex BME PathClear (Trevigen #3432-005-02), a BME preparation derived from murine Engelbreth-Holm-Swarm tumor, was aliquoted and stored according to manufacturer's instructions. Black Microfluor 2 ELISA plates (Thermo) were coated overnight with 1 µg of BME per well in a volume of 150 µl. Coated plates were emptied and treated with 200 µl/well of 1× Blocker BSA (Pierce) in PBS. The block was incubated for 2 hours, with rocking, at room temperature. The plate was then washed twice with PBS plus 0.05% Tween 20 (PBS/Tween; BioRad). To examine direct virus binding to BME, a two-fold dilution series of HPV or MCV capsids beginning at 5 µg of VP1/well in 150 µl PBS/Tween was examined for binding to BME-coated plates. Binding reactions were conducted for two hours at room temperature, with rocking. To analyze the binding of capsids to BME for all experiments, the plate was washed three times with PBS/Tween, and then treated with 150 µl/well Quant-iT PicoGreen dsDNA Reagent (Invitrogen) in TE buffer supplemented with 0.1% Proteinase K stock (Qiagen). The plate was incubated in a 65°C oven for 1 hour, and then cooled for 15 min at room temperature in the dark before measuring fluorescence in a BMG Labtech POLARstar Optima microplate reader. To analyze the effect of enzymatic cleavage of GAGs on virus binding, a three fold dilution series of heparinase or chondroitinase (prepared as described above in the section on enzymatic removal of cell-surface sialic acids or GAGs) beginning with 4.5 µl of enzyme stock per well in 150 µl of digestion buffer was added to the prepared plate and incubated for 2 hours at 37°C. The plates were then washed twice with PBS/Tween, and 100 ng/well of capsids in 150 µl of PBS/Tween was added to all treated and mock-treated control wells. To measure competitive inhibition of capsid binding with heparin and chondroitin A/C, a five fold dilution series of each GAG, beginning with 100 µg of GAG per well was mixed with 50 ng of capsids in 150 µl of PBS/Tween, and then added to the BME-coated plate. MCV virions were produced by co-transfecting 293TT cells [30] with recombinant MCV isolate R17a genomic DNA, reconstituted by intramolecular re-ligation at 4 µg of plasmid DNA per ml using T4 DNA ligase (NEB). The re-ligated MCV genomic DNA was co-transfected with expression plasmids carrying the MCV Large T (pADL*) and small t (pMtB) antigen genes. Cells were expanded for five days after transfection and virions were harvested using the methods outlined above for recombinant capsid production. The virions were purified by Optiprep gradient centrifugation, as above, and fractions were screened for the presence of encapsidated DNA using Quant-iT Picogreen dsDNA Reagent (Invitrogen) and by western blot for MCV VP1. The characteristics of a representative stock of native virions are shown in Figure S10. 293-4T cells were detached with trypsin and 2×105 cells/well were added to an untreated 96 well plate. Cells were washed once with digestion buffer (see above section on enzymatic removal of cell-surface GAGs), and then incubated for 45 minutes at 37°C with or without 5 µl each of heparinase and chondroitinase stock solution in 150 µl digestion buffer/well. Next, native MCV virions (production described above) or BKV virions (kindly provided by Gene Major, NINDS, NIH [66]) diluted in 50 µl OptiMEM were added and the cell suspensions were incubated at 37°C for an additional 45 minutes. Cells were then washed once with culture medium and again with either PBS or culture medium. The PBS suspension was collected and frozen immediately, with the goal of establishing the initial baseline number of bound MCV genomes derived from the virus inoculum. The culture medium suspension was plated in a 24 well plate and cultured for 5 to 6 days. The cultured population was trypsinized and harvested for modified Hirt extraction ([67] protocol at our laboratory website) to isolate low molecular weight DNA. Baseline samples were also subjected to modified Hirt extract. One-fiftieth of the eluted DNA sample was used in a twenty microliter reaction with DyNAmo HS SYBR Green Kit (New England Biolabs) according to manufacturer's instructions in a 7900HT Fast RT PCR System (Applied Biosystems) with ROX reference dye. The primers targeting the MCV genome are 5′-GCTTGTTAAAGGAGGAGTGG-3′ and 5′-GATCTGGAGATGATCCCTTTG-3′. The BKV-specific primers are 5′-TGGTGCTCCTGGGGCTATTGC-3′ and 5′-GCCATGCCTGATTGCTGATAGAGG-3′. A dilution series of known quantities of MCV and BKV genomic DNA were analyzed simultaneously and used to form a standard curve and calculate the number of genome copies present in each sample. An average of 12 million copies of MCV DNA and 29 million copies of BKV DNA were measured from mock-treated baseline samples collected 45 minutes after inoculation of native virions. An average of 465 million copies of MCV DNA and 895 million copies of BKV DNA were measured 5 or 6 days later. Net values conclusively showing viral amplification were calculated by subtracting the baseline number of bound viral genomes observed 45 minutes after inoculation from the number of viral genomes observed after 5 or 6 days. Annotated nucleotide maps of all plasmids used in this work are posted on our laboratory website < http://home.ccr.cancer.gov/Lco/plasmids.asp>. The plasmids and their sequences will also be made available via Addgene.org. Accession numbers for previously-reported sequences are: MCV-R17a (HM011555), BKV-A-66H (AB369093), SLC35A1 (NM_006416).

10.1371/journal.ppat.1007957

Dietary zinc and the control of Streptococcus pneumoniae infection

Human zinc deficiency increases susceptibility to bacterial infection. Although zinc supplementation therapies can reduce the impact of disease, the molecular basis for protection remains unclear. Streptococcus pneumoniae is a major cause of bacterial pneumonia, which is prevalent in regions of zinc deficiency. We report that dietary zinc levels dictate the outcome of S. pneumoniae infection in a murine model. Dietary zinc restriction impacts murine tissue zinc levels with distribution post-infection altered, and S. pneumoniae virulence and infection enhanced. Although the activation and infiltration of murine phagocytic cells was not affected by zinc restriction, their efficacy of bacterial control was compromised. S. pneumoniae was shown to be highly sensitive to zinc intoxication, with this process impaired in zinc restricted mice and isolated phagocytic cells. Collectively, these data show how dietary zinc deficiency increases sensitivity to S. pneumoniae infection while revealing a role for zinc as a component of host antimicrobial defences.

Zinc deficiency affects one-third of the world’s population and is associated with an increased susceptibility to bacterial infection. Despite this, the molecular basis for how zinc deficiency compromises host control of infection remains to be understood. We show that dietary zinc deficiency impacts host tissue zinc abundances and its mobilization during infection by the major respiratory pathogen Streptococcus pneumoniae. Zinc acts as a direct antimicrobial against the pathogen, mobilized by phagocytic cells as a component of the innate immune response. Although immune activation and infiltration of phagocytic cells is unaffected by host zinc status, the lack of antimicrobial zinc compromises bacterial control in zinc deficient hosts. These findings highlight the importance of zinc sufficiency in resisting bacterial infection.

Zinc (Zn) is the second most abundant transition metal ion in humans and has crucial importance in immune function [1]. Although severe Zn deficiency is rare, mild to moderate Zn deficiency is estimated to affect one third of the world’s population [2]. The impact on human health accounts for ~1.4% of annual global mortalities, manifesting in a variety of adverse clinical outcomes including compromised immune defence and an increased susceptibility to infections [3–5]. Bacterial diseases associated with Zn deficiency are typically caused by respiratory or enteric pathogens, resulting in pneumonia or diarrhoea, respectively [6, 7]. Clinical trials of Zn supplementation therapies have shown that the morbidity and mortality of pneumonia and diarrhoea can be significantly reduced [7–10]. However, the efficacy of Zn supplementation strategies varies, highlighting the fact that the mechanism by which Zn contributes to resistance against bacterial infections remains unclear. Streptococcus pneumoniae (pneumococcus) is a globally significant human pathogen and the leading causative agent of bacterial pneumonia [11–14]. Pneumococcal disease is preceded by asymptomatic colonization of the nasopharynx, which occurs early in life with carriage rates reaching up to 90% in young children [15, 16]. Control of pneumococcal disease has primarily been addressed by vaccines that target the polysaccharide capsule of the bacterium. However, there are more than 90 distinct capsule variants (serotypes) with current vaccines providing protection against only a limited subset of serotypes [17, 18]. Further impacting the efficacy of vaccine strategies, disease monitoring has shown that the vaccine-included serotypes are being replaced by non-vaccine serotypes, against which current vaccines are not protective, and have not significantly reduced carriage rates [19, 20]. In addition, clinical interventions are increasingly hampered by the rising rates of resistance in S. pneumoniae to frontline antibiotic treatments [21, 22]. Pneumococcal pneumonia is highly prevalent in regions of endemic Zn-deficiency with mortality rates reaching up to 12% in children under 5 years of age [23, 24]. Although it is known that resistance to pneumococcal infection is significantly impaired by Zn deficiency [25, 26], the molecular basis for the control of S. pneumoniae proliferation and disease by Zn remains unclear. Recent studies have highlighted the importance of first row transition metal ions, such as Zn, in host control of infection [25–28]. Vertebrate hosts have a diverse array of mechanisms to restrict metal ion availability, processes collectively referred to as nutritional immunity [29], which limit microbial infection. Despite this, pathogenic bacteria can overcome host metal ion sequestration by using metal-recruiting proteins and metallophores [30]. S. pneumoniae employs a combination of ATP-binding cassette (ABC) permeases to facilitate the acquisition of essential manganese (Mn), iron (Fe) and Zn ions from the host environment [31–34]. Zinc uptake is facilitated by the Adc permease, which comprises the ABC transporter AdcCB and two extra-cytoplasmic Zn recruiting proteins, AdcA and AdcAII [32, 35–39]. The extra-cytoplasmic protein components, also referred to as solute binding proteins (SBPs), enable scavenging of Zn ions from the mammalian host environment with active import occurring via the AdcCB transporter. This pathway is essential for Zn uptake as Adc permease deficient mutants have been shown to be attenuated in murine models of pneumococcal infection [32, 37]. Zn homeostasis in the pneumococcus is tightly regulated and sensed by two independent systems: the MarR-family regulator AdcR, which regulates expression of the adc permease and associated Zn-scavenging genes [40, 41]; and the TetR-family regulator SczA, which regulates expression of the Zn-efflux pathway gene czcD [42]. The latter system is also implicated in bacterial survival in the host, with recent studies highlighting the contribution of Zn intoxication in microbial clearance [43]. Here, we investigated host Zn status and how it influences susceptibility to pneumococcal infection. By combining dietary Zn restriction with a murine model of pneumococcal disease, we determined host Zn abundance and how this changes in clinically relevant niches during the progression of infection. Gene transcription profiling revealed that host niche Zn levels were directly sensed by the invading bacteria resulting in changes to metal ion homeostatic processes in the pathogen. Activation and infiltration of immune cells in response to pneumococcal infection was also analysed, in combination with in vitro and ex vivo assessment of phagocytic cells and how Zn influences their ability to kill S. pneumoniae. This study reveals the link between dietary Zn intake and host resistance to bacterial pneumonia, demonstrating the antimicrobial activity of Zn in host niches against invading S. pneumoniae and in potentiating the efficacy of phagocytic cell killing of the pathogen. To investigate the role of host Zn in pneumococcal disease progression, we developed a murine model of Zn-deficiency. Dietary intervention achieved an approximate 70% reduction in serum Zn, compared to mice fed on the standard chow diet (12 μM to 40 μM; p < 0.0001, one-way ANOVA) (S1 Fig). Supplementation of drinking water with 250 ppm Zn was the minimum concentration required to restore serum Zn levels to that of mice fed on standard chow diet (S1 Fig). Analyses of sera of mice fed on the two diets, Zn-replete (250 ppm supplementation) and Zn-restricted (0 ppm supplementation), revealed that the impact on first row transition metal ion levels was restricted to Zn (S1 Table). We then investigated the impact of dietary Zn restriction on murine survival time after S. pneumoniae infection. Here, we performed an intranasal challenge with 1 × 107 colony forming units (CFUs) of the S. pneumoniae serotype 2 strain D39. Mean murine survival time decreased by ~30 hrs for mice fed on the Zn-restricted diet, relative to mice fed on the Zn-replete diet (40.2 hrs vs. 70.9 hrs; p < 0.01, Student’s t-test) (Fig 1A). Together, these data show that dietary Zn influences host control of pneumococcal infection in our murine model. Building on the above framework, we sought to ascertain how murine Zn levels influenced infection by S. pneumoniae. Bacterial burden was assessed at 24 and 36 hrs post-challenge in the clinically relevant niches of the nasopharynx, pleural cavity, lung and blood (Fig 1B–1E). At 24 hrs post-challenge, the Zn-restricted and Zn-replete mice showed no differences in bacterial burden in any of the niches (Fig 1B–1E). By 36 hrs, this had changed with the bacterial burden significantly increasing in the Zn-restricted mice. By contrast, the Zn-replete mice, whilst also demonstrating a general increase in bacterial burden relative to 24 hrs, showed a reduced extent of pneumococcal proliferation (Fig 1B–1E). This was most apparent in the nasopharynx and the blood, where the Zn-replete mice showed no increase in bacterial burden between 24 and 36 hrs, whereas in the Zn-restricted mice the pneumococcal burden increased by 1–2 orders of magnitude in the same niches (Fig 1B and 1E). Pneumococcal burden in murine lungs increased in both Zn-restricted and Zn-replete mice from 24 to 36 hrs (Fig 1D). However, the increase was significantly greater in Zn-restricted mice. The pleural cavity, assessed by lavage, was the only niche in which no differences were observed between the dietary groups at either time point (Fig 1C). We then assessed niche Zn abundance and compared the two dietary groups (Fig 1F–1I). In Zn-restricted naïve mice, the nasopharynx and blood serum had significantly reduced Zn abundance, by comparison with the Zn-replete naïve mice (Fig 1F and 1I). In contrast, no significant differences in total Zn were observed for the pleural cavity or lungs. Mobilisation of Zn is a physiological component of systemic inflammation control and the acute phase response [44, 45]. Consequently, we investigated how Zn abundance changed over the course of pneumococcal infection. We observed that tissue Zn abundance did not significantly change in any niche of the Zn-restricted mice. Abundance of Zn was also unaltered in Zn-replete mice in the nasopharynx and pleural cavity during infection (Fig 1F and 1G). However, consistently higher Zn levels (4-fold) were observed in the nasopharynx of Zn-replete mice, compared to Zn-restricted mice (Fig 1F). Although, blood serum Zn levels decreased in Zn-replete mice upon infection, consistent with prior studies [46], they remained significantly elevated by comparison with Zn-restricted mice at 24 (2.2-fold) and 36 hrs (3.3-fold) post challenge (Fig 1I). The abundance of Zn also increased in the lungs of Zn-replete mice at 36 hrs post-challenge and was significantly greater than Zn-restricted mice (1.2-fold) (Fig 1H). We then sought to investigate the flux of Zn in the lungs during infection in greater detail. Elemental bio-imaging was used to map the spatial distribution of Zn in murine lung tissue (Fig 1J–1M). This revealed relatively even distribution of Zn throughout the lung tissue of naïve mice with mean Zn concentrations of 1.35 ± 0.53 and 7.86 ± 1.08 μg.g-1, for the restricted (Fig 1J) and replete (Fig 1L) tissues sections, respectively. This trend and the relative differences were consistent across multiple distinct analyses (S2 Table) and corresponded with dietary intervention and the whole-organ ICP-MS analyses. Upon infection, the mean Zn concentrations for the restricted and replete tissue sections were 2.56 ± 0.25 and 9.55 ± 1.21 μg.g-1, respectively, (Fig 1K and 1M and S2 Table), but distribution of Zn within the tissue sections was no longer uniform with the emergence Zn-enriched regions [region of interest (ROI)]. In Zn-restricted mice, the Zn enriched ROI (ROI 1 in Fig 1K) had a mean Zn concentration of 4.95 ± 0.08 μg.g-1. This reflects a ~2-fold increase in Zn abundance compared to the mean tissue section concentration and a ~3-fold increase by comparison to the naïve tissue. Conversely, in a region distant from Zn enrichment the concentration of Zn was reduced to 1.72 ± 0.04 μg.g-1 (ROI 2 in Fig 1K). This was significantly less than the average tissue concentration suggesting that, in addition to Zn influx to the lungs, Zn was redistributed within lung tissue. The emergence of Zn-enriched regions was also observed in the lungs of Zn-replete infected mice (Fig 1M). Here, the Zn concentration increased to 15.75 ± 0.04 μg.g-1 in an enriched region (ROI 1 in Fig 1M), ~1.6 fold greater than the mean tissue section concentration. Similar to Zn-restricted mice, Zn abundance also decreased in areas distant from the enriched regions with a concentration of 4.19 ± 0.09 μg.g-1 determined for a representative area (ROI 2 in Fig 1M). These data show that regions of selective enrichment and depletion within murine lungs were not uniform across the tissues and most likely varied due to bacterial infection. However, the spatial complexities occurred irrespective of diet, with the extent of change directly related to dietary Zn intake which in turn influenced Zn flux. The location of S. pneumoniae within the lung tissue of Zn-replete mice was then assessed using a strain expressing GFP [47]. Elemental bio-imaging revealed that regions with increased Zn abundance (>10 μg.g-1) corresponded with regions of increased GFP fluorescence (ROI 1 and ROI 2 in Fig 2A and 2B; ROI 1 and ROI 2 in S2A and S2B Fig). In contrast, regions with lower fluorescence corresponded to regions where Zn abundance was not significantly different from the average tissue concentration (ROI 3 in Fig 2A and 2B; ROI 3 in S2A and S2B Fig) or naïve Zn-replete lung tissue (ROI 1 in Fig 2C and 2D; ROI 1 in S2C and S2D Fig) (< 10 μg.g-1). Taken together, these data show that Zn co-localises with the invading pathogen in murine lungs. Collectively, our analyses reveal an inverse correlation between host tissue Zn levels and the overall burden of pneumococci at 36 hrs post-challenge. Further, our data directly shows that Zn abundance increases in the lung upon infection and is predominantly localised in regions where S. pneumoniae is present and thus may contribute to control of pneumococcal proliferation. We then sought to ascertain whether control of pneumococcal proliferation could be linked to host Zn abundance. This was addressed by profiling the transcription of two S. pneumoniae Zn homeostasis genes, czcD and phtE, during infection of mice in both dietary groups. CzcD is the primary Zn efflux transporter of S. pneumoniae, and its transcription is highly sensitive to increased intracellular Zn abundance resulting in up-regulation of expression [42, 48]. Conversely, PhtE is a component of the pneumococcal Zn-starvation response and is also highly responsive to intracellular Zn limitation, but is down-regulated [32]. Here, we isolated bacterial RNA from murine lungs, the pleural cavity and blood of Zn-replete and Zn-restricted mice at 36 hrs post-challenge. Transcriptional analyses revealed that czcD was significantly up-regulated in pneumococci in the lungs (6.6-fold) and blood (4.8-fold) of Zn-replete mice (p<0.01, Student’s t-test) (Fig 3A). Concordantly, phtE was significantly down-regulated in pneumococci colonising these same niches (lungs, 3.0-fold, p<0.05; blood, 3.1-fold, p<0.01; Student’s t-test) of Zn-replete mice (Fig 3B). By contrast, transcriptional analysis of pneumococci from the pleural cavity showed no differences in Zn-dependent gene transcription between the two dietary groups (Fig 3A and 3B). These findings are highly consistent with the niche tissue Zn analyses, which showed consistently higher Zn abundance in the Zn-replete mice for the lungs and blood and no differences for the pleural cavity (Fig 1). Further, these data reveal that pneumococci in the lungs and blood of Zn-replete mice are directly exposed to increased Zn levels in these niches. To further examine how murine Zn abundance influenced bacterial infection we employed pneumococcal strains compromised, to differing extents, in Zn uptake (ΔadcA, ΔadcAII, ΔadcAΔadcAII) and Zn efflux (ΔczcD) [32, 48]. Comparison of the bacterial loads (wild-type D39 versus mutant strains) in the lungs, pleural cavity and blood of Zn-replete and Zn-restricted mice were then performed. Deletion of the Zn-recruiting proteins in isolation, i.e. ΔadcA or ΔadcAII, had minimal effect on pneumococcal proliferation in either dietary group (S3A–S3C Fig). Mutation of both genes (ΔadcAΔadcAII), which abrogates Zn uptake, significantly impaired invasive pneumococcal disease (Fig 3C–3E). The ΔadcAΔadcAII mutant strain was not detected in the pleural cavity of mice from either dietary group (Fig 3D). This finding, when considered with the tissue Zn determination (Fig 1G) suggests that the pleural cavity is a niche of low Zn abundance irrespective of diet. In Zn-restricted mice, the mutant strain was also significantly impaired for infection of the lungs and blood by comparison to the wild-type strain (Fig 3C and 3E). By contrast, deletion of the pneumococcal Zn efflux transporter had strikingly different effects. In the pleural cavity, the ΔczcD strain colonised mice from both dietary groups to the same extent as the wild-type (Fig 3D). In the blood of Zn-restricted mice, the ΔczcD strain had 17-fold greater abundance than that of the wild-type (Fig 3E) suggesting that Zn is poorly available to S. pneumoniae in this niche. This inference is consistent with serum Zn analysis (Fig 1I) of Zn-restricted mice and the reduced fitness of the ΔadcAΔadcAII strain in this niche. By contrast, the ΔczcD strain colonised the blood of Zn-replete mice to a lesser extent than the wild-type, consistent with increased niche Zn abundance in this dietary group (Fig 1I). In the lungs of Zn-restricted mice, the ΔczcD strain showed no significant difference in infection compared to the wild-type strain. However, in the lungs of Zn-replete mice the ΔczcD strain was 2.6-fold lower in abundance, compared to the wild-type strain. Decreased survival of the ΔczcD strain in the lungs of Zn-replete mice was consistent with the increased competitiveness of the ΔadcAΔadcAII strain in this niche and greater Zn abundance shown by ICP-MS and elemental bio-imaging (Fig 1H and 1J–1M). Taken together, these data indicate that during infection the abundance of Zn in the lungs increases and this is directly encountered by invading pneumococci and influences infection kinetics. Collectively, these data show that murine dietary Zn intake directly impacts the abundance of the metal ion encountered by invading pneumococci in various host niches and thereby, influences the Zn homeostatic pathways of the pathogen. Manganese (Mn) acquisition in S. pneumoniae is facilitated by the ABC permease, PsaBCA [35, 49]. The function of this pathway can be disrupted by Zn, which has been shown to compete with Mn for binding to PsaA and block its import at physiologically relevant concentrations (S3D Fig) [31, 50–52]. Transcription of psaBCA is negatively regulated by Mn abundance and positively regulated by Zn abundance via the DtxR family regulator PsaR (S3E Fig) [53]. Here, we examined psaA transcription to ascertain whether murine niche Zn abundances were influencing pneumococcal Mn homeostasis (Fig 3F). We observed that psaA transcription was significantly up-regulated in pneumococci isolated from the lungs and blood of Zn-replete mice, by comparison with Zn-restricted mice (lungs 3.0-fold, p<0.01; blood 2.6-fold, p<0.01), but not in the pleural cavity. Analysis of the Mn abundance in these niches revealed that mice from the two dietary groups had similar Mn levels in their respective tissues (S3F–S3H Fig). Further, elemental bio-imaging of murine lungs from the two dietary groups also showed the same abundances and distribution of Mn both prior to and post-infection (S3I Fig and S2 Table). Therefore, up-regulation of psaA is not attributable to differences in host tissue Mn abundance. Bioavailability of Mn can be altered during in vivo infection. One prominent mechanism is the S100-family protein calprotectin (S100A8/S100A9), which has a major role in nutritional immunity and has been identified in the bronchoalveolar lavage fluid of patients with pneumococcal pneumonia [54, 55]. Here, transcription of murine S100A8, which is co-transcribed with S100A9, was examined in the lungs and blood of mice from both dietary groups. In naïve mice, transcription of S100A8 in the lungs and blood was not significantly different between dietary groups. Upon infection, expression of S100A8 increased in both Zn-restricted and Zn-replete mice in the lungs and blood (S4A–S4D Fig). Although expression of calprotectin did not significantly differ in the lungs between the two dietary groups (S4A–S4C Fig), S100A8 was up-regulated to a greater extent in the blood of Zn-restricted mice (2-fold; p<0.0001) (S4D Fig). This increase in calprotectin transcription may arise from the higher bacterial burden in this niche in Zn-restricted mice. Collectively, these findings show that host Mn abundance is not affected by dietary Zn intake, nor does it appear to undergo significant changes in total abundance in the lungs or blood during infection. It therefore follows that host Zn abundance, influenced by diet, impacts pneumococcal Mn homeostasis during infection. In the context of Zn-replete mice, this results in a pattern of bacterial gene transcriptional responses consistent with Zn intoxication, i.e. up-regulation of psaA and czcD (S3E Fig), in the lungs of Zn-replete mice. Hence, these findings provide a plausible explanation for the lower bacterial burden observed in Zn-replete mice, by comparison to the Zn-restricted mice. Building on these findings, we sought to assess whether murine Zn levels influenced the immune response since Zn deficiency has previously been associated with dysregulated innate immune activation [56]. Here, we investigated the contribution of dietary Zn on the immune response to pneumococcal infection. Transcriptional profiling and cytokine analyses of the lungs and blood of infected Zn-replete mice revealed up-regulation of interleukins (IL) 1β and 6, and the chemokine CCL2, by comparison to naïve mice (Fig 4A–4F). This pattern of expression indicates an innate immune response associated with pneumococcal infection control and host defence [57–62]. Innate immune activation, and the extent thereof, is controlled by multiple negative feedback pathways [63]. NF-κβ, a crucial immune response modulator, has been shown to be negatively regulated by the IκB kinase (IKK) complex in a Zn-dependent manner. Prior studies of polymicrobial sepsis have shown that under conditions of Zn deficiency, the inability to supress NF-κβ activation leads to overproduction of the pro-inflammatory cytokines IL-1β and IL-6, resulting in septic shock [46]. Consequently, we examined whether Zn deficiency dysregulated this inflammatory axis in the context of pneumococcal infection by assessing transcription of nfkb-1 (p105 subunit) and nfkb-2 (p100 subunit). Comparison of the transcriptional profiles between the dietary groups revealed that expression of nfkb-1 was not significantly influenced by diet or by infection in either the lungs or the blood (S5A and S5D Fig). By contrast, during infection transcription of nfkb-2 was up-regulated to a similar extent in the blood, irrespective of diet, as well as in the lungs of Zn-restricted mice (S5B and S5E Fig). We further interrogated the role of NF-κβ by examining the phosphorylation status of the P65 (RelA) subunit of this protein, which is crucial for its activation. Quantitative immunoblotting revealed that there were no significant differences in the relative abundance of the phosphorylated form of NF-κβ between the dietary groups or upon infection (S5C and S5F Fig). Building on these analyses we compared the transcriptional and cytokine profiles between the two dietary groups post-infection. This revealed differences in the extent of up-regulated responses between the two dietary groups (Fig 4A–4F). Notably, IL-1β was elevated in the lungs and blood and CCL2 in the lungs of Zn-restricted mice, by comparison to Zn-replete mice (Fig 4B, 4C, and 4E). Taken together, these data show that Zn-restriction is associated with a greater inflammatory cytokine response, although without dysregulation of NF-κβ signalling, to pneumococcal infection, albeit in the presence of a greater bacterial load. Collectively, the above data indicate that the failure of Zn-restricted mice to control S. pneumoniae infection was not due to a failure in activation of the innate immune response. Consequently, we sought to examine whether, upon dietary restriction of Zn, innate immune cell recruitment was affected and the impact, if any, on S. pneumoniae clearance. We first examined the abundance of alveolar macrophages, monocytes and neutrophils in the Zn-restricted and -replete mice both prior to and 36 hrs post-infection. Residing alveolar macrophages play a primary role during the very early stages of infection by identifying and opsonizing the invading pathogen and activating the required immune response that will lead to an influx of phagocytic cells [64]. Our data shows that their abundance is not affected by dietary intervention (Fig 4G). Upon infection, a relative decrease in the abundance of alveolar macrophages was observed in both the Zn-restricted and Zn-replete mice, possibly due to the infiltration of other major phagocytic cells such as monocytes and neutrophils. Despite this, there was no significant difference between the two dietary groups. Monocyte and neutrophil abundance increased by more than 2-fold post-infection in murine lungs with no apparent differences between the dietary groups (Fig 4H and 4I). In the blood, monocytes and neutrophils were also observed to increase, ~4-7-fold and ~2-fold, respectively, in both Zn-restricted and Zn-replete mice in response to pneumococcal infection. Nonetheless, there were no significant differences between the dietary groups (Fig 4J and 4K) indicating that murine Zn status does not affect the relative abundances of phagocytic cells prior to or post-infection. Collectively, these data indicate that although dietary Zn deficiency does influence, to a minor extent, murine innate immune activation, when considered in the context of phagocytic cell recruitment the results do not suggest a generalised failure or dysregulation of the innate immune response to pneumococcal infection. Therefore, it follows that the increased pneumococcal burden arises from an impaired ability to prosecute clearance of the pathogen. Zinc has been shown to contribute to immune-cell mediated killing of various bacterial pathogens [45]. Although dietary Zn intervention did not impair immune-cell recruitment, we ascertained the effect on the intracellular Zn-status of the major phagocytic cells. Blood leukocytes were incubated with the cell permeable Zn-sensing probe, FluoZin-3 AM, followed by flow-cytometric quantitation of fluorescence in PMNs. These analyses indicated that Zn levels were significantly higher within polymorphonuclear (PMN) cells of Zn-replete mice compared to Zn-deficient mice (p<0.0001) (Fig 5A). To examine how intracellular Zn influenced pneumococcal survival, we pre-treated primary human polymorphonuclear cells and THP-1 derived macrophages with Zn (50 μM Zn-pyrithione) and subsequently examined pneumococcal loading compared to untreated cells. Here, we observed that Zn supplementation significantly reduced bacterial survival in PMNs (p<0.001) and THP-1 cells (p<0.0001) by comparison with untreated cells (Fig 5B). We then sought to ascertain whether the phagocytosed bacteria were being directly exposed to intracellular Zn. This was addressed by transcriptional profiling of the pneumococcal genes czcD, phtE and psaA, which are indicative of Zn intoxication, Zn limitation and Mn starvation, respectively (Fig 5C). In Zn-supplemented THP-1 macrophages, czcD and psaA transcription were significantly up-regulated by comparison to untreated controls (both p<0.05). In human PMNs, Zn supplementation resulted in up-regulation of czcD and down-regulation of phtE (p<0.05 and p<0.01, respectively). These data show that Zn-supplemented phagocytic cells expose the captured pneumococci to higher levels of intracellular Zn than untreated cells. We then investigated the impact of the increased Zn exposure on survival of phagocytosed S. pneumoniae. Here, we compared mutant strains deficient in either Zn-uptake (ΔadcAΔadcAII) or Zn-efflux (ΔczcD) in PMNs and THP-1 macrophages. We observed that the ΔczcD strain had reduced survival in both THP-1 macrophages and PMNs compared to the wild-type (p<0.05 and p<0.01, respectively; Fig 5D). By contrast, the ΔadcAΔadcAII strain was unaffected (Fig 5E). Collectively, our data show that restriction of dietary Zn reduces phagocytic cell Zn abundance, which in turn compromises the efficacy of S. pneumoniae killing due to the inability to employ Zn as a direct antimicrobial agent. Hence, these findings provide a mechanistic basis for the link between dietary Zn deficiency and increased susceptibility to pneumococcal infection. Zinc deficiency compromises immune defence and leads to increased susceptibility to bacterial infections [65]. Here we have built upon previous observations, wherein murine Zn deficiency has been shown to promote invasive pneumococcal lung disease [25, 26], to elucidate how host Zn contributes to antimicrobial control of bacterial infection. We report that dietary Zn restriction significantly impacts Zn homeostasis in most, but not all, of the niches colonised by S. pneumoniae. These changes were specific to Zn, with other transition metal ion abundances unaltered by dietary intervention. Niches in which tissue Zn levels were reduced showed increased bacterial burden with the overall impact of dietary zinc restriction reducing murine survival time in response to S. pneumoniae challenge. Concordantly, in niches where the dietary intervention did not impact Zn status, such as the pleural cavity, pneumococcal burden remained unchanged between dietary groups. Collectively, our findings show that host Zn contributes to control of bacterial burden, although its flux is spatially and temporally complex. Investigation of these changes revealed that they were niche-specific and influenced the progression of disease. In murine blood there was no level of reduction in serum Zn in either dietary group that prevented infection of this niche. Therefore, the pneumococcal Zn-specific uptake pathway is necessary and sufficient to acquire Zn from the blood, irrespective of host nutritional status and/or metal-withholding mechanism(s). Mobilisation of Zn from the blood into other tissues has previously been associated with host control of infection [5]. In this work, the only niche that had a concomitant increase in Zn abundance was the lungs. Elemental bio-imaging of this organ revealed infection-associated spatial flux of metal ions with the emergence of discrete regions enriched for Zn, but unaltered with respect to Mn. These analyses suggest that reduction in serum Zn levels, due to dietary intervention, impairs the mobilisation of Zn into the lungs during infection, thereby compromising host resistance. These findings are starkly different to observations from elemental bio-imaging of tissue abscesses associated with Staphylococcus aureus infection, wherein calprotectin-mediated withdrawal of Mn ions has been shown [66]. The formation of tissue micro-environments enriched for Zn, but unaltered with respect to other ions, suggests that regions of Zn intoxication occur within the lung during pneumococcal infection. Co-localisation, transcriptional and mutant infection studies show that the pathogen is directly exposed to these Zn-enriched regions, with pneumococcal metal ion homeostasis impacted in a Zn-dependent manner. Collectively, these findings indicate that Zn contributes to antimicrobial control within the lung during pneumococcal infection. The influx of antimicrobial Zn also provides a plausible explanation for the prior observation that murine control of S. pneumoniae lung infection was enhanced in calprotectin deficient mice [55]. The absence of calprotectin would permit Zn to act without sequestration by this host metal-withholding protein. However, this inference remains speculative as loss of S100A9 has also been shown, albeit in a distinct model, to diminish control of pneumococcal infection due to impaired recruitment of neutrophils to the lung and reduced expression of pro-inflammatory cytokines [67]. Nonetheless, the broader significance of this work is that it highlights not only how diet influences infection dynamics and elemental abundances in niche-specific ways, but also reveals that whole-organ analyses can mask dramatic and complex changes within tissue microenvironments that are relevant to host control of infection. Our findings suggest that mobilisation of Zn from serum to the lungs may be facilitated, at least in part, by phagocytic cells as a component of the innate immune response to infection. The Zn status of PMNs was significantly lower in Zn-restricted mice compared to the Zn-replete group. Despite the difference in cellular Zn status, the relative abundances of innate immune cells did not differ between the dietary groups prior to, or 36 hrs post infection. This was consistent with the transcriptional profiling of murine responses to pneumococcal infection, which revealed alterations in a subset of inflammatory-response associated genes between the dietary groups. This observation is notable as it is in stark contrast to prior studies of bacterial sepsis, wherein host Zn has been shown to play a key regulatory role in modulating NF-κβ-mediated inflammatory responses [46]. Here, restriction of dietary Zn intake resulted in only an increase in the activation of IL-1β and IL-6, but without infection resulting in a generalised dysregulation of inflammatory responses. However, as the pneumococcal burden was 1–2 orders of magnitude higher in Zn-restricted mice, the heightened immune activation is likely influenced by both host Zn status and bacterial burden. Nevertheless, the distinct differences in immune response between S. pneumoniae, a respiratory pathogen, and polymicrobial sepsis, induced by caecal ligation and puncture, reflect the differing modes of infection and host response. These data also show that control of pneumococcal infection was compromised in Zn-restricted mice, despite the lack of dietary impact on immune activation or phagocytic cell abundance. Impaired resistance to pneumococcal infection can be attributed to the contribution of Zn to phagocytic cell killing of bacterial pathogens [43, 68, 69]. Although the number and type of phagocytic cells present at the sites of pneumococcal infection did not differ between mice fed on the two diets, the Zn status of these cells was reduced in Zn-restricted mice. Thus, this study establishes a direct connection between host dietary intake, Zn status of phagocytic cells and their efficacy of bacterial killing. Investigation of how phagocyte Zn status influenced the killing of wild-type and mutant pneumococci revealed that bacterial Zn and Mn homeostasis was directly impacted, consistent with a Zn intoxication mechanism. Hence, this work links dietary Zn intake to the mechanism of bacterial control at a cellular level. Although Zn is not the sole determinant of bacterial control, acting in concert with other factors [68, 69], this work reveals how its deficiency compromises host control of infection and its antimicrobial mechanism against S. pneumoniae. Collectively, this study elucidates how dietary Zn intake influences susceptibility to pneumococcal infection. Our work shows that this arises from Zn-poor diets providing inadequate Zn for mobilisation into tissues, particularly during infection, and the associated failure of phagocytic cells to efficaciously prosecute killing of pneumococci. Hence, this work highlights the importance of ensuring dietary Zn sufficiency as a critical component of population-wide resistance against the global burden of pneumococcal disease in conjunction with vaccination and other antimicrobial strategies. All murine experiments were approved by the University of Adelaide Animal Ethics Committee (Animal Welfare Assurance number A5491-01; project approval number S-2013-053) and were performed in strict adherence to guidelines dictated by the Australian Code of Practice for the Care and Use of Animals for Scientific Purposes. Mice were anaesthetized by intraperitoneal injection of pentobarbital sodium (Nembutal; Rhone-Merieux) at a dose of 66 μg.g body weight-1. Mice were euthanized by CO2 asphyxiation. All experiments involving humans were performed in accordance with guidelines for low risk/negligible risk projects and approved by the Faculty of Science sub-committee of the University of Queensland Human Research Ethics Committee. Human neutrophils were obtained from healthy adult volunteers who provided informed written consent. Two-week old outbred female CD1 (Swiss) mice were fed a low Zn diet (Specialty Feeds, WA, Australia) for two weeks, with the acidified drinking water supplemented with either 0 ppm or 250 ppm ZnSO4. For challenge with S. pneumoniae strain D39, S. pneumoniae D39 pVA838-GFP, or its mutant derivatives, treatment cohorts, comprised of at least 5 mice, were anaesthetized by intraperitoneal injection of pentobarbital sodium at a dose of 66 μg.g body weight-1, followed by intranasal administration of 30 μL bacterial suspension containing approximately 1 × 107 CFU. The challenge dose was confirmed retrospectively by serial dilution and plating on blood agar. For survival experiments, mice were monitored regularly for signs of illness and euthanized once moribund. For determination of bacterial loads, at 24 or 36 hrs post-challenge, mice were euthanized by CO2 asphyxiation. Blood was collected by syringe from the posterior vena cava. The pleural cavity was lavaged through the diaphragm and the lungs through the trachea, both with 1 mL sterile PBS. Pulmonary vasculature was perfused by infusion of sterile PBS through the heart and lungs were subsequently excised. Lastly, the nasopharynx/upper palate was excised (nasopharyngeal tissue). Tissues were homogenized (Precellys Homogeniser) and all samples were serially diluted and plated on blood agar for bacterial counts. For metal ion determination by inductively coupled plasma mass spectrometry (ICP-MS), the tissues were homogenized, and all samples were boiled in HNO3 at the highest percentage achievable for the relevant niche, i.e. blood, and lung and nasopharyngeal tissue in 35% HNO3, and PL in 17.5% HNO3. Following removal of debris by centrifugation, samples were diluted to a final concentration of 3.5% HNO3. Metal ion detection by ICP-MS was perform on an Agilent 7500cx inductively coupled plasma-mass spectrometer (Adelaide Microscopy). For RNA extraction and qRT-PCR analyses of pneumococci in infected mice, blood, PL and homogenized lung tissues were diluted in Bacterial RNA Protect (Qiagen) and centrifuged at 400 × g for 5 min at 4°C to remove the majority of eukaryotic cells and tissue debris. Bacterial cells were then collected by centrifugation at 3000 × g for 15 min at 4°C. For analysis of transcription levels of phagocytosed pneumococci, THP-1 cells or isolated human neutrophils were coincubated with S. pneumoniae strain D39 for 90 min, then washed and treated with 10 μg.mL-1 penicillin and 200 μg.mL-1 gentamycin for 30 min. The bacterial RNA was stabilized, and the eukaryotic cells lysed by incubation with Bacterial RNA Protect (Qiagen) for 5 min at RT. For all sample types, the bacterial RNA was extracted using hot-phenol followed by further purification over a RNeasy spin column (Qiagen). The total RNA samples were treated with DNase I (Roche) and qRT-PCR was carried out using a SuperScript III One-Step RT-PCR kit (Thermo Fisher Scientific) on a LC480 Real-Time Cycler (Roche). Transcription levels of genes analysed, were normalized to those obtained for 16S rRNA. Primer sequences are available in S3 Table. Results represent the mean (± S.E.M.) of two independent experiments each comprised of treatment cohorts with at least 6 mice, with statistical analyses performed using a one-way ANOVA (Graphpad Prism V7.0d). Bacterial growth for metal content analyses was performed in cation-defined media (CDM), which corresponded to the C + Y media without transition metal-ion supplementation [51]. The base ion content of the CDM was ascertained by ICP-MS, as described previously [51]. CDM was then supplemented with MnSO4 and ZnSO4 at the concentrations specified. Cells were grown in 50 mL of culture to mid-log phase (absorbance at 600 nm = 0.3), harvested and prepared for analysis by ICP-MS, as described previously [51]. Peripheral blood was collected into heparinised tubes from mice by cardiac puncture. Red blood cells were lysed for 5 min at 37°C using mouse red cell lysis buffer, and cells were then extensively washed and then resuspended at 4 × 106.mL-1 in PBS with 1% BSA. Following transcardial perfusion, lungs were excised followed by treatment with 125 μg.mL-1 liberase and 100 μg.mL-1 DNase I in HBSS for 30 min at 37°C. The homogenized mix was passed through a 40 μm cell strainer and washed through using RPMI. Following washing, red blood cells were lysed at 37°C for 10 min using red blood cell lysis buffer. Cells were washed and resuspend in PBS with 1% BSA. The viability of lung and blood cells was examined using a LIVE/DEAD stain (Molecular Probes) at RT for 20 min. Following washing, Fc receptors were blocked by incubation with 400 μg.mL-1 murine gamma globulin (Rockland) for 15 min on ice before addition of fluorochrome conjugated monoclonal antibodies (CD45 (APC Cy7), CD11b (PE-Cy7), Gr1 (PE), MHCII (AF647), Siglec-F (BV421) [BD Biosciences] and F4/80 (AF488) [Invitrogen] or CD45 (APC), CD11b (PE-Cy7), Gr1 (FITC) and NK1.1 (PerCP Cy5.5) [BD Biosciences]) using the gating strategies shown in S6 and S7 Figs. Cells were washed after incubation for 30 min at 4°C. Cells were acquired on a BD LSR II and analysed using Flowjo software. Blood inflammatory monocytes: CD45+CD11b+F4/80+Gr1+, Blood neutrophils: CD45+CD11b+F4/80-Gr1+MHCII+ upon inflammation, Alveolar macrophages: CD45+F480+CD11blow/negCD11c+Siglec F+, Lung inflammatory monocytes: CD45+CD11b+F480+Gr1+ and Lung neutrophils: CD45+CD11b+F4/80-Gr1+MHCII+ upon inflammation. Peripheral blood was collected into heparinized tubes from mice by cardiac puncture. Red blood cells were lysed for 5 min at 37°C using mouse red cell lysis buffer, and cells were then extensively washed and then resuspended at 4 × 106 cells.mL-1 in PBS with 1% BSA. To determine the Zn status of PMNs, cells were labelled with Fluozin-3 AM (Molecular Probes) for 30 min at 37°C and washed in PBS with 1% BSA. Following this, Fc receptors were blocked by incubation with 400 μg.mL-1 murine gamma globulin (Rockland) for 15 min on ice before addition of fluorochrome conjugated monoclonal antibodies (CD115-APC, CD11b-BV421, Ly-6C-BV510, CD45-PE/Cy7 [Biolegend] and Gr1 (PE) [BD Biosciences]) and fixable live/dead reagent (Molecular Probes). Cells were stained for 30 min on ice in the dark, washed twice in cold PBS and then fixed in PBS with 1% PFA before analysis. Cells were acquired on a BD FACSAria and analysed using Flowjo software using the gating strategy shown in S8 Fig. PMNs were identified as live CD45+CD11b+Ly6CintGr-1hi. The fluorescence intensity of these cells with respect to Fluozin-3 staining was quantified. Prior to extraction of RNA using a RNeasy spin column (Qiagen), blood from the posterior vena cava was treated with red blood cell lysis buffer and lung tissues were treated with RNAlater solution (Ambion). All samples were DNase I treated (Qiagen) on the column during purification. For examination of SA100A8 transcripts, qRT-PCR analyses were performed using a SuperScript III One-Step RT-PCR kit (Thermo Fisher Scientific) on a QuantStudio 7 Real-Time Cycler (Thermo Fisher Scientific). Transcription levels of genes analysed were normalized to those obtained for ACTB. Data represent the mean (± S.E.M.) of two independent experiments with treatment cohorts comprised of at least six mice. The examination of mouse immune markers by qPCR was conducted using the TaqMan Array Mouse Immune Panel (Thermo Fisher Scientific) on a QuantStudio 7 Real-Time Cycler (Thermo Fisher). The TaqMan Arrays analyses were performed in triplicate, with each sample representing RNA pools from at least three mice. The results were analysed using the ThermoFisher Cloud Software. For the TaqMan Arrays analyses, data represent the mean (± S.E.M.), with statistical analyses performed using a one-way ANOVA. The IL-6 and IL-1β levels were determined in mouse plasma and homogenized lung tissue supernatant, all according the manufacturer’s recommendations (ELISAKIT.COM). The results were analysed on a PHERAstar plate reader (BMG Labtech). Data represent the mean of at least four independent biological samples (mean ± S.E.M.), with statistical analyses performed using a one-way ANOVA (Graphpad Prism V7.0d). Blood samples for analyses of Phospho-P65 and P65 immunoblots were collected from the posterior vena cava and treated with red blood cell lysis buffer, while lung tissues were homogenized using a tissue homogeniser (Precellys). Both blood and lung samples were further treated by sonication in a Bioruptor (Diagenode), following the manufacturer’s recommendations. Protein concentrations were determined (DC Bio-Rad protein assay; Bio-Rad), and 20 μg total protein was loaded into each lane of a 4–12% NuPage Bis-Tris gel (Invitrogen). After electrophoretic separation by SDS-PAGE, the proteins were transferred to a nitrocellulose membrane using an iBlot system (Thermo Fisher Scientific). Blots were incubated with rabbit anti-mouse phosphor-P65 or P65 (Cell Signalling Technology). This was followed by incubation with goat anti-rabbit IRDye 800 and analysis using an Odyssey infrared imaging system (LI-COR). Band intensities were measured using the manufacturer’s application software. Results are the mean ratios between phosphor-P65 and total P65 from four independent experiments (mean ± S.E.M.), with statistical analyses performed using a one-way ANOVA (Graphpad Prism V7.0d). THP-1 cells (ATCC TIB-202) were grown under atmospheric control (95% air and 5% CO2) at 37°C in complete RPMI medium (RPMI with phenol red [Gibco], supplemented with 10% fetal bovine serum, 10 mM HEPES, 30 μg.mL-1 penicillin and 50 μg.mL-1 streptomycin). Cell culture flasks (25 cm2; BD Falcon) were seeded with 3.5 × 106 THP-1 cells and differentiated by adding 100 ng.mL-1 phorbol 12-myristate 13-acetate (PMA; Sigma-Aldrich) and incubated for three days. Attached differentiated THP-1 cells (macrophages) were washed in complete RPMI and incubated with complete RPMI without added PMA to allow resting for a minimum of 2 days. Two hours before challenge with S. pneumoniae, a subset of macrophages was treated with 50 μM zinc-pyrithione. Following thorough washing, THP-1 cells were detached using 1 mL StemPro Accutase (Thermo Fisher Scientific), washed in Hank’s Balanced Salt Solution (HBSS; Thermo Fisher Scientific) and diluted to 1.1 × 105 cells.mL-1 in HBSS. Wild-type S. pneumoniae D39 and the czcD and adcA/AII mutants were grown overnight on blood agar plates at 37°C with 95% air and 5% CO2 and subsequently inoculated into C+Y media to an optical density at 600 nm (OD600) of 0.05. The cultures were grown until the OD600 reached 0.3, after which the cells were washed, resuspended in HBSS, and colony forming unit (CFU) counts determined by plating on blood agar. The macrophages and S. pneumoniae cells were co-incubated at a ratio of 1:10 for 90 min. The macrophages were then washed, and extracellular bacteria were killed by incubation with 200 μg.mL-1 gentamycin and 10 μg.mL-1 penicillin for 30 min. The macrophages were washed in HBSS without antibiotic and incubated for a further 60 min prior to analysis of intracellular bacteria by lysing the macrophages with 0.0625% Triton-X-100. The lysate was then plated onto blood agar. The CFUs were enumerated and corrected for input. Data represent the mean (± S.E.M.) of three experiments with five independent biological replicates. The statistical differences between pneumococcal survival in zinc-pyrithione treated and untreated THP-1 cells and between survival of wild-type and czcD or adcA/AII mutant pneumococci were examined using an unpaired Student’s t-test (Graphpad Prism V7.0d). Human neutrophils were purified from venous blood from healthy candidates using PolyMorphPrep (Axis-Shield, Norway) as previously described [70]. The neutrophil killing assay was performed using mid-logarithmic phase (OD600 = 0.25–0.3) S. pneumoniae diluted in RPMI + 10% heat-inactivated human serum that was incubated with the purified neutrophils at a multiplicity of infection of 10:1 (S. pneumoniae:PMN) for 60 min. The PMNs were then lysed by diluting them in distilled water and S. pneumoniae plated to enumerate colonies. PMNs were preloaded with Zn by incubation with 50 μM zinc-pyrithione (Merck) for 30 min at room temperature, followed by washing with 1× PBS prior to use in the killing assays. Data represent the mean (± S.E.M.) of three experiments with three independent biological replicates. The statistical differences between pneumococcal survival in zinc-pyrithione treated and untreated PMNs and between survival of wild-type and czcD or adcA/AII mutant pneumococci were examined using an unpaired Student’s t-test (Graphpad Prism V7.0d). LA-ICP-MS experiments were performed with a CETAC LSX-213 G2+ laser ablation system (Teledyne CETAC Technologies, USA) and coupled to a Thermo iCAP RQ ICP-MS (Thermo Fisher). Helium was used as the carrier gas (99.999% purity). The LA-ICP-MS system was tuned for maximum sensitivity prior to each experiment using the reference material NIST 612 “Trace Elements in Glass”. The ICP-MS was operated in standard mode and also tuned to minimize the formation of oxides by monitoring the oxide ratio (232Th16O+/ 232Th+, m/z 248/232 < 0.3%). Furthermore, the isotope ratios were monitored to confirm the absence of interfering polyatomic species. Instrument parameters are summarised in S4 Table. Images were analysed and created with the imaging software MassImager 3.49 (University of Muenster, Germany). Fluorescence images were analysed with a ZEISS AxioScan Z.1 slide scanner. Each specimen was analysed under individual conditions to maximise sensitivity. Fluorescence images were normalised to optimise contrasts. External calibration using matrix matched gelatine standards was used for the quantification of Zn. Calibration curves were constructed by plotting the signal intensity of 66Zn+ obtained by LA-ICP-MS against the standards concentration (1, 5, 10, 15, 20 and 25 ppm). The correlation coefficient as a measure of linearity was determined to be 0.9992. Using the obtained linear regressions, each data point (voxel) recorded by LA-ICP-MS was converted into concentrations. The exact Zn levels of the different gelatine standards were determined in triplicate by solution nebulisation ICP-MS. Statistical analyses were performed with the Prism software (GraphPad Prism V7.0d). Grouped data were analysed by one- or two-way ANOVA followed by multiple comparisons (Tukey or Sidak post-hoc tests). Non-grouped analyses were performed using the Mann-Whitney test. Kaplan-Meier survival curves were analysed by Log-Rank (Mantel-Cox) test. Statistical significance was computed at P ≤ 0.5. *P ≤ 0.05; **P ≤ 0.01; ***P ≤ 0.001; and ****P ≤ 0.0001. Number of animals and replicates for each experiment are indicated in the figure legends.

10.1371/journal.pcbi.1005892

Optimal occlusion uniformly partitions red blood cells fluxes within a microvascular network

In animals, gas exchange between blood and tissues occurs in narrow vessels, whose diameter is comparable to that of a red blood cell. Red blood cells must deform to squeeze through these narrow vessels, transiently blocking or occluding the vessels they pass through. Although the dynamics of vessel occlusion have been studied extensively, it remains an open question why microvessels need to be so narrow. We study occlusive dynamics within a model microvascular network: the embryonic zebrafish trunk. We show that pressure feedbacks created when red blood cells enter the finest vessels of the trunk act together to uniformly partition red blood cells through the microvasculature. Using mathematical models as well as direct observation, we show that these occlusive feedbacks are tuned throughout the trunk network to prevent the vessels closest to the heart from short-circuiting the network. Thus occlusion is linked with another open question of microvascular function: how are red blood cells delivered at the same rate to each micro-vessel? Our analysis shows that tuning of occlusive feedbacks increase the total dissipation within the network by a factor of 11, showing that uniformity of flows rather than minimization of transport costs may be prioritized by the microvascular network.

Arterial trees shuttle red blood cells from the heart to billions of capillaries distributed throughout the body. These trees have long been thought to be organized to minimize transport costs. Yet red blood cells are tightly squeezed within the finest vessels, meaning that these vessels account for as much as half of the total transport costs within the arterial network. It is unclear why vessel diameters and red blood cell diameters are so closely matched in a network that is presumed to optimize transport. Here, we use mathematical modeling and direct observations of red blood cell movements in embryonic zebrafish to show that occlusive feedbacks—the pressure feedbacks that alter the flows into a vessel when it is nearly blocked by a red blood cell—can optimally distribute red blood cells through microvessels. In addition to revealing an adaptive function for the matching of vessel and red blood cell diameters, this work shows that uniformity of red blood cell fluxes can be a unifying principle for understanding the elegant hydraulic organization of microvascular networks.

Vascular networks transport oxygen, carbon dioxide and sugars within animals. Exchange of both nutrients and gases occurs primarily in narrow vessels (e.g. capillaries) that are typically organized into reticulated networks. The narrowest vessels are comparable in diameter to red blood cells, forcing cells to squeeze through the vessels. Accordingly, hereditary disorders or diseases affecting the elasticity of cells and preventing them from contorting through narrow vessels can disrupt microvascular circulation [1]. The cost of blood flow transport in the cardiovascular system is thought to dominate the metabolic burden on animals [2]. The rate at which energy must be expended to maintain a constant flow of blood through a vessel is inversely proportional to the 4th power of the vessel radius. Red blood cells occlude the vessels that they pass through, further increasing the resistance of those vessels [3]. Accordingly capillaries and arterioles account for half of the total pressure drop within the network, and thus half of its total dissipation [4]. Experiments in which cells are deformed using optical tweezers, or by being pushed through synthetic micro-channels have shown that the extreme deformability of mammalian red blood cells requires continous ATP powered-remodeling of the connections between membrane and cytoskeleton. ATP released by deformed cells may induce vasodilation facilitating passage of cells through the narrowest vessels [5]. Thus, chemical as well as hydraulic power inputs are needed to maintain flows through microvessels [6, 7]. Why do micro-vessels need to be so narrow? A textbook answer to this question is that smaller, more numerous capillaries allow for more uniform vascularization of tissues—ensuring that “no cell is ever very far from a capillary” [4]. If smaller vessels are favored physiologically and red blood cell diameter acts as a lower bound on capillary diameters, then networks in which capillary diameters match those of red blood cells may be selected for. However, red blood cell sizes do not seem to be stiffly constrained—for example measured red blood cell volumes vary over almost an order of magnitude (19 to 160 femto-liters) between different mammals [8]. Since for a fixed capillary diameter, a small decrease in red blood cell diameter would greatly reduce rates of energy dissipation for red blood cells traveling through capillary beds [9], the evolutionary forces maintaining red blood cells and capillary diameters remain unclear. There is a natural analogy between occlusion of vessels by red blood cells, and the congestion that occurs in data or road networks [10, 11]. Efforts to construct efficient transport networks often focus on reducing congestion [10], yet although cardiovascular networks are thought to be organized to minimize transport costs (i.e. the viscous dissipation occuring within the network) [12, 13]; the presence of congestion at the finest scales seems at odds with minimizing these costs. Could the extreme deformation of cells passing through capillaries be an adaptive feature of the cardiovascular network? By directly stretching cells using optical tweezers Rao et al. [14] showed that deforming red blood cells releases oxygen. But it remains an untested hypothesis that squeezing cells so that they may pass through capillaries accelerates oxygen release, and therefore contributes to the function of the network. Indeed, earlier models suggest that alterations in the shape of the red blood cell surface decrease rates of oxygen exchange [15]. In this work we use mathematical modeling to reveal a previously unreported contribution of occlusive dynamics to the efficient functioning of the cardiovascular network. Moreover we link occlusive dynamics to a different open mystery of cardiovascular function. Specifically given that microvessels are distributed throughout the body and at very different distances from the heart, there is surprising consistency among measured flow rates in different capillaries [16–18] (with some exceptions [19]). Indeed consistency in flow rates may be biophysically necessary: if flow rate in a capillary is too low, the cells surrounding the capillary may not receive enough oxygen, but if the flow rate is too high, then red blood cells may leave the capillary bed before surrendering their oxygen to the surrounding cells. If the cardiovascular system is treated as an idealized symmetric branching network (such as in [2]) then flows are automatically uniformly partitioned at each level of the network, including among capillaries. But real cardiovascular networks have complex topologies, and it is not clear how the uniform flow can be achieved among billions of capillaries whose distances from the heart can range over several orders of magnitude. In this work we show that in the embryonic zebrafish, a model system for studying cardiovascular development [20], answers to these two questions may be closed linked. Tuned occlusion—i.e. small differences in the resistance that vessels present to cells—ensures that red blood cells are uniformly partitioned between the finest vessels within the zebrafish trunk. Although zebrafish red blood cells have quite different morphologies from mammalian red blood cells, the matching in sizes of red blood cell and narrow vessel means that occlusive dynamics occur in the zebrafish network. Our experimental observations confirm previous measurements that red blood cells are uniformly partitioned between fine vessels [18], yet in the absence of tuned occlusion, we demonstrate that the vessels closest to the zebrafish heart would receive 11-fold higher rates of flow that vessels furthest from the heart. In other words these vessels would act as hydraulic short-circuits. In further support of the hypothesis that occlusion is an adaptive feature of the network we calculate optimal occlusive dynamics—i.e. the distribution of occlusive feedbacks (the negative feedbacks each cell exerts on cells trying to enter the same vessel) that leads to the most uniform partitioning of red blood cells between the smallest vessels. The occlusive feedbacks within the real zebrafish conform very closely to this optimal distribution. Microvascular networks have been postulated to be organized to minimize the cost of transport (i.e. the total viscous dissipation associated with blood flow) [12, 21–23]. Certainly in larger vessels within both the arterial and venous vascular network, vessel radii appear to be organized to minimize dissipation [13, 24]. Yet, our results suggest that rather than eliminating cellular congestion, fine vessels make use of it. As a direct demonstration of the tradeoff between minimizing the cost of transport and tuning occlusion to route red blood cells uniformly, we show that the optimal distribution of occlusive feedbacks that uniformizes red blood cell partitioning increases hydraulic dissipation in the network 11 fold compared with a network in which the smallest measured occlusive feedbacks occur within each vessel. Thus, taken together, our results advance a potential new optimization principle—uniform routing of red blood cells—that may underlie the organization of microvascular networks generally. All animal experiments performed at Academia Sinica were approved by the Animal Use and Care Committee of Academia Sinica (protocol # 12-12-482). Zebrafish were bred and maintained at the UCLA Core Facility. Zebrafish experiments were performed in compliance with the Institutional Animal Care and Use Committees (IACUC) at the University of California, Los Angeles (UCLA) (under animal welfare assurance number A3196-01) To measure the red blood cell fluxes in zebrafish trunk vascular network we cultured double transgenic Tg(fli1:GFP; gata1:ds-red) zebrafish embryos, in standard E3 medium supplemented with 0.05% methylene blue solution at 28.5°C. In this transgenic fish line, fli1, a transcription factor associated with blood vessel morphogenesis is tagged with green fluorescent protein, causing the endothelial cells surrounding blood vessels to fluoresce green. Additionally, GATA-1, a transcription factor associated with erythrogenesis is tagged with red fluorescent protein, so that the red blood cells traveling through the GFP-labelled network fluoresce red. Zebrafish larvae were sedated with neutralized 0.02% tricaine solution(Sigma, MO) and mounted in 1-2% low melting agarose (Sigma-Aldrich, MO) for imaging. Erythrocytes were imaged at 4 day post fertilization (dpf) under a fluorescent microscope (Zeiss, Germany) with 50 ms exposure time. To measure detailed geometry and occlusive feedback of zebrafish trunk network we re-imaged a single 4 dpf zebrafish. We measured vessel lengths and radii from GFP-images taken under 10× magnification using a Zyla sCMOS camera on a Zeiss Axio Imager A2 fluorescent microscope. To measure the flow velocity, the same scope was used to take images in the DsRed channel at time intervals of 0.078 − 0.107 sec. Red blood cells were manually tracked in image sequences using ImageJ [25]. Flow is laminar within each zebrafish microvessel [26, 27]. The Womersley number [28] that characterizes the importance of unsteadiness effects in time-dependent flow, which for a vessel of diameter d, carrying blood with kinematic viscosity ν, and with heart rate f, is given by W o = 2 π f d 2 ν. Within the largest trunk vessels d ≈ 12 μm, the viscosity of whole blood is ν ≈ 5 × 10−6 m2/s [29], and the heart-rate is approximately f = 2 s−1, so Wo = 1.9 × 10−2 ≪ 1, meaning that we may neglect pulsatile effects. Flow is uniform along each vessel, except within an entry region whose length is ℓ ≪ Ud2/ν for a vessel of diameter d, through which blood travels at a speed U [30]. Maximum blood velocities are on the order of 0.3 cm/s [31], so using the diameter of the largest trunk vessels we obtain: ℓ ≪ 0.3 μm. Since the entry region is much smaller than the typical vessel length, we treat the flow in each vessel as being uniform along its length. Putting these ingredients together, we find that the flow through each vessel is inversely proportional to the resistance of the vessel, and the resistance may be calculated using Stokes’ equations (i.e. the equations for slow-creeping flows [30]) from the geometry of the vessel and from the number of red blood cells that it contains. Mechanistic models to predict the motions of red blood cells through micro-vessels or through microfluidic channels with comparable diameters have been developed in previous works [3, 32, 33]. Throughout this work we adopt a simple model for red blood cell occlusion in which the resistance of each vessel increases linearly with the number of red blood cells present. That is, if the number of red blood cells in a narrow vessel is given by n, then its resistance is given by an equation: R ( n ) = R 0 + n α c . (1) where R0 is the resistance of the vessel in the absence of red blood cells, i.e. is given by the Hagen-Poiseuille law relating the pressure drop and flow rate in a tube carrying viscous fluid, so that for a vessel of length ℓ and radius r: R 0 = 8 μ p l ℓ π r 4, where μpl ≈ 1cP is the viscosity of the non-red blood cell component of the flood. Here the parameter αc, which we call the occlusion strength in this paper, gives the increase in vessel resistance per red blood cell. Eq (1) represents a form of non-Newtonian rheology, the deviation of resistance from simple viscous fluid. In particular, the apparent viscosity of blood, i.e. R(n)πr4/8μpl ℓ, increases with hematocrit, i.e. with the concentration of red blood cells. Eq (1) can be derived from the micromechanical model of [34]. Indeed any model in which the pressure drop across the red blood cell is proportional to the velocity of the cell will produce a relationship like Eq (1), and so identical equations are also used to model the traffic of droplets or particles through microfluidic channels [35, 36]. In all of these models, αc, which we may think of as the intrinsic resistance of a single cell [34, 35, 37, 38], depends on the specific details of how the movements of cells, droplets or particles along the walls of the capillary or channel are lubricated. αc therefore depends on parameters that we can not measure experimentally, including the thickness and porosity of the glyocalyx that coats the endothelial wall of the capillary, as well as being sensitive to changes in vessel radius [33, 34] that are too small to be detected in light microscopy. It also depends upon the elastohydrodynamic deformation of both the cells and the capillary wall [32]. Accordingly we treat αc as a phenomenological constant, to be measured directly by fitting Eq (1) to real flow data. Specifically for each micro-vessel, we can measure both the velocity of flow within the vessel and the number of red blood cells that it contains. We note that due to the Fahraeus effect [36, 39] the velocity of red blood cells is in general larger than the flow velocity. However in human vessels whose diameters are comparable relative to human red blood cells to the diameter of the zebrafish vessel relative to the zebrafish’s red blood cells, the ratio of red blood cell velocity to whole blood velocity is less than 1.09 [39]. Hence we approximate the flow velocity by the velocity of the red blood cell in this measurement. The pressure difference across each vessel varies in time due to the variable pressure within the aorta, and also, less predictably because, since the resistance of all vessels changes from moment to moment, there are pressure feedbacks across the entire network. But we assume that there is an overall average pressure drop across each vessel that is constant in time but changes from vessel to vessel. Under conditions of time-independent pressure drop, the velocity of cell movement, v, in each vessel will be inversely proportional to the vessel resistance R(n). Thus Eq (1) predicts that a plot of 1/v against n will give a straight line, the slope of which can be used to calculate αc. Here we used the modeled flows in the fine vessels where no red blood cell is present to determine the intercepts, which can be calculated by using Hagen-Poiseuille formula (see Results, Absence of occlusion …). By regressing 1/v against n for each micro-vessel we calculate the variation of occlusive effects through the network (see S1 Text for more details of the regression). To study how varying occlusive effects between different microvessels may affect distribution of red blood cells, we incorporated Eq (1) into both continuum and discrete models of transport through the network. For continuum level modeling, we assumed that the concentration of red blood cells was a constant, ρ, in each vessel. Phase separation of red blood cells can occur when flows divide at vessel junctions—that is red blood cells may split in different proportions than whole blood [40]—but separation was not seen in our data (i.e. all Se vessels had the same average red blood cell concentration of number per volume), and cannot account for the uniformity of red blood cell flows, as we discuss in the Results section. Thus if the constant concentration (number/volume) of red blood cells is ρ, then a vessel of volume V is expected to contain n = ρV cells. Once each vessel in the network has been assigned a resistance, then we can solve for the flows in the entire network, by applying Kirchoff’s first law (conservation of flux) to calculate the pressure at each branching and fusion point, and then using the pressure difference across each vessel to calculate flows [12, 41, 42]. We discuss the geometry of the network and boundary conditions in the Results section. Since each micro-vessel is so small, typically each vessel contains no more than one or two cells at a time (but occasionally 3-5 cells were present in a vessel, see S2 Fig). For this reason we expected Poisson noise effects (i.e. fluctuations in the number of cells contained within each vessel) to influence red blood cell fluxes. We therefore built a discrete model, in which the trajectories of every single red blood cell traveling through the trunk network were directly simulated. Our discrete model is based on the droplet traffic model of [35]. Initially 990 cells are distributed uniformly through aorta according to measured zebrafish red blood cell concentrations [43]. At each step we calculate the resistance for each capillary by Eq (1), and then use the hydraulic resistor network model to calculate the whole blood flow rates within each vessel. We then let cells travel according to the predicted whole blood velocity in their vessel. Again we assume that the velocity of cell matches with flow velocity in Se vessels. The diameter of the dorsal aorta (DA) is larger and this mismatch may be significant in the DA. Since the cell velocity depends linearly on the flow velocity we expect this effect to increase the cell fluxes in all Se vessels equally and to therefore influence the partitioning of cells only weakly. While for precise prediction of cell fluxes the inclusion of this velocity mismatch will be necessary, here we are developing a minimal model that singles out the effect of occlusive feedbacks, and hence we assume that the cell velocity is the same as flow velocity in all vessels. When a cell arrives at a node of the network; i.e. at a point where a vessel branches into two, which vessel it enters is determined randomly by a Bernoulli process; that is the probability of cell entering a vessel is determined by the flow rate ratio of the two vessels. We therefore suppress the Zweifach-Fung effect [44]. The Zweifach-Fung effect characterizes the uneven distribution of red blood cells at a branching point, depending, amongst other factors, on stream lines at the branching point, and exibility of the cell [45–47]. Here we use a minimal model that neglects the Zweifach-Fung effect because we see that only occlusive feedbacks can account for uniform partitioning of cells. Indeed, we found no difference between the red blood cell concentration concentration (number / unit volume) of vessels in the rostral Se artery (2.88 × 10−4 ± 2.19 × 10−4 1/μm3) and in the caudal Se artery (2.18 × 10−4 ± 2.72 × 10−4 1/μm3). Flows are then recomputed for the new distribution of cells. Cells that leave the network, i.e. reach the end of one of the vessels within the simulated part of the network are immediately reintroduced into the network via the aorta. For each combination of parameters, we simulated 1000 s of red blood cell movement, with a time step of 0.1 s. Using fluorescence microscopy to track red blood cells meant that our measurement frame rate was too low to directly measure cell velocities within the aorta. So we fit total inflow into the trunk via the aorta to match the mean flux across all fine vessels to the experimentally measured mean flux. The 4 day post fertilization zebrafish trunk vasculature is topologically simple. Oxygenated red blood cells (henceforth RBCs) flow into the zebrafish trunk via the dorsal aorta (DA) and return the heart via the posterior cardinal vein (PCV). The microvasculature consists of a series of parallel intersegmental vessels (Se) that, if the vasculature were laid flat, would span between the aorta and cardinal vein like the rungs of a ladder (Fig 1A). Se are divided into intersegmental arteries (SeA) that connect to the aorta, and intersegmental veins (SeV) that connect to the posterior cardinal vein. SeA and SeV connect via another vessel called the Dorsal Longitudinal Anastomotic Vessel (DLAV), and in different parts of the DLAV, red blood cells flow toward the tail of the fish or toward its head. Red blood cells can enter the PCV by flowing along one of the SeAs, through a section of the DLAV, and then along one of the SeVs. Significantly, however, they can also flow directly from the DA into the PCV, since the two connect at the far end of both vessels in the tail of the fish. The positions of SeAs and SeVs vary from embryo to embryo [48]. In particular, SeVs and SeAs do not strictly alternate their connections with the DLAV. To form a model that does not depend on any specific A-V pattern we choose to connect SeAs and SeVs directly in a pairwise manner (Fig 1B), reducing the model to a bilaterally symmetric network in which no flow occurs in the DLAV (which can therefore be suppressed). Then we assign the same conductances for directly connected SeAs and SeVs and the same conductances for sections of DA as for the symmetric matching segments of PCV. Under these symmetry assumptions the pressures at the intersection of SeA and SeV is the same for each SeA/SeV pair, and we can shift this pressure to zero without affecting the calculations. Solving flows in this network reduces to solving flows in the lower half of Fig 1B with fixed inflow in the beginning of the aorta and zero pressures at the intersections between SeAs and SeVs, and between DA and PCV at the tail. As a first step we calculated the RBC flux in intersegmental arteries (SeA) with no occlusion or untuned occlusive effects and compared to experimental measurements. That is we approximated the resistance of each vessel using (1) with αc = 0 and treating the blood as a continuous phase, so that μpl replaced by μwb, the viscosity of whole blood (μwb ≈ 5 cP in zebrafish [29]). This reduced model serves as a motivation and readers interested in the full model may skip to Results, Occlusive feedbacks with …. We measured the lengths of each vessel directly from fli1a-EGFP images. SeAs were all assigned the same radius (2.9 μm), while because the DA tapers from the head to the tail, we independently measured DA radii between each SeA (see S1 Table). Although ultimately tuned variation in SeA radii will be one way to explain changes in occlusive feedbacks, these variations strongly affect the parameter αc in Eq (1) but have little effect on R0. To model flows without feedbacks we can therefore neglect SeA radius variations. We focus on the arterial half of the network made up of SeA and DA vessels. We identify the vertices in this network, i.e. the points at which vessel branch or fuse, as points i = 1, 2, … n, with respective pressures pi (Fig 1B). The number of SeAs, n, increases as the fish grows: for the 4 dpf zebrafish in our experiments n ranges from 9 to 13. For definiteness in modeling, we assume n = 12. If vertices i and j are connected by a vessel, with resistance Rij, then the total flow of blood along this vessel will be (pi − pj)/Rij. Solving for the flows in the network is equivalent to finding the pressures {pi}. For the zebrafish cardiovascular network we labeled vertices along the DA as i = 1, 2, …, n. A vertex, i = n + 1, represents the end of the DA in the tail of the zebrafish, where it connects directly to the PCV, and we label the vertices where the SeA meet the DLAV as i = n + 2, n + 3, … 2n + 1. At vertices i = n + 1, … 2n + 1, our symmetry boundary conditions require that pi = const., and we set arbitrarily the value of this constant to be 0. Thus only the pressures {pi ∶ i = 1, … n} need to be determined. We find these pressures by applying Kirchoff’s First Law (conservation of flux), at each point where the pressure is determined, i.e. ∑j ∈ n(i)(pi − pj)/Rij = 0, except at i = 1 (the vertex closest to the heart). At this vertex, ∑j ∈ n(1)(p1 − pj)/R1j = F, where F is the total supply of blood to the trunk which is fit to real data (see Materials and methods). All summations are taken over the neighbor set, n(i), i.e. over all vertices that are linked to i. The model of the zebrafish trunk microvasculature as an hydraulic resistor network (neglecting occlusive effects) follows many previous capillary network models (see e.g. [12, 41, 42]). The equations are formally identical to those for an electrical resistor network, with pressures replacing voltages, and flow rates replacing currents. Just as placing a wire across the terminals of a battery in an electrical resistor network will short circuit the network (i.e. divert current from higher resistance paths), the first SeA is predicted to receive a larger-than-even share of the blood flow from the zebrafish trunk, with flow rates decreasing exponentially rapidly with distance from the heart. In total there is a predicted 11-fold difference between the flows through the first and last SeA (Fig 1C). A simplified resistor network model that treats each SeA as having the same resistance, and assigns same resistances to each segment of DA between SeAs (i.e. ignores DA taper) quantitatively reproduces the exponential decay. To build the simplified model we assume that each segment of the DA has the same hydraulic resistance, and that each SeA has the same resistance. Using the measured mean radii and lengths, each DA has the same conductance, written as: κ1 = 1/R1 = 9.4 × 105 μm4 s/g, while all Se vessels have the same conductance, written as: κ2 = 1/R2 = 3.9×104 μm4 s/g. Then conservation of flow at vertex i = 2, …, n gives: - κ 1 p i - 1 + ( 2 κ 1 + κ 2 ) p i - κ 1 p i + 1 = 0 , (2) This is a second order recurrence equation with constant coefficients. Its general solution is: p i = C + ξ + i + C - ξ - i , (3) where ξ± are the roots of the auxiliary polynomial ξ2 − (2 + λ)ξ + 1 = 0, in which there is a single dimensionless parameter: λ = κ 2 κ 1 = 0 . 04. This equation has two roots, with ξ+ > 1 and ξ− < 1. In general C+ and C− must both be non-zero to satisfy our boundary conditions (namely pn+1 = 0 and F = κ2 p1 + κ1(p1 − p2)). However the two components give rise to exponentially growing and decaying pressures respectively. Typically the first term will negligible, except potentially in a small boundary layer region consisting of the vertices in the tail. Therefore over most vertices p i ∼ C - ξ - i, i.e. the pressure decays exponentially with distance from the heart, causing flows in the SeAs to decay exponentially as a result. For the real zebrafish network: ξ− = 0.81. Despite the simplification in geometry, the analytic formula agrees quite well with the solution to the full system of linear equations (compare gray and black curves in Fig 1C). Additionally, we note that for any λ > 0, it is impossible to organize an auxiliary polynomial without having one root ξ− < 1, so exponential decay in fluxes is an inescapable feature of the ladder-like geometry of the trunk vasculature. Although embryonic tissues receive oxygen primarily by diffusion through the skin [49, 50], vascular transport of oxygen becomes essential to embryo development after 2.5 weeks [51]. So we expect that a zebrafish with the large predicted difference in fluxes between trunk vessels would be disadvantaged. But because oxygen can diffuse through the zebrafish tissues, we first verified that the differences in fluxes predicted by the model lacking occlusive feedbacks would actually lead to differences in oxygenation in the trunk tissues. To do this, we modeled oxygen diffusion through the trunk by a reaction-diffusion equation, using the formulation and oxygen consumption coefficients derived by [52], and treating the vessels as oxygen sources (Fig 1D, and see S1 Text for details of the model). Note that our model includes only the contribution of oxygen perfusion from the blood to trunk oxygenation. For a real zebrafish at 4 dpf, these uneven oxygen levels would be compensated for by diffusion through the skin. However, our model shows that diffusion of oxygen within the zebrafish trunk can not compensate even at 4 dpf for uneven flows within the Se vessels. In contrast with the resistor network model, which predicts that the first Se vessel short circuits the network, measured RBC fluxes are nearly uniform between Se-vessels in living zebrafish. We tracked fluorescently tagged red blood cells moving through each of the 9∼13 SeAs within 6 living, sedated, zebrafish (see Materials and methods), over a total time interval of 26s per SeA. Fluxes in individual vessels varied greatly in time, due to the rapid change of blood pressures within the DA over the zebrafish cardiac cycle [31] and likely also due to nonlinear dynamics of the cells themselves within vessels [54], so the variability of flow rates was large for each vessel. However, mean fluxes varied little from vessel to vessel (Fig 2A). Each embryo exhibited variable RBC fluxes throughout the trunk. However the envelope of the lines of best fit for all six fish showed no consistent differences in RBC fluxes between first and last Se. Specifically from the six sets of zebrafish data we used bootstrapping method (generating replicate measurements for each Se vessel from the measured mean and standard deviation over all six fish) to estimate regression statistics. The gray envelope in Fig 2A shows the 95% confidence interval on all regressions thereby generated. We found that over all regressions m = 0.012 ± 0.032 (mean ± standard deviation), showing no statistically changes in RBC flux from vessel to vessel. There are two major ingredients missing from the hydraulic resistor network model that could explain the anomalies between the predictions of that model and the real zebrafish flow rate data: phase separation of red blood cells and occlusive feedbacks effects [40, 55]. Separation occurs because red blood cells do not divide in the same ratios as whole blood when blood vessels branch: When a red blood cell passes through a junction at which a vessel branches into two daughter vessels of different sizes, it is more likely to enter the larger daughter vessel than would be expected based on the ratio of fluxes in the two daughter vessels. Phase separation cannot explain the uniform distribution of red blood cells seen across real zebrafish microvessels: to correct for an 11-fold difference in flow rates between first and last Se vessels, there would need to be an 11-fold increase in hematocrit between these vessels, in the absence of occlusive effects (since then hematocrit must increase exponentially to compensate for exponentially decreasing flow rates). This was not observed in our experiments. Indeed Pries et al. [33] explicitly fit measurements of red blood cell fluxes at the branch points of blood vessels, and parameterized the amount of phase separation that occurred. When we applied their model to the zebrafish microvasculature, only minute variations in hematocrit were predicted between different SeAs (see S1 Text and S1 Fig). By contrast, we observed large feedback effects within the SeA, i.e. the presence of a red blood cell reduces the flow in the vessel and hence the entering probability of the next cell. We individually tracked red blood cells in a single 4dpf zebrafish, and plotted the inter-entry intervals, i.e. the times between consecutive red blood cells entering each vessel, condensing data from all SeAs since all vessels have the same approximate rate of blood cell entry (see Fig 3). In the absence of feedbacks, we would expect the inter-entry times to be distributed randomly, i.e. as an exponential random variable. Our red blood cell tracking shows that a single red blood cell passes through an SeA in a mean time of 0.3s. Inter-cell entry intervals larger than 0.3s (i.e. cell entries into unoccupied SeAs) were distributed exponentially (see the inset to Fig 3). However, inter-entry intervals less than 0.3s were not exponentially distributed, and we saw far fewer cells entering vessels less than 0.3s apart (i.e. while the vessels were already occupied by other cells) than would be expected based on the exponential distribution (Fig 3, main panel). In fact we found that inter-entry intervals less than 0.3s were approximately uniformly distributed. These observations are suggestive of a negative feedback mechanism, whereby entry of a red blood cell into an SeA reduces for some time afterward the probability of another red blood cell entering the same SeA. We tested for statistical support for the presence of negative feedback by two methods. First, we extrapolated the exponential fit for time intervals greater than 0.3s to estimate the number of cells that should enter the SeA between 0 and 0.3s, if cell entries into SeA were independent events. For the zebrafish trunk data this amounted to 533 cell entries, compared to the 261 actually observed, and the difference in statistically significant by the Fisher’s exact test (p = 3.9 × 10−22 against independence). Secondly, we fit the distribution of cell entry times directly, to compare an independent model with an exponential probability density function (pdf), with a model in which the feedbacks were modeled by a composite pdf, with uniform probabilities for inter-cell entry intervals less than 0.3s, and an exponential pdf for cell entry intervals greater than 0.3s. The Akaike Information Criterion score corrected for small samples (AICc) [56] for the composite pdf was 4.02 × 103, whereas the AICc for the pdf assuming independence was 4.07 × 103, supporting the inclusion of feedback effects. In mammals red blood cells must squeeze through narrow capillaries. Passage through these narrow vessels is facilitated by specific cellular adaptations—cells are un-nucleated, and have a biconcave shape, assisting cell deformation. By contrast zebrafish red blood cells are almost spherical and are nucleated. However, since the diameters of SeAs are closely comparable to red blood cell diameters (both 6 μm), we speculated that zebrafish red blood cells may also fit tightly within the SeAs. We directly measured these dynamics by measuring the dependence of the velocity within a SeA upon the number of red blood cells contained in the vessel (see Materials and methods). Velocities within each SeA are affected by the phase of the cardiac cycle as well as by nonlinear cell-cell and cell-wall dynamics [57, 58], so there is large variability in these velocities, and pressures are also affected by changes in conductances throughout the network (Fig 4A). However, in each vessel we found that 1/v increased linearly with the number of cells, n, consistent with the model for occlusion in Eq (1). In physical terms, when a cell travels through a vessel, it almost blocks the vessel. Because a large pressure difference must be maintained over the cell to push it forward through the SeA, flow within the vessel slows, so that fewer red blood cells enter a vessel once it contains a cell. We measured the occlusive effect within each SeA, i.e. the parameter αc in Eq (1) by fitting the slope of the graph of 1/v against n (see Fig 4A). The intercept of the graph is given by the speed within the SeA when it contains no red blood cells. We get that speed from the model of flow without occlusive feedbacks, described above, so there is only one free parameter to be estimated for each SeA. Eq (1) represents a form of non-Newtonian rheology, since it gives that the resistance of each vessel increases as hematocrit (i.e. n) increases. The parameter αc represents the intrinsic resistance per cell [34, 35, 37, 38], and it depends on the relative size of the cell and SeA (i.e. how tightly the red blood cell must be squeezed to travel along the vessel), cellular deformation due to elastohydrodynamic effects [32], as well as upon the surface chemistry of both. In particular, [34] built a physically informed model of cells moving through a narrow vessel, including both cell deformation, and interactions between the cell and the vessel glycocalyx: a polymer brush that covers and lubricates the endothelial lining of the vessel. They found that αc is highly sensitive to biophysical parameters: the thickness of the glycocalyx layer and its porosity (i.e. to the concentration of polymer), as well as to small changes in vessel radius. To assay the potential for controllability for the occlusive effect, αc, we measured αc independently in each of the twelve SeAs, in all cases by fitting the data for the dependence of 1/v upon n (see S1 Text for more details of the fit). The experimentally measured occlusion strength decreased from first to last SeA (Fig 4B), over a range of αc = 3.0 × 10−7 ∼ 2.8 × 10−5 g/μm4 s. In physical terms, red blood cells occlude closer vessels to the heart more than distal vessels. These values are consistent with the range given in Secomb et al.’s model [34] in which αc could range from αc = 1.8 × 10−7 to 1.6 × 10−5 g/μm4 s. Our measurement of αc also agrees with an earlier theoretical model of Secomb et al.’s which did not consider glycocalyx (αc = 4.7 × 10−7 ∼ 3.8 × 10−6 g/μm4 s [37]), a numerical model of Pozrikidis’ which simulated the time course of cell deformation (αc = 2.4 × 10−7 ∼ 1.1 × 10−6 g/μm4 s [38], as well as an experimental fit to earlier data (αc = 1.4 × 10−7 g/μm4 s [36])). Note however, that the micromechanical and numerical models of [34, 37, 38] was created for mammalian red blood cells in capillaries and must be applied with caution here; indeed glycocalyx parameters have not been measured in zebrafish. Although the differences between zebrafish and mammalian RBCs mean that we must allow that the parameters controlling occlusive feedback αc may be different in zebrafish than in mammalian vessels, the mammalian data generally support the possibility of tuning feedbacks over a large range of values. The intrinsic resistance αc depends on many factors, including cell velocity, thickness of glycocalyx layer, and the deformation of the cell. Here we focus on the effect of αc on the partitioning of the cells rather than the detailed mechanism that causes the variation. We simulated around 17 min of red blood cell flow through the zebrafish vascular network, assuming the same occlusive effect for every microvessel, using a discrete model in which every red blood cell trajectory was tracked and in which vessel resistances were modeled using Eq (1) (see Materials and methods) using the same occlusive feedback parameter (αc = 1.01 × 10−6 g/μm4 s) for each vessel. The model continued to predict that red blood cell fluxes within vessels decrease exponentially with distance from the heart (Fig 1C). This can be rationalized as follows: If αc is identical between intersegmental vessels, and phase separation is assumed to be negligible, then the model predicts that the resistance of each vessel will increase on average from the value given by the Hagen-Poiseuille law by αc · Hct · V/Vc, where V is the volume of the vessel, Vc is the volume of a single cell and Hct is the hematocrit. The approximate model derived in Results, Absence of occlusion … demonstrates that variation in SeA length from head to tail of the zebrafish contribute very little to partitioning of red blood cell fluxes between SeAs, so changing the resistance of each vessel by an amount simply proportional to its length, will similarly not prevent exponential decay of red blood cell fluxes. The potential effect size of including occlusive feedbacks is much larger than the effect of phase separation: predicted red blood cell flux decreased by a factor of more than 7 in the phase separation model (see S1 Text). We therefore hypothesized that varying occlusive effects between different SeAs may uniformly distribute red blood cells through the network. To probe how variations in occlusive feedback could be used to control the distribution of red blood cells, we studied a reduced model of the vascular network (readers who are mainly interested in simulation results may skip this analysis by going straight to Observed variation in …). Specifically, we built a mean field model for the flows in a model network including only the first and last SeAs, as well as the direct connection between the DA and PCVs (the labeling of vessels and branching points is shown in Fig 5A). In each vessel the cells were assumed to be well-mixed and cell fluxes are divided in proportion to flow rates at all nodes. Then the hematocrit will be the same in all vessels. For simplicity we express our equations in terms of the concentration of red blood cells (number / volume), ρ, rather than the hematocrit. ρ and hematocrit (Hct) are simply related by ρ = Hct/Vc where Vc is the volume of a cell. Let Ri be the modified resistance of the ith vessel according to Eq (1). Then by applying Kirchoff’s first law at the branching points at which first and second Se vessel branch off from the aorta, we obtain the pressures at these points, i.e. p1 and p2: F = p 1 - p 2 R 1 + p 1 R 2 , p 1 - p 2 R 1 = p 2 R 3 + p 2 R 4 , (4) Here F is the total flux of blood into the network, and we can solve Eq (4) by linear algebra (see S1 Text). Of particular interest is is the ratio of fluxes in the two Se, which measures how uniformly the different vessels are kept supplied with cells: Q 4 Q 2 = R 2 0 + V 2 ρ α 2 R 4 0 + V 4 ρ α 4 ( 1 + R 1 R 3 + R 1 R 4 + V 4 ρ α 4 ) − 1 (5) Here α2, α4 are respectively the values of αc in the first and last SeA, R 2 0, R 4 0 are the resistances of the two SeAs in the absence of red blood cell occlusion, and Vi is the volume of vessel i. Most of the parameters in Eq (5) are tightly constrained: the dimensions of the two Se vessels are similar (in fact R 2 0 ≈ 2 R 4 0 and V2 ≈ 2V4), moreover, since the vessel network extends during development and supplies the tail fin in adult zebrafish [59, 60], the aorta must maintain approximately the same radius along its length, leading to R1 ≈ 11R3. Thus the second factor of Eq (5) ( 1 + R 1 R 3 + R 1 R 4 + V 4 ρ α 4 ) − 1 has an upper bound 1 12. Therefore the only parameters that can be used to increase Q4/Q2 (i.e. eliminate short-circuiting of the network by the first SeA) are the relative sizes of α2 and α4. Q4/Q2 is largest if α2 ≫ α4, i.e. if occlusion effects are stronger in the first SeA. Thus uniform flow requires stronger occlusion in vessels close to the heart, consistent with experimental observations in real zebrafish (Fig 4B). However our reduced model also shows that varying occlusion strengths between vessels creates trade-offs between uniformity and the transport efficiency, measured by the dissipation: D network = 8 μ w b π r a 4 ( ℓ 1 Q 1 2 + ℓ 3 Q 3 2 ) + 8 μ p l π r c 4 ( ℓ 2 Q 2 2 + ℓ 4 Q 4 2 ) + ρ ( Q 2 2 V 2 α 2 + Q 4 2 V 4 α 4 ) . (6) (See S1 Text for derivation). Here ℓi is the length of the ith vessel, ra is the radius of the DA, and rc is the radius of the Se vessels. To compare equivalent networks as we vary α2 we also vary F, the total flow into the trunk, to keep the total flux through the pair of Se vessels (Q2 + Q4) constant. Dissipation in the thin layers of fluid surrounding each RBC dominates Dnetwork, so as α2 increases Dnetwork increases. The highest ratios of Q4/Q2 are therefore also the most dissipative networks (Fig 5B). We modified our simulation from Results, Tuning occlusive effects … to incorporate the observed variations in occlusive effects; i.e. using the different measured values of αc in each vessel. We used the regressed data (gray line in Fig 4B) to capture the decreasing trend of αc from head to tail. When vessels were assigned the experimentally measured values of αc, red blood cells became uniformly distributed between SeAs, and matched closely to the real flow observations (see Fig 2A and 2B). Are the measured variations in occlusive effects really evidence of adaptive tuning of the zebrafish cardiovascular network, or could they arise from incidental changes caused for example by the different ages of vessels at different distances along the trunk? New SeAs are progressively added to the trunk at the tail of the zebrafish as the trunk elongates, and we wanted to evaluate the alternate hypothesis that the younger vessels farther from the heart had lower occlusive effects simply because they have a thinner glycocalyx coating, or else because structural adaptation of vessels to the flows through them may tend to reduce vessel radii over time [61]. Although neither alternate explanation can be totally ruled out, we were able to test how close the observed distribution of occlusive effects is to one that optimizes the uniform partitioning of red blood cell flows between vessels. Specifically, we ran discrete cell simulations of flow within the network for different distributions of occlusive effects: that is, we varied Δαc, defined to be the difference in αc between the first and last SeAs, assuming a linear variation of αc in the intermediate vessels. For each model network, we calculated the coefficient of variation (CV) in the red blood cell flux, i.e. the standard deviation in red blood cell flow rate over all vessels, normalized by the mean flow rate. Smaller values of CV correspond to a more uniform distribution of red blood cell flows. Using discrete cell simulations, i.e. tracking every cell trajectory, produces more accurate estimates of red blood cell fluxes in principle than the continuum modeling from Results, Tuning occlusive effects …, because cell number fluctuations within each SeA are comparable to the mean number of cells. Since the change in resistance of a vessel depends on the number of cells in the vessel according to Eq (1), the distribution of red blood cell flows for a given distribution of occlusive effects depends on hematocrit. Accordingly, we varied both hematocrit and occlusive effect distributions independently in our simulations. We found for any fixed hematocrit, near uniform flux (CV close to 0) can be achieved only over a narrow range of Δαc (Fig 6A). Too little difference in intrinsic resistance between first and last SeAs, and the first SeA short-circuits the network, as discussed in Results, Absence of occlusion …. But too large a difference in occlusive effects can have the opposite effect, leading to the vessels furthest from the heart receiving more flow than vessels closest to the heart. The optimal distribution the occlusive effects is realized along a single curve in (Δαc, ρ) space. We found that the observed occlusion effect distribution is close to the optimal value for the real zebrafish hematocrit [43] (Fig 6A). Our work shows that feedbacks associated with the occlusion of fine vessels by the red blood cells that pass through may be associated with previously unreported adaptive benefits for control of blood flows within the microvasculature. Although the existence of occlusive feedbacks is well known [54, 58, 62, 63], to our knowledge they have not previously been shown to be associated with adaptive benefits for oxygen perfusion. Although our experimental observations and modeling are focused on zebrafish, which are a model for vascular development, it is likely that similar feedbacks are significant within mammalian microcirculatory systems, where the deformation of cells to pass through capillaries is, if anything, even more extreme than in the zebrafish. Indeed the apparent intrinsic resistance of cells in human blood vessels has a wide range of variability [34, 37], and precise tuning of blood flows is already known to be vital e.g. to maintain perfusion-ventilation balance in the lungs [64–66]. The proposed occlusion feedback mechanism may be able to explain the variation of capillary blood flow and how it affects the ventilation-perfusion ratio, as well as blood flows in other vascular systems such as brain capillary network. Capillary networks have been hypothesized to be organized to minimize the cost of blood transport [12, 13]. Although large vessels seem to conform very closely to this organizing principle [13, 24], the tuning of occlusive effects to uniformly distribute red blood cell flows takes the zebrafish vascular network far from the configuration that minimizes transport costs. In particular, at the physiological hematocrit, if the same (smallest) occlusive effect, αc, is assigned to each vessel then the dissipation in the network could be reduced by a factor of 11 (Fig 6B). At the same time, more uniform partitioning of cell fluxes between different SeAs (i.e. a lower value of the Coefficient of Variation of red blood cell flow rates) is possible but altering physiological parameters further decreases the transport efficiency. For example decreasing blood cell concentration, ρ, increases uniformity of flux, but at the cost of increasing dissipation if the total cell supply to all Se vessels is to kept fixed (Fig 6B). The ability of SeAs to vary the occlusive effect αc over three orders of magnitude is consistent with previous modeling of red blood cell and microvessel mechanics, and endows the network with tremendous control over red blood cell flow rates. It is natural to ask whether and how uniform red blood cell flux partitioning can be maintained against the numerous sources of perturbation that occur in real cardiovascular networks. Microvascular networks may be disrupted by trauma, micro-anneurysms, or by systemic conditions like diabetes mellitus [67–70]. As a first step toward answering this question, we considered the effect of well-characterized natural variability in SeA spacing [48], and of the notch mutation which alters the trunk network connectivity [71] upon the ability of the network to uniformly distribute red blood cell fluxes. We found that under a wide range of vessel spacing variability, red blood cell fluxes remained uniform across all SeAs (see S1 Text and S3 Fig). Indeed vessel spacing variability has no detectable effect on zebrafish growth and maturation. By contrast, in notch mutant zebrafish the cardiovascular network is malformed, with a shunt connection forming between aorta and principal cardinal vein (S4 Fig). Since the diameter of the shunt is much larger than the cell diameter, there is negligible occlusive feedback within the shunt, causing it to irreparably short-circuit the vascular network. Shunt formation is lethal in embryos, and our model shows that it creates conditions under which uniform perfusion of the trunk is impossible. Note however that mechanisms not described in the model can still play significant roles in both developmental process and mutant network phenotypes. For example in the gridlock mutant [51] the blood flow to the tail is impeded by a localized vascular defect, but the collateral vessels not present at 4dpf was observed to redirect the flow around the blockade and rescue the embryo. During the development process both the number of vessels and size of zebrafish embryo change dramatically. Therefore we expect an observable change in occlusive feedbacks to maintain uniform cell partition throughout the developmental stages. Extending our analysis to include the topological changes observed as embryonic zebrafish develop [51] is an ongoing effort. Although we are able to directly demonstrate that occlusive feedbacks vary between different the SeAs, and this variation is consistent with optimization of feedback strengths to ensure uniform distribution of red blood cells across trunk vessels, our model cannot reveal what physical changes within vessels are used with the zebrafish network to modulate the occlusive effect. In our experiments we cannot visualize the glycocalyx lining of the SeAs, and in fact we are aware of no previous works in which glycocalyx was measured in blood vessels simultaneously with flow. However, previous studies have reported large variations in glycocalyx porosity and thickness between different vessels [32, 72]. Since cells must squeeze into SeAs, variations in vessel radius below the resolution limit of our microscopy method could also account for the variation in occlusive effect. Finally elastohydrodynamic effects associated e.g. with changes in the speed of cells, [32], may affect feedback models. The analysis is also silent on the mechanisms for coordinating occlusive effects across the network. Recent works have dissected structural adaptations in microvascular networks [61], as well as in biological transport networks generally [73–75]. These works have focused on the question of how a set of vascular elements that have information only about their own flows can alter their resistances in response to these cues to minimize dissipation within the network. This question is directly relevant to other objective functions i.e. to networks that maximize uniformity rather than maximizing hydraulic efficiency—can vessels adapt their occlusive effects to the their flow to achieve uniform red blood cell transport? The use of tuned occlusive effects creates uniform distribution of red blood cell fluxes through the zebrafish vascular network, but at the cost of increasing transport costs. Indeed if the network simply used the same value of αc in every SeA we found that an 11 fold decrease in transport costs would be possible within the zebrafish trunk vasculature (Fig 6B). Physically feedbacks from occlusion represent a form of congestion, and efficient transport networks, both natural [76] and artificial [10, 11], are often organized to avoid congestion. Previous works have provided algorithms for constructing minimally dissipative networks given a prescribed set of sources and sinks [22, 23]. Our work suggests that other optimizing principles may govern microvascular network organization. Extending network optimization algorithms to include flow uniformity is likely to further reveal the tradeoffs between uniformity and efficiency.

10.1371/journal.pbio.2004303

Communication across the bacterial cell envelope depends on the size of the periplasm

The cell envelope of gram-negative bacteria, a structure comprising an outer (OM) and an inner (IM) membrane, is essential for life. The OM and the IM are separated by the periplasm, a compartment that contains the peptidoglycan. The OM is tethered to the peptidoglycan via the lipoprotein, Lpp. However, the importance of the envelope’s multilayered architecture remains unknown. Here, when we removed physical coupling between the OM and the peptidoglycan, cells lost the ability to sense defects in envelope integrity. Further experiments revealed that the critical parameter for the transmission of stress signals from the envelope to the cytoplasm, where cellular behaviour is controlled, is the IM-to-OM distance. Augmenting this distance by increasing the length of the lipoprotein Lpp destroyed signalling, whereas simultaneously increasing the length of the stress-sensing lipoprotein RcsF restored signalling. Our results demonstrate the physiological importance of the size of the periplasm. They also reveal that strict control over the IM-to-OM distance is required for effective envelope surveillance and protection, suggesting that cellular architecture and the structure of transenvelope protein complexes have been evolutionarily co-optimised for correct function. Similar strategies are likely at play in cellular compartments surrounded by 2 concentric membranes, such as chloroplasts and mitochondria.

Antibiotic resistance is one of the greatest threats that humanity faces today. In particular, the emergence of multidrug-resistant gram-negative bacteria has become a pressing issue. Infections caused by resistant gram-negative bacteria are indeed difficult to treat because of the presence of a double-membraned envelope that renders these bacteria less permeable to many antibiotics. Bacteria invest a great deal in protecting their envelope. They have evolved complex stress-sensing systems that allow them to monitor envelope integrity in order to detect and fix damage when it occurs. Here, using the model bacterium Escherichia coli, we show that the ability to sense and respond to stress depends on the architecture of the cell envelope, in general, and on the distance between the 2 membranes, in particular. Thus, searching for drugs targeting envelope architecture is an attractive strategy for new antibacterials. In addition, our work reveals the physiological function of the Braun’s lipoprotein (Lpp), the numerically most abundant protein in E. coli and the first lipoprotein that was discovered several decades ago.

Although the multilayered architecture of the cell envelope of gram-negative bacteria was first described in the 1960s, we are still unraveling the links between the structure of this cellular component and its functions in the cell. The envelope of these bacteria consists of an inner membrane (IM), a classical phospholipid bilayer around the cytoplasm, and an outer membrane (OM), an asymmetric structure with phospholipids in the inner leaflet and lipopolysaccharides in the outer leaflet [1]. The space between the IM and the OM defines the periplasm, a cellular compartment that contains the peptidoglycan, a polymer of glycan strands cross-linked by short peptides that provides shape and osmotic protection to cells. This multilayered envelope, which contributes to cellular integrity and modulates permeability, is required for life and serves as an interface to the external milieu. Several essential protein machineries are present in the cell envelope, where they engage in processes that are important for envelope assembly and protection [2–5]. Many of these machineries span the periplasm, with components in both the IM and OM. The pathways that assemble the envelope and that monitor envelope integrity are tightly coordinated in order to sense and respond to damage. However, while the mechanisms of these pathways and how they cooperate to ensure envelope homeostasis have been the focus of numerous studies, the extent to which they depend on the architecture of this compartment for correct functioning is unknown. To close this gap, here we investigated a key player in the organization of the cell envelope of Enterobacteriaceae: the small, alpha-helical protein Lpp, also known as Braun’s lipoprotein [6]. Lpp is numerically the most abundant protein in Escherichia coli, with >1 million copies per cell [7]. Such an extreme abundance suggests that it plays a major role in the cell. Importantly, Lpp provides the only covalent connection between the OM and the peptidoglycan: it is anchored to the OM via a lipid moiety at its N-terminus and attached to the peptidoglycan via its C-terminus. The attachment of Lpp to the peptidoglycan is catalyzed by a family of L,D-transpeptidases [8] that link the C-terminal lysine of Lpp to a diaminopimelic acid residue in the peptide stems of the peptidoglycan [9]. Three E. coli periplasmic enzymes (YbiS, ErfK, YcfS) exhibit this activity [8]. Both this functional redundancy and the high abundance of Lpp suggest that it must be critically important to covalently connect the OM to the peptidoglycan. However, lpp deletion mutants grow and divide normally in culture [10]. The physiological importance of Lpp and of covalently anchoring the OM to the peptidoglycan therefore remains enigmatic. To control the critical biogenesis of the envelope and maintain its integrity, bacteria have evolved intricate signal transduction systems that induce stress responses to deal with potential dysfunctions and environmental insults to the envelope. In Enterobacteriaceae, the Regulation of Capsule Synthesis (Rcs) system monitors the integrity of the OM and the peptidoglycan [11–13] (S1 Fig). Drugs that interfere with peptidoglycan assembly or alter the lipopolysaccharide leaflet of the OM induce Rcs, which represses cellular motility and produces an extracellular capsule that functions like a protective shield [14]. Rcs is a complex stress-signaling cascade involving at least 6 components [12] (S1 Fig). Most Rcs-inducing cues are sensed by RcsF, an OM-localized lipoprotein, which detects damage caused by chemicals targeting the OM or the peptidoglycan [15, 16] or by mutations in genes involved in envelope assembly [12, 17]. Under stress, RcsF interacts with IgaA, an IM protein, thus constituting a molecular signal that turns on the Rcs pathway via a phosphorylation cascade (S1 Fig) [11, 12]. Rcs therefore mediates signal transduction from the outmost cell layer (the peptidoglycan and the OM) to the control center of the cell (the cytoplasm). To what extent the functioning of a system such as Rcs, with components on both sides of the periplasm, depends on the architecture of the cell envelope remains unknown. Here, we established that the transmission of stress signals from the OM to the IM depends on the size of the periplasm. By manipulating the length of the lipoprotein Lpp, we demonstrated that cells in which the IM-to-OM distance is artificially increased become blind to stress affecting the peptidoglycan and the OM. Remarkably, we restored the line of communication between the 2 membranes in cells with a larger envelope by increasing the length of the stress sensor lipoprotein RcsF. We also established that the OM must be covalently attached to the peptidoglycan for intermembrane communication to occur. In order to interrogate the importance of attaching the OM to the peptidoglycan for the communication of information about cellular stress to the cytoplasm, we triggered the Rcs system in an RcsF-dependent manner [11, 16] with the compound A22, which inhibits the actin-like protein MreB15, and mecillinam, a ß-lactam antibiotic, which inhibits the essential transpeptidase PBP2 [18, 19]. Both drugs cause cells to round and eventually lyse [20, 21]. Under these conditions, relative to wild-type (WT) cells, the Rcs pathway was impaired in a mutant lacking YbiS, the primary L,D-transpeptidase [8], and even more in a mutant lacking all 3 L,D-transpeptidases [8] (ΔybiSΔerfKΔycfS cells, denoted Δldt3 here) (Fig 1A). Note that Δldt3 cells (S2A and S2B Fig) remained susceptible to A22 and mecillinam, becoming round, as expected. The absence of YbiS, ErfK, and YcfS also impaired Rcs signalling in cells lacking mdoG (Fig 1B), a gene involved in the synthesis of periplasmic osmoprotectant sugars. Deletion of mdoG has been reported to constitutively activate Rcs via RcsF [22]. Expression of the L,D-transpeptidase YbiS, but not of a catalytically inactive variant, from a plasmid restored Rcs signalling both in the Δldt3ΔmdoG mutant and in the Δldt3 mutant in response to A22/mecillinam (Fig 1B and 1C). Thus, the inability to induce Rcs directly results from the lack of L,D-transpeptidase activity in these cells. Cells expressing an Lpp variant lacking the residue (K58 in E. coli) required for attachment of the protein to the peptidoglycan (LppΔK58) were also unable to trigger Rcs in response to mecillinam, A22, or mdoG deletion (S3A and S3B Fig). Taken together, these data indicate that cells become blind to insults that normally activate the Rcs system via RcsF when Lpp is detached from the peptidoglycan. We then probed whether the failure to induce Rcs resulted specifically from the inability of RcsF to sense stress in cells with no peptidoglycan-linked Lpp. RcsF forms a complex with BamA, the key component of the β-barrel assembly machinery [23], and with the β-barrel proteins OmpA, OmpC, and OmpF [11, 24]. When in complex with these proteins, RcsF is sequestered from IgaA, its downstream partner, and cannot activate Rcs [11]. We previously reported that Rcs-inducing compounds such as A22 and mecillinam decrease the levels of the BamA-RcsF complex [11], likely because newly synthesised RcsF molecules cannot bind BamA under stress. In our model, failure to bind BamA prevents the sequestration of RcsF by β-barrel proteins, allowing RcsF to interact with IgaA and to turn on Rcs [11]. Here, we determined that A22 and mecillinam decreased the levels of BamA-RcsF in Δldt3 cells, similar to the decrease in WT cells (S4A and S4B Fig). Thus, the absence of Rcs induction in cells with no peptidoglycan-bound Lpp (Fig 1, S3A and S3B Fig) did not result from an impaired ability of RcsF to sense stress. Recent studies in Salmonella concluded that Lpp was a principal determinant of the size of the periplasm [25], prompting us to consider the possibility that in cells with no peptidoglycan-bound Lpp, the structure of the cell envelope is affected in such a way that the OM lipoprotein RcsF cannot reach across the periplasm to contact IgaA in the IM and activate the Rcs system (S1 Fig). Here, cryo-electron microscopy (cryo-EM), a technique in which cells are preserved in a frozen-hydrated, near-native state, revealed that cells expressing LppΔK58 formed OM blebs and had envelope defects (Fig 2A). Similar morphological deformities were observed with the lpp deletion mutant [25, 26], indicating that, in the case of this mutant, deformities specifically resulted from the absence of covalent anchoring of the OM to the peptidoglycan. Interestingly, the average IM-to-OM distance increased by about 3 nm compared to WT cells (Fig 2B), which suggested that the absence of Rcs induction in cells with no peptidoglycan-bound Lpp could result from a larger periplasm. However, the IM-to-OM distance varied considerably along the cell axis, which complicated further interpretation of the results obtained in cells expressing LppΔK58. Increasing the length of Lpp was previously reported to change the IM-to-OM distance in Salmonella without causing OM blebbing [25]. To clearly establish the importance of IM-to-OM distance in Rcs signalling, we engineered E. coli strains expressing longer Lpp variants (insertions of 14 residues (Lpp+14) or 21 residues (Lpp+21)) from the native lpp locus (S5A Fig); with these strains, we sought to increase the IM-to-OM distance without altering Lpp cross-linking to the peptidoglycan (S5B Fig). Strikingly, cryo-EM revealed that the IM-to-OM distance in these strains increased proportionally to the length of Lpp, with increases of 3 nm and 4 nm in cells expressing Lpp+14 and Lpp+21, respectively (Fig 2A and 2B). We next used the engineered strains to test the impact of increasing the IM-to-OM distance on Rcs signalling. Remarkably, adding 14 or 21 residues to Lpp completely abolished Rcs activation: cells expressing Lpp+14 or Lpp+21 did not turn on Rcs when exposed to mecillinam or A22 or deleted for mdoG (Fig 3A and 3B). Stress was still sensed by RcsF in these strains, as BamA-RcsF levels decreased under stress (S6A and S6B Fig). In addition, retargeting RcsF to the IM by altering its sorting signal constitutively induced Rcs (S7 Fig), as expected [15], indicating that Rcs could still be turned on in strains expressing Lpp+14 or Lpp+21. Interestingly, increasing the length of Lpp was well tolerated: cells expressing Lpp+14 or Lpp+21 exhibited no growth defects (S5C Fig), did not bleb (Fig 2A), and had WT morphology except for a larger periplasm (Fig 2A). In addition, expression of Lpp+14 or Lpp+21 fully complemented the sensitivity of the Δlpp mutant to the membrane perturbant dibucaine [27, 28] (S5D Fig) and complemented the growth defect of the ΔmrcBΔlpp double mutant [29] (S5E Fig). The results above are consistent with the hypothesis that RcsF does not activate the Rcs system when the size of the periplasm increases because RcsF cannot reach IgaA. If this hypothesis is correct, then making RcsF longer should allow this protein to span the increased IM-to-OM distance in order to reach IgaA and activate Rcs. RcsF has a 32-residue unstructured linker that is located upstream of the signalling domain [30, 31] and that most likely allows the protein to reach the IM. To make RcsF longer, we added 7 residues to the C-terminal part of the linker (RcsF+7; S8A and S8B Fig). Because this sequence is disordered (S8C and S8D Fig), we predicted that adding it to the linker would increase the length of RcsF by 2–3 nm (S8C Fig), roughly corresponding to the increase in length resulting from the addition of a 14-residue α-helix to Lpp (strain lpp+14). Remarkably, expression of RcsF+7 in cells producing Lpp+14 fully restored Rcs signalling in response to mdoG deletion, A22, or mecillinam (Fig 3C and 3D). It also partially rescued signalling in cells expressing Lpp+21 (Fig 3C and 3D), in which the periplasmic size is larger than WT by 4 nm (Fig 2B). Thus, increasing the length of RcsF restored normal Rcs signalling in cells in which the IM-to-OM distance was increased to a similar extent. These observations support our model that, in cells with a larger IM-to-OM distance, the absence of Rcs induction under stress results from the inability of RcsF to reach the other side of the periplasm (S9 Fig). Taken together, our results reveal the importance of controlling the IM-to-OM distance in the maintenance of intermembrane communication within the cell envelope, establishing the physiological importance of the size of the periplasm and highlighting the exquisite architectural organisation of the envelope. In particular, we demonstrated that strict control over the IM-to-OM distance is required for effective envelope surveillance and protection during exponential growth. Even a slight increase in the size of the periplasm disrupted the line of communication between the OM and the IM, effectively disconnecting the outer part of the envelope from the cytoplasm, where cellular behaviour is controlled (S9 Fig). Thus, envelope components have likely been subjected to evolutionary pressure to properly function in the strictly controlled dimensions of the periplasm. For example, the sizes of RcsF and of the periplasmic domain of IgaA must have been rigorously selected to allow the optimal transmission of stress signals between these 2 proteins across the envelope. In addition, our observation that cells expressing longer Lpp variants had no phenotype in the conditions tested (Fig 2A and S5B–S5E Fig) reveals that the essential envelope-spanning machineries involved in elongation and division have enough intrinsic flexibility to adapt to changes in envelope architecture, suggesting that evolution has selected for robustness in the case of these systems. These results motivate future research to test whether the correct functioning of other envelope-spanning systems in E. coli and other bacteria also depends on strict control over IM-to-OM distance and to determine the evolutionary advantage that makes the Rcs system so susceptible to changes in the size of the periplasm. Recent work revealed that Lpp functions as an OM tether under normal growth conditions in Salmonella [25]. Accordingly, removing the covalent connection between OM-anchored Lpp and the peptidoglycan leads to the formation of OM blebs in E. coli (Fig 2A). Interestingly, although the average IM-to-OM distance was larger in cells expressing LppΔK58 than in cells expressing WT Lpp, WT size was maintained in substantial portions of the envelope (Fig 2B). A likely explanation is that in the absence of peptidoglycan-bound Lpp, envelope proteins such as the lipoprotein Pal and the β-barrel OmpA, which noncovalently bind peptidoglycan [32, 33], partially compensate for the loss of Lpp periplasmic spanners, sustaining WT size in some envelope areas. Strikingly, in cells expressing Lpp+14 or Lpp+21, the IM-to-OM distance increased homogeneously along the cell axis, and no WT size was observed (Fig 2A and 2B), suggesting that these longer versions of Lpp override proteins like Pal and OmpA and impose a larger periplasm. Thus, Lpp may act as both a tether and a support column for the OM. Lpp was the first lipoprotein to be identified [6]. However, the functional importance of this massively abundant protein remained poorly defined. Fifty years later, the crucial role of Lpp in controlling the architecture of the bacterial cell envelope, and in particular the IM-to-OM distance, is finally coming to light. The current investigation has revealed the importance of controlling this distance in the maintenance of intermembrane communication within the cell envelope. Cellular compartments surrounded by 2 concentric membranes occur in all kingdoms of life. By highlighting the importance of controlling the intermembrane distance in gram-negative bacteria, our work suggests that similar strategies are at play in other cellular compartments surrounded by 2 concentric membranes, such as chloroplasts and mitochondria. Bacterial strains and plasmids used in this study are listed in S1 and S2 Tables, respectively. E. coli K12 strain MG1655 was used as WT throughout the study. For all deletion mutants, the corresponding alleles from the KEIO collection [34] were transferred into MG1655 or derivatives carrying a chromosomal PrprA-lacZ fusion at the phage lambda attachment site [35] via P1 phage transduction and validated with PCR. To excise the kanamycin-resistance cassette (kanR), pCP20 was used as described previously [36]. lpp length mutants were constructed in E. coli K12 F- λ- cells harbouring the λRed recombineering plasmid (pKD46) [36], as described previously [25]. Briefly, a tetracycline-resistance cassette (tetRA) was inserted between codons 42 and 43 of E. coli lpp. Replacement of the tetRA cassette with insertions of 14 or 21 residues (2 and 3 heptad repeats, respectively) was accomplished by introducing dsDNA fragments bearing 40 bp of homology to E. coli lpp directly flanking the 5′ and 3′ ends of the tetRA cassette. Tetracycline-sensitive transformants (lpp length mutants) were selected by plating on medium containing anhydrotetracycline (0.2 mg/ml) and fusaric acid (2.4 mg/ml). Chromosomal lpp length mutations were moved into the MG1655 background through standard λRed recombineering as described in [37]. Primer sequences are available upon request. The lppΔK58 mutant was constructed using site-directed mutagenesis of WT lpp cloned into pBad18, as previously described [38]. lppΔK58 was moved onto the chromosome of MG1655 cells via standard λRed recombineering, as described in [37]. To create a longer version of RcsF, we used pSC202 (the RcsF ORF in the low-copy vector pAM238) [36] as a template (S2 Table). The extended linker fragment was synthesised with flanking restriction sites RsrII and PstI using gene fragments (sequence available upon request). The fragment and vector pSC202 were digested and ligated to yield pRcsF+7. β-galactosidase activity was measured as described previously [39]. Briefly, cells harbouring PrprA-lacZ at the attB phage lambda site on the chromosome were diluted 1:100 from overnight cultures in Luria broth (LB), then incubated at 37°C. When needed, the cultures were treated with 0.3 μg/ml mecillinam or 5 μg/ml A22 at OD600 = 0.2. Cells were further grown for 1 h (with mecillinam treatment) or 40 min (with A22 treatment). Otherwise, cells were harvested at OD600 = 0.6. Twenty microlitres of cells were harvested and incubated with 80 μl of permeabilisation solution (60 mM Na2HPO4·2H2O, 40 mM NaH2PO4·H2O, 10 mM KCl, 1 mM MgSO4·7H2O, 50 mM β-mercaptoethanol) for 30–45 min at room temperature. Then, 600 μl of substrate (1 mg/ml O-nitrophenyl-β-d-galactoside, 50 mM β-mercaptoethanol) were added. The mixture was further incubated at 30°C for 20–25 min. Seven hundred microlitres of 1 M Na2CO3 were added to stop the reaction, and the optical density was measured at 420 nm. The standardized amount of β-galactosidase activity was reported in Miller units. The ratio of PrprA-lacZ induction was calculated relative to the basal level in a WT strain. Bar graphs with corresponding statistical analysis were prepared using Prism 7 (GraphPad Software, Inc.). Cells were grown in LB until OD600 = 0.5. One millilitre of cells was harvested via centrifugation at 3,300 × g for 5 min at 4°C, washed, and resuspended in phosphate-buffered saline (PBS). The cell suspension was split in half and DTSSP (Thermo Scientific), prepared fresh in PBS, was added to 1 of the 2 samples to a final concentration of 2 mM. The DTSSP-treated and nontreated samples were incubated at 30°C for 1 h under orbital agitation. Glycine (0.1 M) was added to both samples to quench cross-linking. After 10 min on ice, the cell suspensions were precipitated with 10% trichloroacetic acid (TCA), washed with ice-cold acetone, and resuspended in Laemmli SDS sample buffer (2% SDS, 10% glycerol, 60 mM Tris-HCl [pH 7.4], 0.01% bromophenol blue). The volume of sample buffer used to resuspend the samples was normalised to the number of cells loaded. The samples were subjected to SDS-PAGE and immunoblotting using specific antibodies, as described below. Only data from DTSSP-treated samples are shown in S4 and S6 Figs. Protein samples were separated in 12% or 4%–12% acrylamide gels (Life Technologies) and transferred to nitrocellulose or polyvinylidene fluoride membranes (Whatman, 0.45 μm). Immunoblotting was performed as described previously [11]. Rabbit anti-RcsF (from the Collet laboratory’s collection) and rabbit anti-Lpp (from the Hughes laboratory’s collection) were diluted 1:10,000 in 1% skim milk, TBS-T (50 mM Tris-HCl [pH 7.6], 0.15 M NaCl, 0.1% Tween 20). The membranes were incubated with horseradish peroxidase-conjugated goat anti-rabbit IgG (Sigma; 1:10,000 in 1% skim milk, TBS-T) and washed with dilution buffer. Labelled proteins were detected via enhanced chemiluminescence (Pierce ECL Western Blotting Substrate, Thermo Scientific), which was imaged with a GE ImageQuant LAS4000 camera (GE Healthcare Life Sciences). To measure RcsF levels, band intensities were quantified using ImageJ 1.48v (NIH) and analysed with ImageQuant TL software 1Dv8.1 to ensure that they were within the linear range. Strains were grown to an OD600 of 0.5. Then, 450 μl of culture was collected and mixed with 50 μl of 10× fixing solution (0.4% glutaraldehyde, 25% formaldehyde, 330 mM sodium phosphate [pH 7.6]) for 30 min at room temperature. Cells were spun down at 6,000 × g at room temperature, resuspended in 50 μl PBS, and imaged on agarose pads. Fixed samples were imaged with a Primo Star microscope (Carl Zeiss) equipped with an AxioCam 105 color camera (Carl Zeiss) and a phase-contrast objective (Plan-Achromat 40×/0.65 Ph2; Carl Zeiss). Images were acquired with Zen 2 (blue edition; Carl Zeiss). Exposure times and image scaling were identical for compared conditions. MicrobeTracker [40] was used to obtain cell outlines. Quantitative analysis from cell meshes was done with MATLAB R2014a (Mathworks, Inc.) using custom scripts to plot the distributions of length-to-width ratios for the strains tested. Strains were grown aerobically in LB at 37°C until an OD600 of 0.6 was reached. Cells were spun for 5 min at 6,000 × g at 4°C and resuspended to an OD600 of about 12. UltraAuFoil R2/2 grids (200 mesh; Quantifoil Micro Tools GmbH) were glow-discharged for 60 s at 10 mA. Cells were mixed with a solution of 10 nm colloidal gold (Sigma) immediately before freezing. A 2.5-μl droplet of sample was applied to the grid and plunge frozen using a Vitrobot MkIV (FEI Company) with a wait time of 60 s, a blot time of 5 s, a blot force of 3, and a drain time of 1 s at a constant humidity of 100%. Grids were stored under liquid nitrogen until required for data collection. Projection images were collected on a 200 keV FEI Tecnai TF20 FEG transmission electron microscope (FEI Company) equipped with a Falcon II direct electron detector (FEI Company) using a Gatan 626 cryogenic-holder (Gatan). Leginon automated data-collection software 3.0 [41] was used to acquire images with pixel size of 0.828 nm (nominal magnification 25,000×) with a defocus of −5 μm. Membrane measurements were carried out as previously described [25]. Briefly, 3dmod from the IMOD package [42] and custom scripts were used to manually segment the IMs and OMs of projection images of about 35 cells per mutant, measuring the periplasmic width at 0.5-nm intervals to produce width histograms (Fig 2B). Single colonies were used to inoculate overnight cultures, which were diluted 1:1,000 in round-bottom 96-well plates in 200 μl LB. Absorbance was measured at 600 nm every 30 min in a Synergy H1 microplate reader (BioTek) with constant orbital shaking at 37°C. Graphs were prepared using GraphPad Prism 7. Sensitivity to the membrane perturbant dibucaine was assessed on LB agar plates. Briefly, 4-ml cultures were inoculated with overnight cultures at 1:100 dilution and grown in LB at 37°C until an OD600 of about 0.5 was reached. Cell counts were normalised according to OD600, then serially diluted in LB with seven 10-fold dilutions using 96-well microtitre plates (Corning). Two microlitres of the diluted cultures were manually spotted onto the LB agar plates and incubated overnight at 37°C. When indicated, dibucaine (Sigma-Aldrich) was added to the LB agar plate at a final concentration of 1.2 mM. Cultures were grown until OD600 = 0.7 in LB at 37°C with agitation. Cells (1 ml) were pelleted and resuspended in 1 ml PBS. When indicated, 1 mg lysozyme (Sigma-Aldrich) and EDTA (final concentration 10 mM) were added to the resuspended cells, which were incubated for 1 h at 37°C. Lysates were obtained via TCA precipitation and resuspension in 100 μl 2× SDS Laemmli buffer. Typically, 5–10 μl of sample were loaded into 12% Bis-Tris polyacrylamide gels (Nupage). Lpp was detected using anti-Lpp antiserum. The significance of differences among bacterial strains was assessed using GraphPad Prism 7 according to analysis of variance (ANOVA), followed by the application of Tukey’s multiple-comparison test when the distribution was normal. Otherwise, the Kruskal-Wallis test was used, followed by Dunn’s multiple comparison test. Normality was assessed using the Shapiro-Wilk test.

10.1371/journal.pntd.0004768

Accuracy of Mobile Phone and Handheld Light Microscopy for the Diagnosis of Schistosomiasis and Intestinal Protozoa Infections in Côte d’Ivoire

Handheld light microscopy using compact optics and mobile phones may improve the quality of health care in resource-constrained settings by enabling access to prompt and accurate diagnosis. Laboratory technicians were trained to operate two handheld diagnostic devices (Newton Nm1 microscope and a clip-on version of the mobile phone-based CellScope). The accuracy of these devices was compared to conventional light microscopy for the diagnosis of Schistosoma haematobium, S. mansoni, and intestinal protozoa infection in a community-based survey in rural Côte d’Ivoire. One slide of 10 ml filtered urine and a single Kato-Katz thick smear from 226 individuals were subjected to the Newton Nm1 microscope and CellScope for detection of Schistosoma eggs and compared to conventional microscopy. Additionally, 121 sodium acetate-acetic acid-formalin (SAF)-fixed stool samples were examined by the Newton Nm1 microscope and compared to conventional microscopy for the diagnosis of intestinal protozoa. The prevalence of S. haematobium, S. mansoni, Giardia intestinalis, and Entamoeba histolytica/E. dispar, as determined by conventional microscopy, was 39.8%, 5.3%, 20.7%, and 4.9%, respectively. The Newton Nm1 microscope had diagnostic sensitivities for S. mansoni and S. haematobium infection of 91.7% (95% confidence interval (CI) 59.8–99.6%) and 81.1% (95% CI 71.2–88.3%), respectively, and specificities of 99.5% (95% CI 97.0–100%) and 97.1% (95% CI 92.2–99.1%), respectively. The CellScope demonstrated sensitivities for S. mansoni and S. haematobium of 50.0% (95% CI 25.4–74.6%) and 35.6% (95% CI 25.9–46.4%), respectively, and specificities of 99.5% (95% CI 97.0–100%) and 100% (95% CI 86.7–100%), respectively. For G. intestinalis and E. histolytica/E. dispar, the Newton Nm1 microscope had sensitivity of 84.0% (95% CI 63.1–94.7%) and 83.3% (95% CI 36.5–99.1%), respectively, and 100% specificity. Handheld diagnostic devices can be employed in community-based surveys in resource-constrained settings after minimal training of laboratory technicians to diagnose intestinal parasites.

Handheld light microscopes are new technologies that may be helpful in enabling better access to diagnostic testing for people living in resource-constrained settings in tropical and subtropical countries. Recent studies evaluating the accuracy of such devices have focused on their use by expert microscopists and were mainly conducted in laboratories. We evaluated the operating performance of two handheld microscopes (Newton Nm1 microscope and clip-on version of the reversed-lens CellScope) in comparison to conventional microscopy for the diagnosis of urogenital and intestinal schistosomiasis, when integrated into routine use in a community-based survey carried out in Côte d’Ivoire. Additionally, we evaluated the same microscopist’s diagnostic performance with the Newton Nm1 microscope for intestinal protozoa in a laboratory set-up. The Newton Nm1 microscope demonstrated excellent diagnostic sensitivity and specificity for schistosomiasis and intestinal protozoa. The CellScope had high specificity but only modest sensitivity for schistosomiasis diagnosis. Taken together, handheld diagnostic tools show promise to improve the quality of clinical and public health care delivered in resource-constrained settings.

Neglected tropical diseases have considerable detrimental impacts in resource-constrained settings as they can result in chronic disability and stigmatization, and have profound negative economic consequences [1,2]. Microscopy is an essential tool in the diagnosis and surveillance of many neglected tropical diseases and is a vital component in virtually every clinical and public health laboratory worldwide. Unfortunately even basic microscopy facilities are lacking in many resource-constrained settings where the greatest needs exist, and where neglected tropical diseases are rife [3]. Schistosomiasis is a neglected tropical disease and an important public health threat, with countries in sub-Saharan Africa affected most [4]. Chronic infection with Schistosoma mansoni may result in disability and death due to complications of portal hypertension, while chronic infection with S. haematobium frequently results in genitourinary morbidity and mortality, with bladder cancer as a well-known complication [5,6]. Intestinal protozoa, such as Entamoeba histolytica and Giardia intestinalis, are common pathogens accounting for widespread morbidity and mortality in resource-constrained settings. For example, E. histolytica is responsible for an estimated 40,000 to 100,000 deaths annually. G. intestinalis, a common cause of diarrheal illness, has an estimated prevalence of 20–30% in low-income countries [7,8]. These infections are diagnosed primarily by stool microscopy (or urine microscopy in the case of S. haematobium). Recently, portable handheld microscopes [9–11] and mobile phone-based microscopes [12–14] have been evaluated for the diagnosis of neglected tropical diseases (e.g., schistosomiasis, opisthorchiasis, and soil-transmitted helminthiasis) and malaria. Most of the prior studies evaluating handheld and mobile phone-based microscopy have utilized expert microscopists in field settings, or were implemented under laboratory conditions. Prior to wide-scale utilization, these devices must be validated in real-world clinical and public health settings, and operated by individuals who will use them in routine daily practice. Here, we integrate a handheld light microscope (i.e., Newton Nm1 microscope; Newton Microscopes; Cambridge, United Kingdom) [15] and a handheld mobile phone-based microscope (i.e., clip-on version of the reversed-lens CellScope) [16] for the diagnosis of S. mansoni, S. haematobium, and intestinal protozoa into a community-based survey in rural Côte d’Ivoire. These devices were chosen because of their compact design, ease of use, and sufficient resolution to detect intestinal and urogenital parasites. We assessed the accuracy of these devices by comparing them to routine microscopy. This study was embedded into a larger, cross-sectional, community-based survey in Côte d’Ivoire. Ethical approval was granted by the Ministry of Health and Public Hygiene of Côte d’Ivoire (reference no., 32/MSLS/CNER-dkn). Written informed consent was obtained from adults aged 18 years or older, and parents or legal guardians on behalf of children. Children, in addition, assented orally. Anthelmintic treatment was offered to all participants at the end of the study (i.e., praziquantel, 40 mg/kg of body weight for schistosomiasis and albendazole, 400 mg for soil-transmitted helminthiasis). This cross-sectional study was conducted in the village of Grand Moutcho in southern Côte d’Ivoire (geographic coordinates: 4.181 N latitude and 5.961 E longitude). The village belongs to a region that is highly endemic for schistosomiasis [17]. The study was carried out between April and June 2014. Study participants were between 6 and 19 years of age. Early morning stool and urine samples, collected between 10:00 and 12:00 hours [18], were processed and evaluated on the spot in a community clinic. Fresh stool samples were processed with the Kato-Katz technique [19]. In brief, standard 41.7 mg thick smears were placed on microscope slides for evaluation of S. mansoni and soil-transmitted helminth eggs. In addition, approximately 2 g of unprocessed stool from each individual was fixed in a standard solution of sodium acetate-acetic acid-formalin (SAF) for subsequent laboratory processing and diagnosis of intestinal protozoa infections. Urine samples were first shaken, then 10 ml was extracted and pressed through a 13 mm diameter meshed filter with 20 μm pores (Sefar AG; Heiden, Switzerland). One drop of Lugol’s iodine solution was placed over the filter prior to examination. We selected one Kato-Katz thick smear slide and one filtered urine slide from each individual on their first day of participation in the study. Each slide was subjected to three microscope techniques shortly after collection, and eggs were identified and quantified at the field site. All microscopists were blinded to prior diagnoses on each slide. Slides were first evaluated by ‘gold’ standard microscopy, with an Olympus CX21 microscope under 10x and 40x lenses (Olympus; Volketswil, Switzerland). Laboratory technicians read slides, and 10% of all slides were re-examined by a senior expert microscopist (JTC, IIB) blinded to prior results for quality control and validation. Each slide was subsequently examined by two experimental microscopes; the Newton Nm1 Portable Field Microscope and the mobile phone-mounted reversed-lens CellScope (Fig 1). The Newton Nm1 microscope is a handheld, commercially available device, weighing 480 g with modular objective lenses (10x, 40x, and 100x), and has been described in field use elsewhere [11,13]. The reversed-lens CellScope fits a 3D printed plastic attachment weighing 5.2 g over an iPhone 5s (Apple; Cupertino, California, United States of America), with an embedded lens superimposed over the iPhone lens. This device harnesses the mobile phone’s light source to illuminate a specimen [14,16]. Laboratory technicians were provided with a half-day of training with direction on the operation of each microscope prior to initiating the study. This training consisted of didactic teaching sessions followed by supervised, hands-on training with multiple test slides. Briefly, the Newton Nm1 is operated by placing a slide in an XY translation stage mounted above the objective, focusing the objective on the sample, and then scanning the sample as it is viewed through the eyepiece of the microscope. The reversed-lens CellScope is operated by holding the mobile phone microscope above the sample and manually moving the device above the sample at the same time as maintaining focus and viewing the images on the screen. Since the slide evaluation was performed live using the screen, which displays images with a lower resolution than still photographs that capture the full resolution of the microscope, the effective resolution of the device used in this study was 14 μm. One month after completion of the field study, SAF-fixed stool samples were subjected to an ether-concentration method, performed in the laboratory of the Centre Suisse de Recherches Scientifiques en Côte d’Ivoire near Abidjan, using a standard protocol [20]. In brief, the SAF-fixed stool samples were re-suspended and placed into a centrifuge tube and centrifuged for 1 min at 500 g. The supernatant was discarded and 7 ml of 0.85% NaCl plus 3 ml of ether were added to the remaining pellet. After shaking for 30 sec, the tube and its content were centrifuged for 5 min at 500 g. Finally, from the four layers formed, the three top layers were discarded. The bottom layer (including sediment) was placed on a microscope slide. One slide from each participant was created. Slides were examined by microscopy with an Olympus CX21 microscope (Olympus; Volketswil, Switzerland) by the same laboratory technicians for the presence or absence of intestinal protozoa, with 10% of the slides re-examined by an expert microscopist (JTC) for quality control. Identification of the presence or absence of intestinal protozoa was recorded. Laboratory technicians blinded to earlier results re-examined each slide with the Newton Nm1 microscope and recorded the presence or absence of intestinal protozoa. The clip-on version of the reversed-lens CellScope was not used for intestinal protozoa evaluation in the current study given its effective resolution of 14 μm, as described above. Data were double entered into an Excel spreadsheet, transferred into EpiInfo version 3.2 (Centers for Disease Control and Prevention; Atlanta, Georgia, United States of America) and cross-checked. All analyses were conducted using R (R Foundation for Statistical Computing; Vienna, Austria). Prevalences were expressed as proportion and we calculated sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) of the experimental microscopes for each parasite, using conventional light microscopy as ‘gold’ standard. Using logistic regression models, we determined the sensitivity of the experimental microscopes for detecting any eggs, as a function of the egg count determined by conventional microscopy. Linear association of egg count estimates was assessed by Pearson’s correlation coefficient. One slide of filtered urine and a single Kato-Katz thick smear were examined from 226 individuals by conventional microscopy and the two experimental microscopes. Conventional microscopy identified 12 positive Kato-Katz thick smears for S. mansoni (5.3% prevalence), and 90 positive slides for S. haematobium (39.8% prevalence). No infections with Ascaris lumbricoides or Trichuris trichiura were noted. Hookworm eggs were detected in 22 of the Kato-Katz thick smears (9.7% prevalence). However, these were not further subjected to experimental microscopes given the concerns over rapid egg degradation after collection of stool samples and laboratory work-up, pending microscopy analysis, thus affecting the validity of our results [21]. Fig 2 demonstrates S. mansoni eggs visualized by conventional microscopy, the Newton Nm1 microscope, and a clip-on version of the reversed lens CellScope. Table 1 outlines the operating characteristics of the Newton Nm1 microscope and the reversed-lens CellScope for S. mansoni and S. haematobium diagnosis. Sensitivities for S. mansoni and S. haematobium with the Newton Nm1 microscope were 91.7% (95% confidence interval (CI) 59.8–99.6%) and 81.1% (95% CI 71.2–88.3%) respectively, and specificities were 99.5% (95% CI 97.0–100%) and 97.1% (95% CI 92.2–99.1%). S. mansoni and S. haematobium diagnosis with the reversed-lens CellScope demonstrated sensitivities of 50.0% (95% CI 25.4–74.6%) and 35.6% (95% CI 25.9–46.4%), respectively, and specificities of 99.5% (95% CI 97.0–100%) and 100% (95% CI 86.7–100%), respectively. The diagnostic sensitivity for the Newton Nm1 microscope for S. haematobium was 100% for egg counts ≥10 eggs/10 ml of urine. The clip-on version of the reversed-lens CellScope had limited sensitivity at low egg counts, but sensitivity improved as infection intensity increased, culminating in a sensitivity of >90% at 40 eggs/10 ml of urine or higher (Fig 3). Compared with conventional microscopy, estimates of egg counts for S. haematobium had a Pearson’s correlation coefficient of 0.98 using Newton Nm1 and 0.92 using the CellScope. Overall, 121 slides were examined for evidence of intestinal protozoa infection by both conventional microscopy and the Newton Nm1 microscope, with results outlined in Table 2. Based on conventional microscopy, the prevalence of E. histolytica/E. dispar and G. intestinalis of SAF-fixed stool samples subjected to an ether-concentration method were 4.9% and 20.7%, respectively. The Newton Nm1 microscope demonstrated a sensitivity for E. histolytica/E. dispar and G. intestinalis of 83.3% (95% CI 36.5–99.1%) and 84.0% (95% CI 63.1–94.7%), respectively, while specificity was 100%. For other intestinal protozoa, sensitivity of Newton Nm1 microscopy was variable (39–88%), while specificity was excellent (98–100%), as shown in Table 2. Our study shows that handheld microscopes such as the Newton Nm1 portable field microscope and the mobile phone-based CellScope can be successfully implemented into public health settings, after minimal training of laboratory technicians, for the diagnosis of gastrointestinal parasitic infections in rural African settings. Novel diagnostic approaches for common parasitic infections could have a positive impact on the quality of care delivered in resource-constrained settings [3]. Handheld microscopes may be useful tools in such settings as they are lightweight and easily transportable, enabling the delivery of quality diagnostics to individuals in rural, remote, or under-serviced locations rather than transporting people or specimens to distant laboratories. In addition, these devices are battery powered and are helpful in settings where there is no or only intermittent electricity. Indeed, mobile phone-based microscopes have several attributes that make them attractive for use in epidemiologic and public health settings. For example, mobile phone microscopes have the capacity to digitize images such that they can be saved and easily catalogued, or rapidly sent to other practitioners [22]. Digitization of images also allows for the attachment of geographic coordinates that may aid in mapping of infectious diseases and risk profiling of neglected tropical diseases [23]. Lastly, valuable clinical information associated with each image can be stored and catalogued, enabling healthcare providers for patient management. Additionally, the digitization of samples via mobile phone microscopy allows for computer vision and machine learning technology to aid in automated diagnoses and quantification of infectious diseases, such as malaria [24], schistosomiasis [22], giardiasis [25], and filariasis [26]. One potential barrier to widespread implementation is that handheld and mobile phone microscopy only addresses the issue of enabling microscopy in underserviced settings. Developing simple, reliable, and low-cost approaches to standardized sample and slide preparation are required as well, and have received comparably little attention. To date, virtually all studies have evaluated handheld and mobile phone microscopes either in controlled laboratory settings or as used by expert microscopists. Prior to broader implementation, such devices must be rigorously validated in real-world settings and operated by front-line healthcare professionals. Our data confirm and add to findings from previous studies [9] demonstrating that laboratory technicians can reliably use handheld microscopes after minimal training. The Meade Readview handheld microscope was used by laboratory technicians in a Ugandan field study for Schistosoma diagnosis and demonstrated a sensitivity and specificity of 85% and 96%, respectively, compared to conventional microscopy. However, this device was limited by a smaller field of view, limited movement of the stage, and has not demonstrated widespread scale-up since its introduction [9]. Similarly, Ugandan laboratory technicians were trained to operate the Newton Nm1 microscope (as used in this study), to evaluate a set of pre-selected malaria slides, and demonstrated a sensitivity and specificity of 93.5% and 100%, respectively [11], although this study was not conducted in a true field setting. Our study adds to this prior work by implementing and evaluating the handheld microscope devices in a real-world field setting, demonstrating the utility of this device in day-to-day community-based diagnostic testing. In the current study, the clip-on version of the reversed-lens CellScope demonstrated excellent diagnostic specificity for S. mansoni and S. haematobium infection, but only modest sensitivity for these trematode eggs. This observation is consistent with a prior study evaluating the reversed-lens CellScope for S. haematobium diagnosis [14]. We suspect sensitivities were low because users must manually hold the device and guide the lens over the entire surface area of a slide. The CellScopes used in this study were not anchored to a solid structure nor did they utilize a microscope stage such as with conventional microscopy or the Newton Nm1 microscope (Fig 1), however future iterations of this device will have the ability to reliably read an entire slide. Hence, the operators are likely undercounting schistosome eggs due to the challenges of manually maneuvering the device over a microscope slide. Interestingly, the CellScope rapidly gains sensitivity at higher egg counts. For example the sensitivity of CellScope reaches that of the Newton Nm1 microscope (>95%) for S. haematobium diagnosis when used by laboratory technicians at infection intensities of 40 eggs per 10 ml of urine, which is still considered to be a low-intensity infection (Fig 3) [27]. Diagnosing moderate- and high-intensity infections may be useful in clinical settings where worm burdens closely correlate with symptoms [28, 29]. However it is still crucial to have the most sensitive tests available to diagnose even very low intensity infections for disease mapping, epidemiologic surveys, particularly after drug interventions, and rigorous surveillance. Newer versions of the CellScope are currently in development that will enable automated sample scanning and image interpretation. Limitations of our study include only commenting on the presence or absence of intestinal protozoa rather than quantifying these organisms. Future work should evaluate the diagnosis of intestinal protozoa in field settings with experimental microscopes rather than under laboratory conditions. Also, there were no infections with A. lumbricoides or T. trichiura in this setting, and hence, it would be useful to validate the diagnostic performance of these devices for these nematode eggs in other epidemiologic settings given the considerable global health importance of soil-transmitted helminthiasis [2], in addition to other endemic infections. Lastly, our study was restricted to a community-based setting, and future studies should validate the diagnostic capabilities of these devices in clinical environments. In conclusion, handheld light microscopes have considerable potential for use in clinical and public health settings in resource-constrained environments. The clip-on reversed-lens CellScope, while convenient and low-cost to produce, was only modestly sensitive as currently used, however improvements are under development that could make it more appropriate for field deployment in the future. The Newton Nm1 handheld microscope, on the other hand, had good sensitivity and excellent specificity, and hence, could be readily integrated into real-world public health settings to diagnose intestinal parasitic infections.

10.1371/journal.ppat.1000131

HIV-1 Nef Targets MHC-I and CD4 for Degradation Via a Final Common β-COP–Dependent Pathway in T Cells

To facilitate viral infection and spread, HIV-1 Nef disrupts the surface expression of the viral receptor (CD4) and molecules capable of presenting HIV antigens to the immune system (MHC-I). To accomplish this, Nef binds to the cytoplasmic tails of both molecules and then, by mechanisms that are not well understood, disrupts the trafficking of each molecule in different ways. Specifically, Nef promotes CD4 internalization after it has been transported to the cell surface, whereas Nef uses the clathrin adaptor, AP-1, to disrupt normal transport of MHC-I from the TGN to the cell surface. Despite these differences in initial intracellular trafficking, we demonstrate that MHC-I and CD4 are ultimately found in the same Rab7+ vesicles and are both targeted for degradation via the activity of the Nef-interacting protein, β-COP. Moreover, we demonstrate that Nef contains two separable β-COP binding sites. One site, an arginine (RXR) motif in the N-terminal α helical domain of Nef, is necessary for maximal MHC-I degradation. The second site, composed of a di-acidic motif located in the C-terminal loop domain of Nef, is needed for efficient CD4 degradation. The requirement for redundant motifs with distinct roles supports a model in which Nef exists in multiple conformational states that allow access to different motifs, depending upon which cellular target is bound by Nef.

HIV is unique among viral pathogens in its capacity to cause chronic and progressive disease in almost all infected people. To accomplish this, HIV must evade the host immune response, especially cytotoxic T lymphocytes (CTLs), which normally function to lyse virally infected cells. HIV encodes a factor, Nef, which protects HIV infected cells from lysis by anti-HIV CTLs. To prevent CTL lysis, Nef interferes with the expression of host MHC-I, which is needed for CTL recognition of infected targets. A clear understanding of how Nef works has been hampered by its many complex functions. In addition to MHC-I, Nef protein disrupts the expression of multiple other cellular targets using different mechanisms and it is unclear how one protein can accomplish all these tasks. Here, we provide evidence that Nef acts as a highly flexible adaptor protein that is capable of utilizing different protein binding domains depending on which cellular target it is bound to. For example, we present evidence that Nef binding to MHC-I creates novel motifs that result in the recruitment of AP-1 and subsequently β-COP. This series of events results in the mis-localization of MHC-I from the cell surface to cellular degradative compartments, where MHC-I is destroyed.

The HIV-1 accessory protein, Nef, affects the biology of the infected cell in several ways to achieve conditions optimal for viral replication and spread. Nef alters the intracellular trafficking of important immune molecules, such as class I and II major histocompatibility complex proteins (MHC-I and MHC-II), CD4, CD28, and DC-SIGN [1]–[5]. Nef-dependent reduction of surface MHC-I protects HIV-infected primary T cells from recognition and killing by HIV-specific cytotoxic T lymphocytes (CTLs) in vitro [6]. Moreover, disruption of MHC-I expression by HIV-1 and SIV Nef provides a selective advantage under immune pressure in vivo [7]–[10]. CD4 downregulation by Nef is also essential for efficient viral spread. The rapid removal of CD4 prevents viral superinfection [11], and enables optimal viral particle production by eliminating detrimental CD4/HIV envelope interactions in the infected cell [12],[13]. Mutagenesis of protein-protein interaction domains has revealed that Nef uses genetically separable mechanisms to affect MHC-I and CD4 transport. Specifically, disruption of MHC-I surface expression requires an N-terminal α helix, a polyproline repeat, and an acidic domain in Nef [14],[15], while CD4 downregulation requires an intact dileucine motif, two diacidic motifs, and a hydrophobic pocket in Nef [15]–[18]. Amino acids necessary for the myristoylation [19],[20] and oligomerization [21] of Nef are required for the disruption of both MHC-I and CD4 surface expression. Nef has the capacity to affect MHC-I transport at multiple subcellular locations; Nef blocks the export of newly-synthesized MHC-I from the secretory pathway and Nef expression results in a small increase in the rate of MHC-I internalization [22]. To accomplish this, Nef directly binds to the cytoplasmic tail of MHC-I early in the secretory pathway [23]–[26]. The Nef-MHC-I complex then actively recruits the clathrin adaptor protein complex AP-1, which targets MHC-I from the TGN to the endo-lysosomal network where it is ultimately degraded [25]. Recruitment of AP-1 primarily requires a methionine at position 20 in the N-terminal α helical domain of Nef and a tyrosine residue in the cytoplasmic tail of MHC-I. Additionally, the acidic and polyproline domains of Nef have recently been shown to stabilize this interaction [27],[28]. The normal function of AP-1 is to target proteins into the endosomal pathway and then recycle them back to the TGN. Thus, the AP-1 interaction with the Nef/MHC-I complex explains the targeting of MHC-I containing vesicles to the endosomal pathway and to the TGN. However, it does not explain accelerated degradation of MHC-I, hence other cellular factors may be involved [25]. The mechanism of Nef-induced CD4 internalization and degradation has been derived, in part, from correlating Nef function with the requirement for domains in the C-terminal flexible loop region of Nef that bind to cellular factors. The Nef dileucine motif (ExxxLL165) is needed for CD4 internalization and it binds to adaptor protein complexes AP-1, AP-2, and AP-3 [16], [29]–[36]. In addition, a diacidic motif, which is also required, enhances the interaction of Nef with AP-2 [37]. There is separate evidence that this diacidic motif may recruit the H subunit of the vacuolar ATPase (V1H) [38] to promote AP-2 recruitment [39]. Because the normal role of AP-2 is to link cargo to clathrin and promote internalization, it makes sense that this molecule would be necessary and indeed, the involvement of AP-2 has now been confirmed using RNAi knockdown in a number of cell systems [40]–[42]. After CD4 is internalized, it is targeted to lysosomes for degradation. There is evidence that this step requires β-COP [18], a component of COP-1 coats implicated in endosomal trafficking as well as transport through the early secretory pathway [43]–[45]. Specifically, there are defects in the Nef-dependent transport of CD4 into acidified vesicles at the non-permissive temperature in cells harboring a temperature sensitive ε-COP mutant [18]. Nef directly interacts with β-COP [46], and a second diacidic motif in the C-terminal loop domain of Nef has been demonstrated to mediate this interaction [18],[47], although, this result has not been reproducible by another group [48]. To more clearly understand the mechanism of altered MHC-I and CD4 trafficking observed in Nef-expressing cells, we directly compared these two processes in T cells that expressed Nef. We confirmed that Nef primarily affected MHC-I and CD4 at different subcellular locations and we demonstrated that the cytoplasmic tails of the respective molecules dictated which pathway was utilized. Despite the differences in initial trafficking, we found that HLA-A2 and CD4 co-localized in a discrete subset of vesicular structures. Upon further inspection, we determined that these structures also contained markers of late endosomes (Rab7) and to a lesser extent, the lysosomal marker, LAMP-1. Electron microscopy (EM) revealed that CD4 and HLA-A2 were found within MVBs of Nef-expressing T cells. HLA-A2 (but not CD4) was also found in tubulovesicular structures adjacent to the Golgi. In Nef expressing cells, reduction of β-COP expression reduced the targeting of HLA-A2 from the TGN to LAMP-1+ compartments and stabilized CD4 expression within endosomal compartments. Finally, we identified two separate domains within Nef that were necessary for these activities and for β-COP binding. These data support a model in which both MHC-I and CD4 are ultimately targeted to the lysosomes in Nef expressing cells by a final common pathway. It is known that Nef binds to the cytoplasmic tails of both CD4 and MHC–I, but that it affects them differently. To better understand the similarities and differences governing these two pathways, we examined the trafficking of CD4, HLA-A2 and a chimeric molecule in which the wild type HLA-A2 cytoplasmic tail was substituted with the CD4 cytoplasmic tail (HA-A2/CD4). A flow cytometric analysis of steady state surface expression revealed that Nef dramatically reduced steady state surface expression of all three molecules (Figure 1A). Consistent with prior studies, we found that CD4 was rapidly internalized from the cell surface in Nef expressing T cells, whereas wild type HLA-A2 was not (Figure 1B). Substitution of the CD4 tail for the HLA-A2 cytoplasmic tail was sufficient to confer this phenotype (Figure 1C). Conversely, prior studies have shown that Nef disrupts cell surface expression of MHC-I by blocking the transport of newly synthesized MHC-I from the TGN to the cell surface [22],[23]. As shown in Figure 1D, Nef inhibited HLA-A2 forward transport by approximately 75%, whereas CD4 was unaffected at Nef levels that had a clear effect on HLA-A2 transport. Slight effects on CD4 could be observed at higher Nef levels (Figure 1D, lane 8). The substitution of the HLA-A2 cytoplasmic tail with the CD4 tail reduced the ability of Nef to disrupt forward trafficking (Figure 1E). Thus, sequences in the cytoplasmic tails of CD4 and HLA-A2 determine how Nef disrupts their trafficking. To better understand the similarities and differences between MHC-I and CD4 trafficking in Nef-expressing cells, we compared the steady-state distribution of these molecules in T cells using confocal microscopy (Figure 2A). We found that Nef expression caused the bulk of MHC-I to cluster in the perinuclear region where, in agreement with many other studies [14],[30],[49], it co-localized with markers of the TGN (data not shown). Interestingly, we also identified a subset of HLA-A2 that co-localized with CD4 in vesicular structures (Figure 2A; arrows show example vesicles). To further identify these structures, we simultaneously stained for HLA-A2, CD4, and organelle markers using 3-color confocal microscopy (summarized in Table S1). Our results indicated that CD4 was mainly found in discrete vesicular structures, which also contained HLA-A2 (91.9% of the CD4+ vesicles co-localized with HLA-A2, Table S1) and markers of late endosomes and lysosomes. Overall, the best marker for structures containing both HLA-A2 and CD4 was Rab7 (94%, of CD4+ vesicles co-localized with Rab 7, Table S1 and Figure 2A, arrowheads mark example vesicles). CD4 and HLA-A2 were also found to co-localize with markers of lysosomes, such as LAMP-1. However, the vesicles with the most intense LAMP-1 staining did not contain either HLA-A2 or CD4, possibly because of degradation. Consistent with this, the co-localization of HLA-A2 and CD4 was dramatically increased when the cells were treated with bafilomycin, which inhibits degradation in acidic compartments (Figure S1). Thus, the normal steady-state co-localization of HLA-A2 and CD4 in Nef expressing cells was limited because degradation prevented accumulation in this compartment. To further discern these structures, we also examined them using electron microscopy (EM). In agreement with the confocal data, our EM analysis revealed that compared with control cells in which both HLA-A2 and CD4 were found on the cell surface (Figure 2B, panel 1), in Nef-expressing T cells, the majority of CD4 was found in MVBs, co-localizing with HLA-A2 (Figure 2B, panel 2). In addition, we also noted substantial HLA-A2, but not CD4, accumulating in tubulovesicular structures adjacent to Golgi stacks (Figure 2B, panel 3). In separate experiments these structures were also found to contain AP-1 (Figure 2C). Based on these studies, it appears that the majority of HLA-A2 resides in tubulovesicular structures in the region of the TGN with AP-1, whereas at any given time, a small subset can be found in the endosomal compartment with CD4. To further elucidate the similarities and differences between these pathways, we examined the role of known Nef-interacting proteins implicated in intracellular trafficking. AP-1 is a heterotetrameric adaptor protein involved in protein sorting from the TGN and it has been previously demonstrated to interact with MHC-I molecules in Nef expressing HIV-infected primary T cells and to direct MHC-I into the endolysosomal pathway [25]. Nef is also known to interact with β-COP [46], a component of COP-1 vesicles also involved in endosomal trafficking [43]–[45]. Indeed, expression of wild type COP 1 components is needed for targeting CD4 into acidic vesicles in Nef-expressing cells [18]. To compare and contrast the requirement for these factors in Nef-dependent CD4 and HLA-A2 trafficking, we knocked down their expression using lentiviral vectors expressing short hairpin RNAs (shRNAs) [50]. All of these studies were performed in T cells and new cell lines were generated for each experiment to eliminate the possibility that long term growth in culture would select for cells that had compensated for the defect. Using this system, we obtained good knock down of the μ1 subunit of AP-1 and β-COP (Figure 3A–C). (A small apparent effect of shβ-COP on μ1 levels observable in Figure 3A was not significant when adjusted for protein loading in the experiment shown here or in replicate experiments [Figure 3B]. We also did not observe any effect of another siRNA directed against a different target site in β-COP on μ1 expression [Figure S2].) Because β-COP is known to be important for intra-Golgi and ER-to-Golgi trafficking, we asked whether the Golgi structure or MHC-I trafficking were drastically affected by reduced β-COP expression. We found that there was only a small reduction in the normal transport of MHC-I to the cell surface (35% reduction, Figure 3D). In addition, cells lacking β–COP generally maintained overall Golgi structure as assessed by the intracellular localization of giantin, a transmembrane protein normally residing in the cis and medial Golgi [51] (Figure 3E). In contrast, brefeldin A, an inhibitor of an ARF1 GEF necessary for β-COP activity obliterated the normal Golgi staining (Figure 3E, panel 9). The relatively mild phenotype of this knock-down compared to the drastic effects of brefeldin A, suggests that brefeldin A has effects other than just disrupting COP 1 coats by blocking ARF1 activity. Having established that knocking down β-COP allowed relatively normal forward trafficking of HLA-A2, we proceeded to assess the effect of knocking down β-COP or AP-1 in Nef-expressing cells. Consistent with previous publications [25], we found that knocking down the ubiquitously expressed form of AP-1 (AP-1A [52]) largely reversed the effect of Nef on HLA-A2 (p<10−4), but had a smaller and less significant effect (p<0.02) on CD4 surface expression (Figure 4A and 4B). Surprisingly, we also observed that knocking down β-COP expression inhibited MHC-I downmodulation by Nef and had a small but statistically significant effect on CD4 downmodulation (p<10−3; Figure 4A and 4B). The small effect of β-COP on CD4 surface expression indicated that β-COP was not necessary for CD4 internalization and downmodulation from the cell surface. However, further studies were needed to determine whether β-COP was required to degrade the CD4 after it was internalized. Prior studies had determined that expression of β-COP was necessary for acidification of CD4-containing vesicles and thus it was hypothesized that β-COP was needed to target vesicles containing internalized CD4 for lysosomal degradation. Therefore, we asked whether the role of β-COP in MHC-I trafficking was also to promote MHC-I degradation. To examine this, we utilized an assay we had developed, which measures the loss of mature, endo H–resistant HA-tagged HLA-A2 in Nef expressing cells by western blot analysis. This assay system is based on previous data demonstrating Nef-dependent degradation of the mature form of MHC-I in a manner that is reversible by inhibitors of lysosomal degradation [25]. As shown in Figure 4C, under normal, steady state conditions, most of the HLA-A2 is resistant to endo H digestion, indicating that it has matured through the Golgi apparatus (Figure 4C, lane 2). However, when Nef was expressed, we observed a dramatic reduction in total MHC-I and a decrease in the ratio of endo H resistant to sensitive protein (Figure 4C compare lanes 2 and 18, see also Figure S3). Consistent with a role for AP-1, we observed that AP-1A shRNA largely reversed this effect of Nef (Figure 4C, compare lanes 18 and 20. See also Figure 4D for quantification). To detect degradation of molecules containing a CD4 tail, we used HA-A2/CD4 (Figure 1) and found that Nef expression accelerated the degradation of endo H resistant forms of this molecule (Figure 4C, compare lanes 6 and 22). However, we found that there was no effect of reduced AP-1A expression on Nef-dependent degradation of molecules containing the CD4 tail (Figure 4C, compare lanes 22 and 24. See also Figure 4D for quantification). When β-COP expression was reduced, we observed a small increase in the amount of immature, endo H–sensitive protein (Figure 4C, compare lanes 10 and 12), consistent with the 35% reduction in export of MHC-I to the cell surface shown in Figure 3D. However, we also noted that reduction in β-COP expression reduced the Nef-dependent degradation of the mature, endo H resistant form of these molecules (Figure 4C, compare lanes 26 and 28. See also Figure 4D for quantification) implicating β-COP in this pathway. We were also able to confirm the model that β-COP is involved in Nef-dependent CD4 degradation as treating cells with β-COP shRNA reduced the degradation of the A2/CD4 chimeric molecule (Figure 4C, compare lanes 30 and 32. See also Figure 4D for quantification). We next directly examined the effect of reducing β-COP expression on Nef-dependent trafficking by confocal microscopy. For these experiments, cells were infected with HIV or were transduced with Nef-expressing adenoviral vectors and then the fate of internalized CD4 was assessed by confocal microscopy. Using this assay system, we observed fairly rapid internalization of CD4 in Nef-expressing cells, followed by loss of CD4 staining by 30 minutes (Figure 5A, compare control cells in row 1 to Nef-expressing cells in row 3). However, in T cells expressing β-COP shRNA, there was a three-to-four fold increase in the number of CD4-containing vesicles, consistent with a role for β-COP in promoting maturation of these vesicles into degradative compartments (Figure 5A, compare control treated Nef-expressing cells in row 3 to shβ-COP–expressing cells in row 4). Reduction of β-COP expression yielded similar results whether Nef was introduced using HIV infection or via adenoviral vectors (Figure 5B and 5C). Confocal analysis of MHC-I intracellular localization revealed that expression of β-COP shRNA in control cells increased the intracellular accumulation of MHC-I, consistent with the slowing of export we observed in cells deficient in β-COP (Figure 5D, compare rows 1 and 2). Infection with Nef-expressing HIV resulted in the loss of cell surface MHC-I and an increase in intracellular MHC-I, some of which co-localized with LAMP-1 (Figure 5D, compare rows 1 and 3). Under these conditions, reduction of β-COP expression reduced the degree of colocalization with LAMP-1 (Figure 5D, compare rows 3 and 4). To enhance our ability to observe trafficking of MHC-I into LAMP-1+ compartments, we treated the cells with bafilomycin, which inhibits the vacuolar ATPase and thus acidification and degradation within lysosomal compartments. As previously reported [25], bafilomycin treatment enhanced our ability to detect MHC-I in LAMP-1+ compartments in Nef-expressing T cells (Figure 5D, compare rows 3 and 7). The expression of β-COP shRNA decreased LAMP-1 colocalization with MHC-I, consistent with a role for β-COP in targeting MHC-I for degradation in lysosomal compartments in Nef expressing T cells (Figure 5D, compare rows 7 and 8). Similar results were observed whether Nef was introduced using HIV or adenoviral vectors (Figure 5E and 5F). We also examined co-localization of HLA-A2 and CD4 in cells that expressed β-COP shRNA. We observed that reduction of β-COP expression resulted in increased staining of both proteins, and did not disrupt their co-localization (Figure S4). Thus, β-COP was not necessary for targeting these proteins into a common endosomal pathway, but rather was needed for their subsequent targeting into a degradative pathway. To further explore the molecular mechanism for the similarities and differences in MHC-I and CD4 trafficking in Nef-expressing T cells, we asked whether these molecules differed as to how well they bound Nef or cellular factors. As expected, we found that HIV Nef bound to both the HLA-A2 and the CD4 tail (Figure 6A, right panel). However, AP-1 only co-precipitated with molecules containing the HLA-A2 cytoplasmic tail (Figure 6A, right panel). The chimeric molecule with the CD4 cytoplasmic tail did not bind AP-1 in Nef-expressing T cells (Figure 6A, right panel). In these experiments, we noted that the expression level of A2/CD4 was lower than for wild type HLA-A2, which could explain this difference. Therefore, we confirmed these data using a fusion protein containing either HLA-A2 or A2/CD4 directly fused to full length HIV-Nef protein. In previously published experiments it was shown that the HLA-A2/Nef fusion protein co-precipitated AP-1 in a manner that depended on sequences both in Nef and in the HLA-A2 cytoplasmic tail [25]. Here we show again that the HLA-A2 cytoplasmic tail was necessary for this interaction and, moreover, that the CD4 tail could not substitute for it (Figure 6B, right panel). The Nef-β-COP interaction is well-described in the literature [46] and there is evidence that β-COP interacts with a diacidic motif (E154/155) within the Nef C-terminal loop [18]. However, this region of Nef has never been implicated in MHC-I trafficking. To provide further evidence that β-COP is needed to promote MHC-I degradation, we sought to identify a region of Nef that is needed both for MHC-I degradation as well as β-COP binding. We therefore examined a panel of mutations (M20A, V10EΔ17–26 and E62–65Q) that are specifically defective at disrupting MHC-I trafficking [14],[15],[26],[53]. We also examined a Nef mutant, D123G, that is defective at both CD4 and MHC-I downmodulation [21]. The relative activity of these Nef mutants in MHC-I and CD4 downmodulation is shown in Figure 7A and quantified in Figure 7B. We then examined the relative ability of each of these mutant molecules to co-precipitate with β–COP. As shown in Figure 7C, we found that the V10E Δ17–26-Nef, which is defective at MHC-I downmodulation, was also defective at binding to β-COP (compare lanes 3 and 5). Interestingly, this deletion mutant is also defective at interacting with AP-1 [25]. However, the β-COP binding site was separable from the AP-1 interaction site because M20, which is located within the deleted region, is needed for AP-1 interaction [25],[27]), but was not necessary for β-COP binding to Nef (Figure 7C, compare lanes 3 and 4). Mutation of the Nef dimerization motif [D123G, [21]], which disrupts a number of Nef functions, including MHC-I and CD4 downmodulation, also reduced binding to β-COP (Figure 7C, compare lanes 3 and 7). Finally, mutation of the Nef acidic domain (E62–65Q), which disrupts binding to MHC-I [26], AP-1 [27],[28] and PACS-1 [54], did not affect binding to β-COP (Figure 7, compare lanes 3 and 6). As expected, we found that V10EΔ17–26 Nef, which was defective at β-COP binding, was also defective at inducing the degradation of the endo H resistant form of HLA-A2 (Figure 7D, upper panel, compare lanes 3 and 4 with lanes 5 and 6). In contrast, V10EΔ17–26 Nef was not defective at A2/CD4 degradation based on western blot analysis (Figure 7D, lower panel, compare lanes 3 and 4 with lanes 5 and 6). These data suggested that there may be another interaction domain that recruits β-COP to the Nef-CD4 complex to promote CD4 degradation. This would be consistent with the faint band observable in the V10EΔ17–26-Nef mutant immunoprecipitation (Figure 7C, lane 5, longer exposure) and prior publications demonstrating that mutation of E154/155 also affected β-COP binding [47]. Thus, there may be two independent binding sites for β-COP within Nef, each of which governs the degradation of a different cellular factor. To further define the β-COP binding site, and to determine whether there were indeed two β-COP binding sites, we constructed additional Nef mutants. We focused on the arginine residues (R17ER19MR21R22) within the Nef deletion Δ17–26) because previous studies had indicated that arginine rich regions could form β-COP-binding sites [55]. Flow cytometric analysis of MHC-I levels on cells expressing these mutants revealed that the R17/19 pair was necessary for maximal MHC-I downmodulation (Figure 8A and 8B). In contrast, mutation of R21/22 did not significantly affect MHC-I downmodulation (unpublished data). An assessment of Nef-induced degradation by pulse chase analysis of HA-HLA-A2, revealed that mutating this motif also inhibited Nef-dependent degradation (Figure 8C, compare lanes 5 and 7, quantified in Figure 8D). Additionally, mutation of R17/19 reduced, but did not eliminate binding of β-COP to Nef in a manner similar to the effect of the Δ17–26 Nef mutation (Figure 8E, compare lanes 3 and 4). We next examined the diacidic motif (E154/155) previously implicated in β-COP binding. As shown in Figure 8A and 8B, mutation of this motif did not disrupt MHC-I downmodulation, in fact downmodulation was somewhat enhanced. Additionally, we found that mutation of this motif did not reduce MHC-I degradation (Figure 8C, compare lanes 5 and 11, see also quantification in 8D). However, in agreement with prior results, we observed a partial defect in β-COP binding with this mutant (Figure 8E, compare lanes 3 and 6, [18],[47]. However, this defect was less reproducible (observed in two out of four experiments) than that observed with disruption of R17/19 (consistently observed in five out of five experiments), suggesting that binding to R17/19 can mask the defect observed with mutation of E154/155 under certain conditions. To provide additional data supporting the possibility that both sites contributed to β-COP binding, we constructed a double mutant, R17/19 A and E154/155A (R/E). As shown in Figure 8E, lane 5, binding of R/E to β-COP was further reduced relative to binding of Nef proteins containing single mutations in each motif, strongly implicating both motifs in β-COP binding. The phenotype of the double mutant was highly reproducible in 5 out of 5 experiments. Interestingly, the R/E double mutant was not more defective than R17/19A at downmodulating MHC-I (Figures 8A and 8B) or at promoting MHC-I degradation (Figure 8C, compare lanes 7 and 9, quantified in 8D), indicating that Nef did not utilize the E154/155 binding site to recruit β-COP for MHC-I degradation. Conversely, we confirmed prior reports that the E154/155A mutant was defective at CD4 degradation (Figure 9A, compare lanes 3 and 6) and determined moreover that there was no significant effect of mutating R17/19 on CD4 degradation, either alone or in combination with E154/155A (Figure 9A, compare lanes 3 and 4). It is also worth noting that, in contrast to what was observed with HLA-A2, we did not observe a clear correlation between the relative CD4 surface expression and the relative level of total cellular CD4 (compare Figure 8B and 9B), indicating that there was a complex relationship between total cellular CD4 and the fraction expressed on the cell surface. Because the R17/19 motif is directly adjacent to M20, which is necessary for AP-1 recruitment [25],[27], we also examined whether these mutations, which affect β-COP binding, also disrupted AP-1 co-precipitation. To accomplish this, we used our standard AP-1 recruitment assay in which proteins co-precipitating with MHC-I HLA-A2 were detected by western blot analysis. As shown in Figure S5, mutation of R17,19 (and E154/155) decreased AP-1 binding only slightly. Thus, the defects in MHC-I downmodulation and degradation noted with mutation of R17,19 resulted primarily from defects in β-COP binding. Expression of HIV Nef in infected cells protects them from lysis by CTLs and this activity of Nef is due to downmodulation of MHC-I surface expression. The Nef protein also prevents superinfection and promotes viral spread by removing the viral receptor, CD4 from the cell surface (for review see [56]). We provide evidence that sequences in the cytoplasmic tail of these molecules are important for determining whether Nef disrupts their trafficking from the cell surface or at the TGN. These data, that swapping cytoplasmic domains switches the initial pathways taken by HLA-A2 and CD4 in the presence of Nef, may seem somewhat obvious. Nef is always the same and thus one might conclude that this information has to be contained in the modulated protein. However, it was also possible that the ectodomain affected Nef responsiveness by binding to other transmembrane proteins or by altering intracellular trafficking. This was certainly a possibility for MHC-I for which it is clear that the efficiency of peptide loading can affect trafficking and we have found that trafficking rates affect responsiveness to Nef and AP-1 binding [23]. Prior studies have demonstrated that Nef initially binds to hypo-phosphorylated forms of the MHC-I cytoplasmic tail early in the secretory compartment [23], but binding does not affect normal transit through the Golgi apparatus and into the TGN [25]. The Nef-MHC-I complex then recruits the AP-1 heterotetrameric clathrin adaptor protein using a binding site that is created when Nef binds the MHC-I cytoplasmic tail. This binding site requires a methionine from the N-terminal α helix of Nef and a tyrosine residue in the MHC-I cytoplasmic tail [25]. Additionally, there is evidence that this complex is stabilized by the acidic and polyproline domains of Nef [27],[28]. Formation of this complex results in the re-direction of MHC-I trafficking in such a way that it is targeted to lysosomes for degradation [25]. However, cellular proteins that normally bind AP-1 are not degraded, but rather recycled to the TGN (Figure 9C). Here we present new evidence that Nef utilizes β-COP to promote trafficking to degradative compartments (Figure 9C). Knocking down expression of β-COP inhibited the degradation of MHC-I and it did so by blocking the transport of MHC-I from intracellular vesicles to LAMP-1+ compartments. We also provide results here that confirm β-COP is necessary for degradation of CD4 in lysosomal compartments. Thus, we propose that AP-1 and AP-2 deliver MHC-I and CD4 respectively to endosomal compartments where β-COP displaces AP-1 and AP-2 to target MHC-I and CD4 for lysosomal degradation (Figure 9C). As described above, we found that knocking down β-COP with shRNA resulted in stabilization of internalized CD4, however the effect on CD4 surface expression was small, but still significant. In contrast, there was a greater effect of β-COP knockdown on HLA-A2 surface expression. This might suggest that the role of β-COP in the modulation of these targets was different, rather than the same. However, this apparent paradox can be explained by our model shown in Figure 9C. As indicated, differences in response to β-COP knockdown can be explained by differences in the intracellular pathways of these proteins before they interact with β-COP. MHC-I is engaged in an AP-1-dependent endosome-to-TGN loop, and MHC-I could “leak” out to the cell surface from the TGN in the absence of β-COP, whereas CD4 may be unable to return to the cell surface from its endosomal compartment. Consistent with this, we also noted a lack of correlation between degradation and surface expression of CD4 (but not MHC-I) when Nef mutants that were defective in β-COP binding were examined. These data indicate that there is a complex relationship between total cellular CD4 and the fraction that is present on the cell surface and thus intracellular pools need to be directly examined to assess degradation rather than relying on surface expression as an indicator of the efficiency of this process. It is also noteworthy that shRNA knockdown of β-COP did not fully reverse Nef-dependent MHC-I and CD4 degradation. This may have resulted from incomplete knockdown of β-COP. However, we also observed a similar phenotype with Nef mutants defective at β-COP binding. Failure to fully reverse degradation may be secondary to a default degradative pathway that exists for all proteins delivered to endosomal pathways. Alternatively, there may be other ways Nef targets these proteins to lysosomes, which have yet to be identified. Our studies indicate that there are at least three domains needed for Nef to interact efficiently with β-COP. One of these domains (D123), is required for dimerization of Nef and is needed to affect a variety of Nef functions [21]. Another region lies within the N-terminal α helical domain of Nef that is specifically required for disruption of MHC-I trafficking and for interactions with AP- 1 [25]. This binding site for β-COP is distinct from that used by AP-1, because recruitment of β-COP does not require Nef's acidic domain or Nef M20, whereas AP-1 does [25],[27]. The fact that these Nef mutants bind β-COP, but are still defective at MHC-downmodulation [53] makes sense, because these mutants are also unable to bind the MHC-I cytoplasmic tail [26]. Additional mutants, which focused on the highly conserved stretch of arginines in the N-terminal alpha helical domain of Nef (R17XRMRR22), revealed that the regions involved in AP-1 and β-COP binding were very closely apposed. However, we determined that mutation of R17/19 affected primarily β-COP binding, with only a minimal effect on AP-1 interaction. Thus, these two Nef-interacting proteins have distinct and separable amino acid requirements for binding. The identification of a β-COP binding domain within a region of Nef that is also required for Nef to accelerate MHC-I degradation confirms the requirement for β-COP in this pathway. In addition, the residual binding of β-COP to these Nef mutants provided suggestive data that another binding site for β-COP existed. Indeed, we were able to confirm prior evidence that a diacidic motif within the C-terminal loop of Nef also promoted an interaction with β-COP and that mutation of this motif reduced CD4 degradation [47]. Finally, we demonstrated that mutation of both the RXR and the diacidic motifs resulted in the greatest defect in β-COP binding. The double mutant did not however result in a greater defect in either MHC-I or CD4 degradation, indicating the role of each motif is distinct and not additive. The discovery of two distinct β-COP binding motifs helps explain why some groups could not confirm the role of the diacidic motif in β-COP binding [48] as both motifs need to be mutated to reliably eliminate an interaction between β-COP and Nef. There is precedent for such redundancy. For example, there are two AP-1 binding sites within Nef; a dileucine motif within the C-terminal flexible loop [16],[31],[32],[33] as well as a second site that forms upon binding of Nef to the MHC-I cytoplasmic tail. Despite the presence of two AP-1 signals, only one is active in the context of the natural Nef-MHC-I complex [25],[27]. The dileucine motif in the C-terminal flexible loop can become activated to affect MHC-I transport, but only when Nef is artificially fused to the MHC-I cytoplasmic tail [27]. This result indicates there is no inherent inability of this signal to affect MHC-I traffic but rather that something else, such as the structure of the natural complex, causes the dileucine motif to be inactive [27]. The dileucine motif at position 164 is located close to the diacidic motif at position 154 that binds β-COP to promote CD4 degradation. The fact that both of these motifs are inactive when Nef is bound to MHC-I, suggests that much of the C-terminal flexible loop region of Nef is inaccessible under these conditions. Thus, Nef behaves as though it assumes different structural forms in different contexts to differentially expose distinct trafficking signals. We also present evidence that knock down of β-COP yielded a distinct phenotype from BFA treatment. As described above, BFA is a chemical inhibitor of ARF1, that is known to trigger the reversible collapse of the cis-medial Golgi compartments to the ER [57]–[59] by inhibiting an ARF-specific guanine nucleotide-exchange protein (ARF-GEF) [60],[61]. Because ARF1 activity is necessary for recruitment of β-COP to membranes [62], it was possible that the dramatic effects of BFA resulted from the inability for β-COP to function normally. However, our results demonstrating that knockdown of β-COP had no effect on overall Golgi structure indicate that the dramatic effects of BFA are not due solely to disruption of β-COP function in the Golgi. Given the important role of β-COP in the Golgi, it is surprising that β-COP bound to Nef does not also affect transport of MHC-I through the ER/Golgi. It is possible that our inability to detect an effect of Nef on early transport of MHC-I [25] may be a result of the cell type chosen for these studies. T cells, which are an important natural target of HIV, normally traffic MHC-I through the early secretory pathway slowly [23] and thus it might be difficult to further reduce the trafficking speed through an interaction with β-COP. Interestingly, another group has reported a reduced ER-Golgi exit rate for MHC-I in Nef-expressing HeLa cells [63], which normally transport MHC-I more rapidly than T cells [23]. We have made similar observations in astrocytoma cells expressing higher levels of Nef than typically needed to observe MHC-I downmodulation (Roeth and Collins, unpublished observations). Further studies will be needed to determine whether this effect of Nef plays a role in more physiologically relevant cell systems and whether this effect of Nef might be dependent on β-COP expression. A recent report indicates that the effect of Nef on internalization of MHC-I, which is only minimally apparent in our system, occurs via a PI3-kinase dependent pathway [64]. This publication reported that CEM cells, which were used in our study, have less PTEN (a phosphatase that inhibits PI3-kinase) than another T cell line used in their study (H9). This deficiency might make it relatively more difficult for us to detect an effect of chemical PI3-kinase inhibitors, but would not affect our ability to detect a PI3-kinase-dependent trafficking pathway. In fact, one would expect the opposite, that the PI3-kinase-dependent pathway would be more active in our system. However, we have found that Nef has a relatively small effect on internalization of MHC-I, and mainly affects MHC-I protein export and degradation. These data have been corroborated in HIV-infected primary T cells [22],[26], which were also found to much lower levels of PTEN than H9 cells did [64]. From a teleological perspective, it makes sense that Nef would have evolved to target early forms of MHC-I, which harbor antigens derived from the newly synthesized viral proteins. Older forms of MHC-I already on the cell surface would be bound to normal cellular antigens and would in fact be protective as they would inhibit killing by natural killer cells that are stimulated to lyse cells with abnormally low MHC-I expression. On the other hand, it makes sense that Nef, an early viral protein, would have evolved to target surface CD4 to rapidly and efficiently remove CD4 in order to prepare the cell for rapid release of viral particles and to render the cell resistant to re-infection. Meanwhile, a late protein, Vpu, is expressed in infected cells and specifically targets the newly synthesized CD4 for degradation, preventing any additional CD4 from reaching the cell surface [65]. In sum, we have found that the HIV Nef protein commandeers the cellular trafficking machinery efficiently by utilizing their natural activities for abnormal purposes. The fact that these pathways may end in a final common step raises the important possibility that inhibitors might be developed that could block multiple Nef functions. CEM T cells stably expressing HA-tagged HLA-A2 (CEM HA-HLA-A2) have already been described [25]. Cell lines stably expressing YFP-tagged Rab7 or HA-HLA-A2/CD4 were made by transducing cells with murine retroviral constructs (MSCV YFP-Rab7 or MSCV HA-A2/CD4) as previously described [22], followed by culture in selective media. MSCV YFP-Rab7 was constructed by cloning a filled-in a Kpn I-Xho I fragment from pEYFP-Rab7 [66] into MSCV puro [67]. MSCV HA-A2/CD4 was constructed using PCR mutagenesis. The first round PCR produced two products: the first utilized 5′ primer (primer 1) 5′-CGGGATCCACCATGCGGGTCACGGCG-3′ and 3′ primer (primer 2) 5′-CTCTGCTTGGCGCCTTCGGTGCCACATCACAGCAGCGACCAC-3′ with MSCV HA-HLA-A2 as the template [25]. The second utilized 5′ primer (primer 3) 5′-GTGGTCGCTGCTGTGATGTGGCACCGAAGGCGCCAAGCAGAG-3′ and 3′ primer (primer 4) 5′-CCTCGAGTCAAATGGGGCTACATGTCTTCTGAAATCGGTGAGGGCACTGG-3′ using CD4 as the template. The second round utilized primers 1 and 4 from the previous PCR reactions plus 1 µl of each purified first round PCR reactions as template. The resulting product was digested with BamHI and XhoI and ligated into MSCV 2.2 [67] digested with BglII and XhoI. MSCV A2/Nef has been described [26]. MSCV HA-A2/CD4/Nef was constructed using a PCR mutagenesis approach. The first round PCR produced two products: the first utilized 5′ primer (primer 1) 5′-CGGGATCCACCATGCGGGTCACGGCG-3′ and 3′ primer (primer 2) 5′-CCACTTGCCACCCATACTAGTAATGGGGCTACATGT-3′ with MSCV HA-A2/CD4 as the template. The second utilized 5′ primer (primer 3) 5′-ACATGTAGCCCCATTACTATGATGGGTGGCAAGTGG-3′ and 3′ primer (primer 4) 5′- GCGAATTCTCAGCAGTTCTTGAAGTACTC-3′ with NL4-3 Nef open reading frame as template. The second round utilized primers 1 and 4 from the previous PCR reactions plus 1 µl of each purified first round PCR reactions as template. The resulting product was digested with BamHI and EcoRI and ligated into MSCV IRES GFP [68] digested with BglII and EcoRI. Nef mutants were made by using the PCR mutagenesis approach described previously (Wonderlich et al. 2008). The mutagenesis primers were as follows: R17/19A 5′-TGGCCTACTGTAGCGGAAGCAATGAGACGAGCT-3′ and EE154–155AA 5′-GTTGAGCCAGATAAGGTAGCAGCGGCCAATAAAGGAGAGA-3′. Each primer, plus its reverse complement were utilized together with additional 5′ and 3′ primers to generate the mutated product. Wild type NL4-3 Nef [MSCV A2/Nef IRES GFP (Roeth et al 2005)] was used as a template for the PCR reaction, except for the double mutant, R17/19A/EE154–155AA, in which the MSCV R17/19A Nef IRES GFP was used as the template. Each mutated PCR product was digested and cloned into MSCV IRES GFP [68] as described previously (Wonderlich et al. 2008). The FG12 shRNA lentiviral vectors were constructed as previously described [50]. Briefly, complementary primers were annealed together and ligated into vector pRNAi [69] digested with BglII and HindIII. The sequences of the primers were as follows (the target sequence is underlined): shNC (an siRNA directed at GFP, with several base changes [25])- sense 5′-GATCCCCGCTCACACTGAAGTTAATCTTCAAGAGAGATTAACTTCAGTGTGAGCTTTTTGGAAA-3′, antisense 5′-AGCTTTTCCAAAAAGCTCACACTGAAGTTAATCTCTCTTGAAGATTAACTTCAGTGTGAGCGGG-3′, shβ-COP- sense 5′-GATCCCCTGAGAAGGATGCAAGTTGCTTCAAGAGAGCAACTTGCATCCTTCTCATTTTTGGAAA-3′, antisense 5′-AGCTTTTCCAAAAATGAGAAGGATGCAAGTTGCTCTCTTGAAGCAACTTGCATCCTTCTCAGGG-3′; shμ1A- (a mixture of two lentiviruses was used) (1) sense 5′GATCCCCTGAGGTGTTCTTGGACGTCTTCAAGAGAGACGTCCAAGAACACCTCATTTTTGGAAA-3′, antisense 5′-AGCTTTTCCAAAAATGAGGTGTTCTTGGACGTCTCTCTTGAAGACGTCCAAGAACACCTCAGGG-3′, (2) sense 5′- GATCCCCCGACAAGGTCCTCTTTGACTTCAAGAGAGTCAAAGAGGACCTTGTCGTTTTTGGAAA-3′, and antisense 5′- AGCTTTTCCAAAAACGACAAGGTCCTCTTTGACTCTCTTGAAGTCAAAGAGGACCTTGTCGGGG-3′. The pRNAi constructs were digested with XbaI and XhoI to remove the promoter and shRNA sequence. The resulting fragment was ligated into FG12 [50], digested with XbaI and XhoI. Adenovirus was prepared by the University of Michigan Gene Vector Core facility. Adenoviral and HIV (HXB-EP [6]) transductions of T cells [25] or 373 mg astrocytoma cells [49] have been described previously. Murine retroviral vector (MSCV) expressing Nef was prepared as described previously (Roeth et al. 2005), except that in some cases the retroviral vector supernatants were concentrated by spinning at 14000 RPM for four hours at 4°C. The viral pellet was then resuspended in media to yield a twenty-fold concentrated stock. Lentiviruses expressing shRNA were generated using an approach similar to that already described [50]. Briefly, 293 cells were transfected with the FG12 constructs described above plus pRRE [70], pRSV-Rev [70] and pHCMV-G [71] using Lipofectamine 2000 (Invitrogen). Supernatants from the transfected cells were collected and used to transduce CEM T cells using a spin-transduction protocol. Intact cells were stained for flow cytometry analysis as previously described [24]. Briefly, HLA-A2 was detected with BB7.2 [72] that had been purified as previously described [22]. Endogenous CD4 was detected using RPA-T4 from Serotec. The secondary antibody was goat anti-mouse-phycoerythrin (BioSource, 1∶250). For experiments using the GFP-expressing FG12 lentivirus for shRNA expression, the GFP-positive cells were gated to identify the subset of transduced cells (generally >90% of cells). Endocytosis assays were performed as previously described with minor modification [22]. Briefly, cells were washed once with Endocytosis Buffer [D-PBS, 10 mM HEPES, 10 µg/ml BSA (NEB)], then stained with primary antibody (described above) for 20 minutes on ice. After washing, the cells were resuspended in RPMI supplemented with 10% fetal bovine serum, 10 mM HEPES buffer, 2 mM L-glutamine, penicillin and streptomycin (R10) (pre-warmed to 37°C) and replicate aliquots were removed and placed on ice for each time point. Cells were then washed and stained with goat anti-mouse-phycoerythrin (BioSource, 1∶250) and the samples were analyzed using a FACScan flow cytometer (Becton Dickinson). Flow cytometry data was processed using FlowJo v4.4.3 software (Treestar Corp.). The mean fluorescence at time zero was set to 100%, and this value was used to calculate the relative surface staining at each subsequent time point. CEM cells transduced with adenoviral vectors as previously described [22] were first incubated in pre-label media [RPMI –Cys –Met (Specialty Media, Inc.)+10% dialyzed FBS (Invitrogen)] for 15 minutes at 37°C. Pulse labeling was performed in pre-label media with 150–200 µCi/ml Pro-mix-L [35S] (>1000 Ci/mmol; Amersham Pharmacia) for 30 minutes at 37°C. The cells were then chased in R10 media for 15 minutes at 37°C, followed by two washes with D-PBS. To label the protein that reached the cell surface, the cells were resuspended in D-PBS containing 0.5 mg/ml EZ-Link sulfo-NHS-LC-Biotin (Pierce), and incubated at 37°C for 1 hour. Surface biotinylation was quenched by washing the cells in D-PBS+25 mM Lysine (Fisher). For Figure 1D, immunoprecipitation of proteins from cell lysates was performed as previously described [25], except that one-third of the total lysate was used for the HLA-A2 immunoprecipitation while two-thirds of the material was used to recover CD4. For immunoprecipitations of 35S labeled proteins, 5 µg of BB7.2 and 2.5 µg RPTA4 (BD Pharmingen) were used for HLA-A2 and CD4 respectively. In Figure 1E and 3D, the total cell lysate was immunoprecipitated with anti-HA ascites (HA.11, Covance). For Figures 1D, 1E and 3D, recovered proteins were released from the beads by boiling in 100 µl of 10% SDS. One third was analyzed directly by SDS-PAGE (Total). The remaining two thirds was brought to a total volume of 1 ml in RIPA Buffer [25], and 40 µl of avidin-agarose (Calbiochem) was added to recover biotinylated proteins. After 2 hours at 4°C, the beads were washed three times with 1 ml RIPA buffer and proteins were separated by SDS-PAGE (Surface). Adeno-transduced CEM cells were adhered to glass slides, fixed, permeabilized, and stained for indirect immunofluorescence as previously described [25]. Bafilomycin treatment was performed as described previously [25]. The following antibodies were utilized to localize proteins via microscopy: Figure 2, and Figures S1 and S4; anti-CD4 (S3.5, Caltag Laboratories) and anti-HLA-A2 (BB7.2); Figure 3; anti-giantin (Covance); Figure 5; anti-CD4 antibody (S3.5, Caltag Laboratories), anti-LAMP-1 (H4A3, BD Pharmingen) and anti-HLA-A2 (BB7.2). Secondary antibodies were obtained from Molecular Probes and were used at a dilution of 1∶250: Giantin, Alexa Fluor 546 goat anti-rabbit; CD4, Alexa Fluor 546 goat anti-mouse IgG2a; LAMP-1, Alexa Fluor 546 goat anti-mouse IgG1; BB7.2 (Figures 2, 5D and S4), Alexa Fluor 647 goat anti-mouse IgG2b; BB7.2 (Figure S1), Alexa Fluor 488 goat anti-mouse IgG2b. See Table S2 for a summary of antibodies used to gather data for Table S1. For the microscopy based internalization assay in Figure 5A, CEM T cells were allowed to adhere to glass slides, and placed on ice. The cells were washed once with wash buffer (D-PBS, 10 µg/ml BSA (NEB) and 2% goat serum), incubated with anti-CD4 antibody (S3.5, Caltag Laboratories, IF, 1∶25) for 20 minutes, washed once with wash buffer, incubated with Alexa fluor 546 goat anti-mouse IgG2a (Molecular Probes, 1∶250) for 20 minutes and washed once with wash buffer. The zero time point was fixed with 2% paraformaldehyde, while the remaining time points incubated at 37˚C for the indicated time. The cells were then fixed with 2% paraformaldehyde. Images were collected using a Zeiss LSM 510 confocal microscope and processed using Adobe Photoshop software. Three-dimensional projections of cells were generated from Z-stacks using Zeiss LSM Image Examiner software. Otherwise, single Z sections through the center of the cell were displayed. Electron microscopy with CEM cells transduced with adenovirus was performed by the Harvard Medical School (HMS) Electron Microscopy Facility. Frozen samples were sectioned at −120°C, the sections were transferred to formvar-carbon coated copper grids and floated on PBS until the immunogold labeling was carried out. The gold labeling was carried out at room temperature on a piece of parafilm. All antibodies and protein A gold were diluted in 1% BSA. The diluted antibody solution was centrifuged 1 minute at 14,000 rpm prior to labeling to avoid possible aggregates. Grids were floated on drops of 1% BSA for 10 minutes to block for unspecific labeling, transferred to 5 µl drops of primary antibody and incubated for 30 minutes. The grids were then washed in 4 µl drops of PBS for a total of 15 minutes, transferred to 5 µl drops of Protein-A gold for 20 minutes, washed in 4 µl drops of PBS for 15 minutes and 6 µl drops of double distilled water. Contrasting/embedding of the labeled grids was carried out on ice in 0.3% uranyl acetete in 2% methyl cellulose for 10 minutes. Grids were picked up with metal loops (diameter slightly larger than the grid) and the excess liquid was removed by streaking on a filter paper (Whatman #1), leaving a thin coat of methyl cellulose (bluish interference color when dry). The grids were examined in a Tecnai G2 Spirit BioTWIN transmission electron microscope and images were recorded with an AMT 2k CCD camera. For the western blot analysis in Figures 3A, 4C, 7D, 9A, S2, and S3, cells were lysed in PBS 0.3% CHAPS, 0.1% SDS pH 8, 1 mM PMSF, normalized for total protein and separated by SDS-PAGE. Endo H (NEB) digestion was performed according to the manufacturer's protocol. Staining of the western blot was performed using anti-Nef (AG11, [73]) and anti-β-COP (M3A5 [74]), which were purified as previously described [22]. Additional antibodies used were HA (Covance) and μ1 (RY/1 [75]). The secondary antibody for anti-Nef, β-COP, and HA was HRP-rat anti-mouse IgG1 (Zymed) and for anti-μ1 was HRP-goat anti-rabbit (Zymed). For Figure 6B, the IP-western experiment was performed as previously published [26]. Briefly, parental CEM T cells were spin-transduced with murine retroviral supernatant expressing either empty vector, A2/Nef or A2/CD4/Nef. At 72 hours post transduction, the cells were incubated in 20 mM NH4Cl for 4 hours. The cells were then treated with DTBP (Pierce) for 40 minutes, quenched per the manufacturer's protocol, and lysed in PBS with 0.3% Chaps and 0.1% SDS. The lysate was pre-cleared and immunoprecipitated with HLA-A2 with BB7.2 chemically crosslinked protein A/G beads (Calbiochem) [25]. The immunoprecipitates were washed in TBS with 0.3% CHAPS and 0.1% SDS. A more stringent IP protocol was used in Figures 6A, 7C, 8E, and S5. For these experiments, CEM cells were transduced with control, wild type Nef, or mutant Nef expressing adenovirus (Figure 6A and 7C) or concentrated MSCV (Figures 8E and S5). At 48 hours post-transduction, the cells were incubated in 20 mM NH4Cl for 16 hours. The cells were not crosslinked and were lysed in digitonin lysis buffer (1% digitonin (Wako), 100 mM NaCl, 50 mM Tris pH 7.0, 1 mM CaCl2, and 1 mM MgCl2). After pre-clear, the lysates were immunoprecipitated with either BB7.2 (Figures 6A and S5) or M3A5 (Figures 7C and 8E) crosslinked to beads. The immunoprecipitates were eluted and analyzed by western blot as described previously [26]. A total of 30 million CEM T cells transduced with wild type or mutant Nef using concentrated MSCV as described above were pulse labeled for 30 minutes with [35S]-methionine and cysteine. Half of the cells were collected as the zero time point and stored at −20 degrees. The remaining cells were then chased for 12 hours in RPMI, collected and stored at −20 degrees. Lysates were generated in lysis buffer (PBS 0.3% CHAPS, 0.1% SDS pH 8, 1 mM PMSF) and precleared over night. They were immunoprecipitated for two hours with an anti-HLA-A2 antibody (BB7.2) and washed once in radioimmunoprecipitation assay (RIPA) buffer (50 mM Tris pH 8, 150 mM NaCl, 1% NP-40, 0.5% deoxycholate, 0.1% SDS). The immunoprecipitates were then eluted by boiling in 10% SDS, reprecipitated with an antibody against HA (HA.11, Covance), and washed two times in RIPA buffer. The final immunoprecipitates were then separated by SDS-PAGE, the gel was dried down and analyzed using a phosphorimager.

10.1371/journal.pcbi.1002193

Ligand-Induced Modulation of the Free-Energy Landscape of G Protein-Coupled Receptors Explored by Adaptive Biasing Techniques

Extensive experimental information supports the formation of ligand-specific conformations of G protein-coupled receptors (GPCRs) as a possible molecular basis for their functional selectivity for signaling pathways. Taking advantage of the recently published inactive and active crystal structures of GPCRs, we have implemented an all-atom computational strategy that combines different adaptive biasing techniques to identify ligand-specific conformations along pre-determined activation pathways. Using the prototypic GPCR β2-adrenergic receptor as a suitable test case for validation, we show that ligands with different efficacies (either inverse agonists, neutral antagonists, or agonists) modulate the free-energy landscape of the receptor by shifting the conformational equilibrium towards active or inactive conformations depending on their elicited physiological response. Notably, we provide for the first time a quantitative description of the thermodynamics of the receptor in an explicit atomistic environment, which accounts for the receptor basal activity and the stabilization of different active-like states by differently potent agonists. Structural inspection of these metastable states reveals unique conformations of the receptor that may have been difficult to retrieve experimentally.

G protein-coupled receptors (GPCRs) constitute one of the most important classes of cellular targets owing to their known response to a host of extracellular stimuli, and consequent involvement in numerous vital biological processes. Compelling evidence herein referred to as ‘functional selectivity’ shows that ligands with varied efficacies can stabilize different GPCR conformations that may selectively interact with different intracellular proteins, and therefore induce different biological responses. Understanding how this selectivity is achieved may lead to the discovery of drugs with improved therapeutic properties. We propose here a computational strategy that enables identification of the specific conformations assumed by a GPCR when interacting with ligands that elicit different physiological responses. Not only can these computational models help bridge the information gap in structural biology of GPCRs, but they can be used for virtual screening, and possibly lead to the structure-based rational discovery of novel ‘biased’ ligands that are capable of selectively activating one cellular signaling pathway over another.

G-protein coupled receptors (GPCRs) are versatile signaling proteins that functionally couple a host of extracellular stimuli to intracellular effectors, thus mediating several vital cellular responses. The majority of marketed drugs act as agonists, inverse agonists, or antagonists at these receptors depending on whether they increase, reduce, or have no effect on the so-called ‘basal activity’ that characterizes unliganded GPCRs for diffusible ligands. Not only can a specific GPCR activate different G-protein or arrestin isoforms [1], but a single ligand can display different efficacy for different signaling pathways, an observation that has been dubbed “functional selectivity”, “agonist trafficking”, “biased agonism”, “differential engagement”, or “protean agonism” in the literature [2]–[6]. At the molecular level, a simple explanation for this phenomenon is that ligands with varied efficacies can shift the conformational equilibrium of a GPCR towards different conformations of the receptor, which in turn can activate one or another intracellular protein. Although several spectroscopy studies (e.g., for the β2-adrenergic receptor, herein referred to as B2AR, see [7]–[9]) have been instrumental in showing that ligands with different efficacies stabilize GPCR conformational states that are structurally and kinetically distinguishable, perhaps the most direct evidence of ligand-induced conformational specificity comes from the recent high-resolution crystallographic structures of several different ligand-bound GPCRs. In the majority of cases, these structures were obtained in the presence of an inverse agonist, and therefore in an inactive state. Only very recently have high-resolution crystal structures of agonist-bound GPCRs started to appear in the literature [10]–[15]. However, possibly restrained by crystallization conditions, not all these agonist-bound structures present the features that are usually attributed to an active GPCR conformation, most typically: the large outward movement of transmembrane helix 6 (TM6) with respect to the center of the receptor helical bundle, which is accompanied by the disruption of an important salt bridge between the conserved D/E3.49-R3.50 pair and E6.30, commonly referred to as the “ionic lock”. Residue numbering here and throughout the text follows the Ballesteros-Weinstein notation [16]. According to this notation, each residue is indicated by a two-number identifier N1.N2 where N1 is the number of the transmembrane helix, and N2 is the residue number on that helix relative to its most conserved position, which is designated N2 = 50. We direct the reader elsewhere (e.g., [17], [18]) for recent reviews of all the relevant structural changes that have been attributed by various biophysical techniques to active forms of GPCRs. A different extent of structural rearrangement was noted at the binding site of high-resolution crystal structures of GPCRs depending on the type of ligand to which they were bound. For instance, only minor local structural changes were noted between the high-resolution crystal structures of the B2AR in the presence of inverse agonists such as carazolol [19], timolol [20], ICI-118,551 [21], or a compound deriving from virtual screening [21] and the neutral antagonist alprenolol [21]. Slightly more pronounced differences were noted by comparing these inverse agonist/antagonist-bound binding pockets with those stabilized by full agonists (i.e., either the covalently-bound ligand FAUC50 [11] or BI-167107 [10]). Among them, the most notable differences were the hydrogen bonding contacts that only agonists formed with S5.42 and S5.46 on TM5. Similar interactions helped discriminate between inverse agonist-bound crystal structures of the β1-adrenergic receptor (B1AR) and structures obtained in the presence of full agonists (e.g., isoprenaline or carmoterol) [13]. Notably, only one of these two hydrogen bonds involving TM5, specifically the interaction with S5.42, was also present in structures stabilized by the partial agonists salbutamol or dobutamine, suggesting a distinguishable binding mode between full and partial agonist structures [13]. Analogous to the cases of the B1AR and B2AR, where specific residues (i.e., S5.46) are found to bind uniquely to agonists, key residues (S7.42 and H7.43) that bind agonists (either adenosine or NECA) but not antagonists (ZM241385) were revealed by the very recent crystal structures of a thermostabilized construct of the adenosine A2A receptor [15]. Unlike another recent crystal structure of this receptor stabilized by both T4-lysozyme and the conformationally selective agonist UK-432097 [12], these agonist-bound structures did not display changes at the cytoplasmic side that resemble those of an active state of a GPCR. In addition to the crystal structure of the adenosine A2A receptor bound to UK-432097 [12], these more marked changes at the cytoplasmic side have so far only been observed in the high-resolution crystal structures of opsin [22], [23], Meta II rhodopsin [14], and the nanobody-stabilized B2AR [10]. Despite these recent remarkable achievements in structural biology of GPCRs, the majority of pharmacologically relevant ligands of these receptors do not appear to be ideally suited for the stabilization and crystallization of these receptors, most likely because of their low affinity, slow off-rate, and poor solubility. Not only might this prevent the identification of physiologically relevant conformational states of a given GPCR, but it is considered a limiting bottleneck for the characterization of different structures of these receptors. Molecular dynamics (MD) simulations can help to fill this information gap by enabling an atomic-level characterization of ligand-specific conformations that are impossible or difficult to retrieve experimentally. Moreover, these simulations allow extension of static structural data into dynamic representations, thus laying the basis for a mechanistic understanding of the selective activation of GPCR-mediated signaling pathways. To enable characterization of large conformational changes within the limited timescales commonly accessible to MD simulations, and to evaluate the extent to which ligands with different efficacies affect the free-energy landscape of GPCRs, we implemented a computational strategy employing a combination of different adaptive biasing techniques. Specifically, we used well-tempered metadynamics [24] to identify metastable states of a GPCR along putative activation pathways between inactive and active crystallographic states determined by adiabatic biased MD. We tested the accuracy of this strategy in reproducing crystallographic [19], [21] and/or spectroscopic [7]–[9] data available for the B2AR in its interaction with either a full agonist (i.e., epinephrine), a weak partial agonist (i.e., dopamine), a very weak partial agonist (i.e., catechol), two inverse agonists (i.e. ICI-118,551 and carazolol), or one neutral antagonist (i.e., alprenolol). The results show a clear ligand-induced modulation of the free-energy landscape of the receptor with shifts in the conformational equilibrium towards inactive or active conformations depending on the physiological response elicited by the simulated ligand. A model of the B2AR (Figure S1 was prepared starting from one of the available crystal structures of this receptor (PDB ID: 2RH1), removing the lysozyme insertion, and modeling the missing intracellular loop 3 (IL3) with the Rosetta ab-initio loop modeling protocol [25]. The intracellular loop 2 (IL2), which is probably misfolded [26], [27] in the inactive structure of the B2AR (2RH1), but in a helical conformation in the active nanobody-stabilized crystal (3P0G) of the receptor, was also replaced by the lowest-energy Rosetta model with a helical fold. The resulting receptor model was embedded into an explicit 1-palmitoyl-2-oleoyl-sn-glycero-3-phosphocholine (POPC)/10% cholesterol membrane bilayer using a pre-equilibrated 8×8×10 nm patch hydrated with SPC/E water molecules, and the procedure described in [28]. As found in the crystal structure [19], one palmitoyl group was covalently attached to a C-terminal residue (Cys 341) of the receptor before insertion in the membrane. The system was then hydrated with SPC/E water molecules [29] and Na+ and Cl– ions were added to ensure charge neutrality. The resulting system of ∼50,000 total atoms was equilibrated with unbiased MD simulations for 20 nanoseconds (ns) using the Optimized Potentials for Liquid Simulations all-atom (OPLS-AA) force field [30] for the receptor and united-atoms Berger parameters for the lipids [31]. The Gromacs 4.0.7 [32] package enhanced with the Plumed plug-in [33] was used for all simulations. Specifically, NPT simulations were carried out under periodic boundary conditions, using the Parrinello-Rahman algorithm [34] with a time constant of 1.0 ps and a reference pressure of 1 bar to control pressure, and the Nose-Hoover [35] algorithm with a time constant of 1.0 ps to maintain a constant temperature of 300 K. Prior to production runs (summarized in Table S1), the system was equilibrated by a series of three 0.2 ns runs with progressively weaker restraints on the protein backbone followed by a 3.0 ns unconstrained equilibration. We used the standard Gromacs leap-frog [32] algorithm with a time step of 2.0 fs, LINCS algorithm [36] to preserve the bond lengths, and SETTLE algorithm [37] to maintain the geometry of the water molecules. Lennard-Jones interactions were treated with a twin-range cutoff of 0.9∶1.4 nm and an integration time step of 2.0 fs; the neighbor list was updated every 10 steps. Electrostatic interactions were described using the particle-mesh Ewald method [38], with a cutoff of 0.9 nm for real-space interactions, and a 0.12-nm grid with fourth-order B-spline interpolation for reciprocal-space interactions. The starting equilibrated unliganded conformation of the B2AR within the lipid bilayer was subjected to ten independent Adiabatic Biased MD (ABMD) simulations [39], [40] to obtain transition pathways of the receptor from an inactive to an active conformation, built using the coordinates of the nanobody-stabilized crystal structure (PDB code: 3P0G). Briefly, this method biases the system towards a given value χ0 of a predefined order parameter χ(R), where R represents the coordinates of the atoms in the system. A harmonic bias acts only when the distance of χ(R) from the target χ0 is bigger than its minimum value previously reached during the simulation (i.e. if χ(R(t))- χ0 > mins<t χ(R(s))- χ0), according to the following equation:(1) The order parameter χ measures the distance from the putative activated conformation of the receptor, and is defined as the Cα root mean square deviation (RMSD) from the active conformation of the B2AR (all residues were included except the long flexible IL3). In order to obtain activated final states, the simulation was run with χ0 = 0. After carrying out 10 independent ABMD runs with an elastic constant of k = 10 kcal/(mol⋅nm2), the trajectories were pooled and clustered using an average linkage agglomerative algorithm and the same dissimilarity measure used to run ABMD. Bonded and van der Waals interactions for the ligands were assigned manually choosing the appropriate OPLS-AA atom types [30] for each atom in the molecule. Coulomb point charges were obtained according to the RESP approach [41] from quantum chemical calculations (i.e., geometry optimization using Gaussian 03 [42] and restricted Hartree-Fock calculations with the 6-31G* basis set). Ligands for which an experimental crystal structure in complex with the B2AR is available, i.e. 2RH1 for carazolol [19], 3NY8 for ICI-118,551 [21], and 3NYA for alprenolol [21], were positioned in the binding pocket accordingly. The other ligands, i.e. the full agonist epinephrine and the partial agonists dopamine and catechol, were docked into the initial inactive model of B2AR, using a standard Autodock 4.0 protocol [43], [44]. Inferences from agonist-bound crystal structures of B2AR [10], [11] and B1AR [13] were taken into account when selecting the most accurate initial binding poses of these ligands for free-energy calculations. Notably, simulations of initial conformations comprising slightly different binding poses produced similar free-energy profiles (data not shown). The free-energy profiles of liganded and unliganded systems were estimated using metadynamics [45]–[47]. Briefly, this technique enables an efficient reconstruction of the free-energy as a function of a set of k predetermined order parameters, referred to as collective variables si(R), 1≤i≤k. A history-dependent bias potential is added to the force-field driving the system dynamics so as to discourage the re-visiting of regions of the si phase space that have already been explored. Specifically, the bias potential is(2)where t′ is a multiple of a deposition time τ and the values of wt′ and σi regulate the shape and size of the Gaussian bias contributions. In the original metadynamics algorithm, wt′ = w is constant, and the free-energy profile can be estimated up to an insignificant additive time-dependent constant as W(R)  =  − limt→∞ V(R,t). Here, we used well-tempered metadynamics [24], a variant of the original metadynamics algorithm that enables assessment of simulation convergence while keeping the computational effort focused on physically relevant regions of the conformational space. In this variant of the method, the value of wt′ depends on the bias accumulated up to t′ according to the equation: (3)where ΔT is a constant with the dimension of a temperature, kB is the Boltzmann constant, and w is a constant energy representing the maximum height of the Gaussian biases. Since in the regions where the bias is higher the exponential factor reduces the rate of the bias update, the bias potential smoothly converges to a constant value in time, and the underlying free-energy can be derived by(4)where T is the temperature at which the simulation is performed. To efficiently sample the conformational space along the activation pathway, reference states from the clustered ABMD runs were selected by cutting the agglomerative tree at 30 clusters, and selecting from them n = 10 clusters homogeneously covering the pathway. The reference states Rj (1≤j≤n) were numbered assigning j = 1 to the cluster closer to the inactive state (Cα RMSD from 2RH1 ∼0.4 Å) and j = 10 to the one closer to the active state (Cα RMSD from 3P0G ∼0.3 Å). Two path collective variables describing the position along (s) and the distance from (z) the pathway were defined [48] as follows:(5)(6)where d(R,Rj) is the squared Cα RMSD (excluding IL3) with respect to the reference structure Rj, and Z  =  ∑1≤j≤n exp(− γ d(R,Rj)). The simulations were performed choosing γ  =  1/0.25 Å-2, σs = 0.1, and σz = 1 Å2, and well-tempered metadynamics was used with a bias factor ΔT = 10 T, an initial value of w = 0.4 kcal/mol, and a deposition interval τ = 8 ps. Metadynamics simulations were run for 300 ns, time at which the reconstructed free-energy difference between the metastable states converged to 0.2 kcal/mol. Since the trajectory was generated adding the metadynamics bias, the resulting conformations cannot be used to obtain statistical information on order parameters other than the collective variables. However, it is possible to unbias the distribution of any given function of the system coordinates using the algorithm described in [49]. This so-called reweighting technique was used to estimate the free-energy surface of the complexes as a function of three important descriptors of receptor activation, namely the distance between R3.50 and E6.30 (the “ionic lock”), the rotamer of residue W6.48 (the so-called “toggle switch”), and the outward displacement of the intracellular segment of TM6. Three order parameters were defined to monitor the behavior of these changes upon activation. For the ionic lock, we defined dIL =  ||〈R3.50〉 − 〈R6.30〉||, where 〈R3.50〉 and 〈R6.30〉 represent the center of mass of the η-nitrogens of R3.50 and the δ-oxygens of E6.30, respectively. For the toggle switch, we monitored the first dihedral angle χTS of the side chain of W6.48. Finally, the movement of TM6 was measured by aligning the receptor to the inactive crystal structure (2RH1) and calculating the distance dTM6  =  ||M − 〈R6.35〉|| (angled brackets indicate the centroid of all the atoms of the residues) between the midpoint of an imaginary line connecting residues K6.35 and Y2.41 in the inactive structure, M  =  ½[〈R2.41〉 + 〈R6.35〉] (located roughly at the center of the intracellular exposed surface of the receptor), and residue K6.35. The outward movement is described by the difference in dTM6 values between any given conformation and the reference inactive crystal structure, i.e. by ΔdTM6 = dTM6–dTM6 (2RH1). Representative conformational states of the metastable energy basins identified by metadynamics were selected and their structural stability analyzed. Specifically, standard, unbiased, NPT molecular dynamics simulations of these conformational states were initiated by randomizing new initial starting velocities with the Maxwell distributions at 300 K, and were run for ∼50 ns using the same simulation parameters described above. We calculated the free-energy profile of the B2AR in an explicit POPC/10% cholesterol membrane bilayer along an activation pathway connecting two recently determined inactive [19] and active [10] crystallographic states of the receptor. Specifically, the receptor was studied in its unliganded form as well as in complex with the full agonist epinephrine, the weak partial agonist dopamine, the very partial agonist catechol, the inverse agonist ICI-118-551, the inverse agonist carazolol, or the neutral antagonist alprenolol. All free-energy values at the active and inactive states, and the barriers between them, are summarized in Table 1. An activation pathway from the inactive to the active B2AR crystal structures (PDB codes 2RH1 and 3P0G, respectively) was obtained by ABMD following the protocol described in the Materials and Methods section. This pathway was used to define the s and z collective variables (see the Materials and Methods section for corresponding equations) that were employed for the metadynamics simulations. Panel A of Figure 1 illustrates the free-energy ?G of the unliganded receptor as a function of the position s along the activation pathway following integration of the dependence on z. Specifically, s = 0.0 and s = 1.0 indicate the inactive and fully activated extreme conformations of the pathway, respectively. This free-energy profile shows two minima, one at s∼0.2 that is close to the inactive state and the other at s∼0.6 that is shifted towards the active state. The two states are separated by a barrier of ∼2.5 kcal/mol, but they have a similar overall stability (ΔG<kBT), and are therefore equally populated at equilibrium. Inspection of the entire two-dimensional free-energy profile ΔG(s,z) reported in the supplementary material (see panel A of Figure S2) shows that these states correspond to conformations along the activation pathway with values of z close to 0. Visual inspection of a representative structure of the s∼0.2 energy basin confirms that the corresponding transmembrane bundle is very close to the inactive B2AR crystal structure (Cα RMSD excluding IL3 ∼0.6 Å), as substantiated by the very small outward movement of TM6 (ΔdTM6 ∼ 0.4 Å) with respect to the inactive crystal (see panel B of Figure S2). In contrast, a representative structure of the second energy basin at s∼0.6 (RMSD ∼1.6 Å and ∼1.1 Å from the inactive and active crystal structures, respectively) displays a more pronounced outward movement of TM6 (ΔdTM6 ∼2.5 Å in Figure S2). Figure 1B shows the free-energy of the unliganded B2AR as a function of order parameters that monitor molecular switches which have traditionally been reported as descriptors of GPCR activation. Specifically, these molecular switches are: 1) the ionic lock between TM3 and TM6, herein monitored using the distance dIL between R3.50 and E6.30 and 2) the W6.48 rotamer toggle switch, herein monitored using the first dihedral angle χTS of the residue side chain. Whilst the latter has not been observed in recent activated crystal structures of GPCRs, compelling spectroscopic data exist supporting a rotamer change of the W6.48 side chain upon activation [50]. Two different energy basins can be identified in the plot of Figure 1B: a more stable one, labeled a, in which both molecular switches are in their inactive conformation (dIL∼3 Å and χTS∼163°), and a second basin, labeled c, where both switches are in their activated conformation (dIL∼12 Å and χTS∼55°). The two basins are separated by a barrier of ∼3.0 kcal/mol. A transition state at χTS∼65° and dIL∼5 Å (labeled b on the free-energy map) suggests a preferential rotamer toggle switch prior disruption of the ionic lock. The neutral antagonist alprenolol, consisting of an “aromatic head” (a 2-allyl-pheniloxyl moiety) and an “aliphatic tail” (oxy-propanol-amine) (see chemical structure in Figure 2A), was docked in accordance to the binding mode assumed by the ligand in the crystal structure of the corresponding ligand-bound receptor [21]. The results of the simulations for the alprenolol-bound receptor are illustrated in panels A-C of Figure 2. As shown in Figure 2A, the overall shape of the free-energy profile of the alprenolol-bound B2AR as a function of the position (s) along the activation pathway is qualitatively similar to the profile obtained for the unliganded receptor, and reported in Figure 1A. A similarity is also noted between the two-dimensional energy surfaces of the alprenolol-bound (Figure S3A) and the unliganded B2AR (Figure S2A). In spite of these qualitative similarities, the inactive state at s∼0.2 is more stable (∼1 kcal/mol) than the intermediate state at s∼0.6 for the alprenolol-bound receptor compared to the unliganded one. Given the relatively higher stability of the alprenolol-bound receptor conformation with no significant outward movement of TM6 (ΔdTM6 ∼0.4 Å at s∼0.2 in Figure S3B), these results suggest an energy profile that is more suitable for a very weak inverse agonist rather than a neutral antagonist. Notably, data are available in the literature in support of an inverse agonist [51], [52] (or even a partial agonist [53]) role for alprenolol. Figure 2B shows a representative conformation of the lowest energy basin identified for the alprenolol-bound receptor. In this conformation, and similar to the corresponding crystal structure [21], the alprenolol charged moiety in its aliphatic tail forms interactions with polar residues D3.32 and N7.39, while the ligand aromatic head interacts with residues V3.33, V3.36, F6.51, N6.55, Y5.38, and S5.42, which define a cleft formed by TM3, TM5 and TM6. Figure 2C shows that the energetically most stable alprenolol-bound inactive state is characterized by inactive molecular switches (χTS∼160° and dIL∼5 Å). This state, labeled a in Figure 2C, is separated by an energy barrier of ∼3 kcal/mol from the second most stable energetic minimum at χTS∼50° and dIL∼12 Å (c in Figure 2C), with a transition state (b in Figure 2C) at χTS∼85° and dIL∼5 Å. Thus, the presence of alprenolol in the binding pocket does not appear to disrupt the free-energy profile seen in the unliganded receptor, further confirming a possible rotamer toggle switch of the W6.48 residue prior breaking of the ionic lock. The stability of alprenolol in a representative state of the ligand-receptor complex extracted from the most stable energy basin at s∼0.2 was confirmed by carrying out ∼50 ns unbiased MD simulations. The evolution of the ligand and the protein RMSD during these simulations is reported in Figure S4. We assessed the effect of two different B2AR inverse agonists, namely ICI-118,551 and carazolol, on the free-energy landscape of the receptor during transition from inactive to activated experimental states. Carazolol and ICI-118,551 share important structural features with alprenolol, e.g., they both have an “aliphatic tail” (oxy-propanol-amine for carazolol and oxy-butanol-amine for ICI-118,551) and an “aromatic head”. The results of the simulations for the carazolol-bound and the ICI-118,551-bound receptor are illustrated in panels A-C and D-F of Figure 3, respectively. In the presence of either carazolol or ICI-118,551, the B2AR free-energy profiles (Figure 3A and 3D, respectively) show a single lowest-energy basin at s∼0.18 close to the inactive state of the receptor. These much more stable energy basins are also present in the two-dimensional energy surfaces of the carazolol-bound (Figure S5A) and the ICI-118,551-bound (Figure S5C) B2AR, and comprise inactive conformations as further illustrated by the lower energy values for states characterized by the absence of outward movement of TM6 (ΔdTM6 ∼ 0.4 Å in Figures S5B and S5D). Representative conformations extracted from the lowest energy basins of either the carazolol-bound (Figure 3B) or the ICI-118,551-bound (Figure 3E) receptors show that the energy-optimized binding poses of these ligands are very similar to their positions in the corresponding crystal structures [19], [21]. Similar to the binding mode of alprenolol, the charged moieties contained in the aliphatic tails of these ligands interact with polar residues D3.32, and N7.39, while their aromatic heads are oriented toward TM3, TM5, and TM6, thus directly interacting with residues in these helices (see Figures 3B and 3E). To assess the stability of the ligands in these representative conformations, we performed standard, unbiased MD simulations. As shown in Figures S6A-D, which report the time evolutions of the RMSD of the protein, as well as those of the heavy atoms of carazolol and ICI-118,551, after superposition of the receptor Cα atoms, the receptor conformations are stable and the binding modes of the ligands are conserved over a simulation time of ∼50 ns. The intermediate state at s∼0.6 that was significantly populated in the unliganded and neutral antagonist-bound receptor is much less stable at ΔG∼4.0 kcal/mol in the case of the carazolol-bound or ICI-118,551-bound receptors (see Figures 3A and 3D, respectively). However, these are still metastable states, as judged by the presence of shallow minima at s∼0.6 in both the free-energy profiles, and are separated from the inactive states by multiple barriers. In terms of modulation of the toggle switch and the ionic lock, the free-energy as a function of χTS and dIL (Figures 3C and 3F for the carazolol-bound and ICI-118,551-bound complexes, respectively) features only one minimum in the inactive region of these molecular switches (χTS∼160° and dIL∼3 Å). To study the effects of full agonists on the free-energy landscape of B2AR, we docked epinephrine into the receptor, and performed metadynamics calculations. Figure 4A shows the free-energy profile of the epinephrine-bound B2AR with the lowest energy state (s∼0.9) likely to correspond to an activated conformation. The same observation is possible by inspection of the two-dimensional free-energy surface (Figure S7A) as well as the TM6 outward movement (ΔdTM6 ∼5.9 Å in Figure S7B) as a function of the position s along the activation pathway. However, a second low-energy metastable state is present in these free-energy profiles, close to the inactive state (s∼0.2), and with a free-energy difference of only ∼1 kcal/mol with respect to the most stable activated state. As illustrated in Figure 4B, our proposed binding mode of epinephrine within a fully activated B2AR (energy basin at s∼0.9) is consistent with the binding poses displayed by full agonists in the B2AR [10] and B1AR [13] crystallographic structures. Specifically, the ligand amino group forms hydrogen bonds with D3.32 and N7.39 of B2AR, the ligand β-hydroxyl group interacts with D3.32, and the ligand catecholamine hydroxyl groups interact through hydrogen bonding with the side chains of both S5.42 and S5.46. In this state, the B2AR helix bundle is structurally very similar to the corresponding nanobody-activated crystal structure of the receptor (C? RMSD from 3P0G is ∼1.6 Å). The stability of the epinephrine binding pose and the specific receptor conformation were verified by carrying out ∼50 ns standard MD simulations (see corresponding time evolutions of RMSD in Figure S8). On the other hand, representative structures of the energy basin at s∼0.2 (data not shown) corresponded to conformations of the helix bundle very similar to the inactive crystal structure of B2AR (Cα RMSD from 2RH1 is ∼1.0 Å). Two energy basins (labeled a and c) were identified from the free-energy as a function of the order parameters describing the ionic lock and rotamer toggle switches (Figure 4C). Specifically, the basin comprising conformations in which both the ionic lock and rotamer toggle switches are in the ‘active’ (dIL∼16 Å and χTS∼50°) positions appear to be more stable than the basin with receptor conformations with ‘inactive’ (dIL∼3 Å and χTS∼160°) molecular switches. Also in this case, the minimum free-energy path between these two energy basins suggests activation of the toggle switch prior breaking of the ionic lock interaction along the path to full receptor activation. The weak and very weak partial agonists, dopamine and catechol, were also simulated in the context of the B2AR activation pathway. Figures 4D and 4G illustrate the free-energy profiles of the catechol-bound and dopamine-bound receptors, respectively. In both cases the receptor is most stabilized in an intermediate state (s∼0.6) along the pathway to activation. Inspection of the free-energy as a function of the position (s) along and the distance (z) from the activation pathway (see Figures S9A and S9C for the catechol-bound and dopamine-bound receptors, respectively) confirms that these two ligands stabilize a state different from the inactive or fully activated ones as judged by the lowest energy values at z∼2 in Figure S9A for catechol, and at s∼0.6, z∼0.0 in Figure S9C for dopamine. This difference is also evident from the free-energy surfaces as a function of the TM6 outward movement and the position along the activation pathway (see Figures S9B and S9D, respectively), as well as from the structural superpositions shown in Figure 5. Specifically, Figure 5 illustrates the structural differences between the TM regions of the predicted inverse agonist- and partial agonist-specific conformations (Figure 5A), the inverse agonist- and full agonist-specific conformations (Figure 5B), and the partial agonist- and full agonist-specific conformations (Figure 5C). Figures 4E and 4H show the binding modes of catechol and dopamine, respectively. These binding poses were proven to be stable during ∼50 ns of unconstrained MD simulations (see Figures S10A and S10B for the time evolution of the RMSD of catechol and dopamine, respectively, and Figures S10C and S10D for the time evolution of the RMSD of the corresponding protein Cα atoms). In agreement with inferences from recent B1AR structures co-crystallized with either full or partial agonists, these two B2AR partial agonists formed stable hydrogen bonds (through the catechol moiety) with the side chain of S5.42, but do not with S5.46. In terms of the ligand-induced modulation of the molecular switches, the catechol-bound B2AR state with a broken ionic lock (located at χTS∼50° and dIL∼16 Å in Figure 4F) is relatively less stable than the corresponding larger basin identified in the presence of dopamine (see Figure 4I), consistent with spectroscopy data suggesting that catechol is unable to disrupt the ionic lock [9]. Understanding the molecular mechanisms underlying GPCR functional selectivity is extremely important in modern drug discovery, since it provides a unique opportunity for the identification or rational design of ‘biased’ ligands as novel more effective therapeutics. Epitomizing an emerging paradigm in current drug discovery [54], native states of GPCRs can be assumed in a dynamic equilibrium between different conformational sub-states [11], [18], [55], which correspond to the valleys of an energy landscape, the barriers of which reflect the timescales of the conformational exchange. The relative populations of these sub-states follow statistical thermodynamics distributions and are shifted towards specific conformations as a consequence of ligand binding and/or other allosteric events such as those induced by protein-protein interactions. Thus, ligands with varied efficacies are believed to modulate the free-energy landscape of a GPCR, shifting the conformational equilibrium towards active or inactive conformations of the receptor, depending on their pharmacological action. A reliable characterization of the specific conformations that inverse agonists, agonists (both full and partial), or antagonists can stabilize in a given GPCR is highly desirable for the structure-based discovery of novel ligands eliciting selected functional responses. This is difficult to achieve by X-ray crystallography for the majority of GPCRs due to their intrinsic structural instability, and the realization that the majority of pharmacologically active ligands are not ideal compounds for receptor stabilization that is suitable for crystallization. The enhanced sampling approach we describe here provides atomic-resolution information of receptor conformations along pre-determined activation pathways that are differentially stabilized by ligands with different efficacies. Our approach also provides a quantitative description of the thermodynamics of the B2AR basal activity, with the unliganded receptor being able to sample both an inactive state and an intermediate state that is shifted towards the activated conformation. This latter state is structurally different from the fully active state of B2AR captured by the nanobody-stabilized crystal structure. Although it exhibits a broken ionic lock and a cytoplasmic opening that is able to accommodate the camelid antibody, a few clashes are produced by the much smaller outward movement of TM6 (∼2.5 Å compared to the ∼5.9 Å that can be achieved by a full agonist). Given the small free-energy difference between the two lowest energy minima identified for the unliganded B2AR, these two states are almost equally populated at equilibrium, in agreement with the high basal activity of the B2AR. Moreover, the relatively low energy barrier between the two states is consistent with the flexible nature of the unliganded B2AR, and the consequent difficulty in obtaining crystals of the native receptor. We observed a more or less pronounced perturbation of the free-energy profile of the unliganded B2AR depending on the ligand considered for binding. Although alprenolol has often been described as a neutral antagonist of B2AR, its presence in the B2AR binding pocket slightly modifies the free-energy profile of the receptor, making the inactive state more stable in spite of the small difference in free-energy (∼kBT). This result is not completely surprising in light of the evidence existing in the literature for a role of alprenolol as an inverse agonist or even a weak agonist, depending on the assay used 56,57. Our results show that the selection of a single conformational state is particularly effective in the case of inverse agonists. The docking of either carazolol or ICI-118,551 in the receptor dramatically changes the free-energy landscape of B2AR and reduces it to a funneled profile with a single major basin corresponding to the inactive conformation. This result is consistent with the greater availability of crystals of B2AR in an inactive conformation stabilized by potent inverse agonists in the binding pocket, and with the observation that the structural features of the inactive states of the various receptors obtained so far are similar. The situation is different when we study the free-energy landscape in the presence of agonists. The computational experiment with epinephrine shows that a full agonist is capable of stabilizing a state of B2AR presenting structural features that have been found in the nanobody-stabilized agonist-bound crystal structure of B2AR. However, in addition to this active state, we obtain a relatively stable agonist-bound inactive state that is structurally similar to the inverse agonist-bound crystal structure of B2AR. This is not surprising, given the absence of TM6 outward movements noted in both the B2AR crystal structure with a covalently-bound agonist [11], and the agonist-bound B1AR crystal structures [13]. Moreover, the relatively small difference in free-energy between the fully active and the inactive agonist-bound conformations is probably due to the lack of the G-protein in the simulation setup, in line with the observation deriving from the two recent agonist-bound B2AR crystal structures [10], [11] that a ligand alone is not sufficient to stabilize a fully active crystallographic state of the receptor, but a G-protein mimicking nanobody is necessary to trap this conformation. Different from the crystallographic information, but in line with experimental evidence from fluorescence spectroscopy [9], we find that metastable states corresponding to fully (and partial) activated conformations of the receptor favor the rotamer change of the W6.48 side chain. The partial agonism elicited by dopamine and catechol shifts the conformational equilibrium towards states that are different from that stabilized by the full agonist, and captured in the nanobody-stabilized crystal structure. In particular, the two ligands affect the free-energy landscape in different ways. While the intermediate dopamine-bound state always features a broken ionic lock, the receptor samples conformations that have different ionic lock states when catechol is in the binding pocket. Notably, experimental evidence from fluorescence spectroscopy [9] also suggested that the very weak partial agonist catechol is not able to completely disrupt the interaction between the charged residues at the cytoplasmic end of TM3 and TM6. Structurally, the two conformations stabilized by catechol and dopamine are different in the degree of separation between the extracellular ends of TM5 and TM6 and between the intercellular ends of TM3 and TM6. Consistent with the hypothesis that global structural features of the receptor, such as the tilt of the extracellular half of TM5, can optimize the binding to agonists [58], we see a larger TM5 tilt in the presence of dopamine (as well as for epinephrine) and a smaller one in the presence of catechol. Owing to the greater ability of catechol to stabilize a state with a formed ionic lock, the intracellular ends of TM3 and TM6 also appear slightly closer (by ∼1 Å) together. In summary, we have designed a strategy using a combination of different adaptive biasing techniques that enables characterization of reliable ligand-specific conformations as demonstrated here in the case of B2AR. The strategy is completely general and may be of practical use for the structure-based design of ‘biased’ ligands that selectively activate signaling pathways, and may therefore exhibit improved therapeutic properties.

10.1371/journal.pgen.1003582

Joint Molecule Resolution Requires the Redundant Activities of MUS-81 and XPF-1 during Caenorhabditis elegans Meiosis

The generation and resolution of joint molecule recombination intermediates is required to ensure bipolar chromosome segregation during meiosis. During wild type meiosis in Caenorhabditis elegans, SPO-11-generated double stranded breaks are resolved to generate a single crossover per bivalent and the remaining recombination intermediates are resolved as noncrossovers. We discovered that early recombination intermediates are limited by the C. elegans BLM ortholog, HIM-6, and in the absence of HIM-6 by the structure specific endonuclease MUS-81. In the absence of both MUS-81 and HIM-6, recombination intermediates persist, leading to chromosome breakage at diakinesis and inviable embryos. MUS-81 has an additional role in resolving late recombination intermediates in C. elegans. mus-81 mutants exhibited reduced crossover recombination frequencies suggesting that MUS-81 is required to generate a subset of meiotic crossovers. Similarly, the Mus81-related endonuclease XPF-1 is also required for a subset of meiotic crossovers. Although C. elegans gen-1 mutants have no detectable meiotic defect either alone or in combination with him-6, mus-81 or xpf-1 mutations, mus-81;xpf-1 double mutants are synthetic lethal. While mus-81;xpf-1 double mutants are proficient for the processing of early recombination intermediates, they exhibit defects in the post-pachytene chromosome reorganization and the asymmetric disassembly of the synaptonemal complex, presumably triggered by crossovers or crossover precursors. Consistent with a defect in resolving late recombination intermediates, mus-81; xpf-1 diakinetic bivalents are aberrant with fine DNA bridges visible between two distinct DAPI staining bodies. We were able to suppress the aberrant bivalent phenotype by microinjection of activated human GEN1 protein, which can cleave Holliday junctions, suggesting that the DNA bridges in mus-81; xpf-1 diakinetic oocytes are unresolved Holliday junctions. We propose that the MUS-81 and XPF-1 endonucleases act redundantly to process late recombination intermediates to form crossovers during C. elegans meiosis.

Meiotic recombination generates joint molecules that ensure chromosomes segregate correctly. Failure to generate or resolve joint molecules can have profound effects on fertility and on the viability of resulting progeny. The generation and resolution of joint molecules is carefully regulated. Generation of joint molecules is highly similar across a broad range of organisms, from yeast to mammals. Yet, the resolution of the resultant joint molecules varies across organisms, with helicases and endonucleases contributing to varying extents in different organisms. We used the genetically tractable model organism, Caenorhabditis elegans (C. elegans) to uncover redundancies between joint molecule processing proteins. Specifically, we investigated the contribution of the C. elegans BLM helicase ortholog, HIM-6, and the endonucleases MUS-81, XPF-1, GEN-1 and EXO-1 to the resolution of meiotic joint molecules. We found that MUS-81 and HIM-6 act redundantly to resolve joint molecules early in meiosis, presumably to form noncrossovers. Late in meiosis, MUS-81 and XPF-1 act redundantly to resolve joint molecules to form crossovers. When both MUS-81 and XPF-1 are absent, joint molecules are not resolved, resulting in disorganized chromosomes in the oocyte and embryonic death. Joint molecules in mus-81;xpf-1 animals are rescued by microinjection of the human GEN1 protein, indicating these intermediates are Holliday junctions.

Meiotic recombination generates chiasmata that join homologous chromosomes together to ensure proper meiotic chromosome segregation. The efficient generation and resolution of joint molecules (JM) is essential for meiosis; therefore, JM formation and resolution is carefully regulated. In most organisms, meiotic recombination is initiated by the generation of Spo11-induced double strand breaks (DSBs). DSBs are resected to produce a 3′ single-stranded stretch of DNA onto which Rad51 is loaded, forming a nucleoprotein filament. Rad51 catalyzes invasion of the homologous chromosome and JM intermediates physically linking homologous chromosomes are formed (reviewed in [1]). JMs must be resolved before homologs segregate at meiosis I. JMs can be resolved to form crossover (CO) products in which flanking markers are exchanged, or they can be resolved to form non-crossover (NCO) products. The overall progression of JM resolution appears to be similar in diverse organisms, however the proteins and their relative involvement in JM resolution vary from species to species. Thus, the same initiating lesion (Spo11-induced DSB) is repaired through diverse mechanisms. Study of meiotic DSB repair in a range of organisms illuminates the modularity of repair and how different organisms have evolved to favor distinct endonucleases to repair Spo11 generated DSBs. There are two main pathways that process meiotic JMs to COs, which are utilized to varying extents in different organisms. Much of what we know about meiotic crossover resolution comes from studies in budding yeast and fission yeast. The predominant pathway in budding yeast involves the synaptonemal complex-associated ZMM (Zip1/2/3/4, Msh4/5, Mer3) proteins. It has been proposed that the ZMM proteins protect recombination intermediates (RIs) from NCO resolution, ensuring CO resolution of ZMM-associated RIs [2], [3]. ZMM-dependent COs are subject to crossover interference, which occurs when the presence of one CO decreases the probability that another CO will occur nearby. A second ZMM-independent pathway is characterized by the structure specific endonuclease Mus81/Mms4 that resolves RIs to either CO or NCO outcomes [4], [5]. These ZMM-independent COs are not subject to crossover interference. Fission yeast, which lack both a synaptonemal complex and ZMM proteins, represent an extreme case in which all COs are ZMM-independent and are resolved by Mus81/Eme1. Loss of Mus81 in S. pombe results in spore inviability due to meiotic chromosome segregation defects and a profound decrease in the frequency of COs [6], [7]. Consistent with Mus81 being the major meiotic Holliday junction (HJ) resolvase in fission yeast, the meiotic chromosome segregation defects of Mus81 mutants can be rescued by the expression of the bacterial HJ resolvase RusA [7] or by the expression of the human Holliday junction resolvase GEN1 [8]. Loss of Mus81 in budding yeast, which has a synaptonemal complex and ZMM-proteins, results in only a minor reduction in spore viability and a modest decrease in the frequency of COs [9]. These data are consistent with a model in which Mus81 is only responsible for ZMM-independent COs and that ZMM-dependent COs predominate in budding yeast [5]. This bias towards the ZMM-dependent CO pathway in budding yeast is mediated by the BLM-helicase homolog, Sgs1 [10]–[13]. Sgs1 is thought to prevent the accumulation of JMs by channeling most double strand breaks towards NCO resolution and ensuring that remaining DSBs are associated with ZMM-proteins and resolve as COs [10], [13]. When Sgs1 is absent, JMs accumulate and require Mus81 for resolution. Cells lacking both Sgs1 and Mus81 during meiosis accumulate JMs and cannot segregate at meiosis I [11], [12]. Although, loss of Mus81 does not greatly affect meiotic segregation in budding yeast, it appears that the Yen1 and Slx1/Slx4 endonucleases act redundantly with Mus81 to resolve persistent JMs. Loss of Yen1 alone does not result in meiotic phenotypes whereas mus81 yen1 double mutants exhibit a profound decrease in spore viability due to a failure of JM resolution and chromosome segregation at meiosis I [14]. Similarly, the Slx4/Slx1 endonuclease processes some RIs in the absence of Mus81 [10], [13]. The structure-specific endonucleases are also critical for crossing over in more complex organisms. In Drosophila, MUS81 does not appear to play a role in the formation of COs [15]. Most COs are catalyzed by the Mus81-related structure specific endonuclease MEI-9 (fly nucleotide excision repair endonuclease XPF ortholog), and MUS312 (fly SLX4 ortholog). Loss of function of either MEI-9 or MUS312 results in a severe decrease in crossing-over [16] suggesting that MUS312 and MEI-9 resolve meiotic HJ intermediates. The MUS81 and XPF endonucleases also play a role in mouse meiosis. Although Mus81 knockout mice are viable and fertile [17], [18], MUS81 appears to be required for the repair of at least some meiotic DSBs during murine meiosis. Mus81-/- male mice exhibit significant meiotic defects in the germ line: mature sperm are depleted, MLH1 foci, which mark ZMM-dependent crossovers, are increased, and a subset of meiotic DSBs is not repaired [19]. Similar to Mus81 mutant mice, mice lacking ERCC1 (the binding partner of XPF) or BTBD12 (Slx4 ortholog) exhibit sperm defects, persistent unrepaired DSBs in the germ line, and increased MLH1 foci [20], [21]. C. elegans is a powerful model for study of metazoan meiotic DSB repair. The germ line is temporally and spatially polarized with respect to meiotic progression, allowing detection of subtle but significant alterations in the kinetics of repair events. Synaptonemal complex formation and DSB induction are independent during C. elegans meiosis, allowing separation of homolog pairing and DSB repair. Finally, crossover interference is incredibly robust in C. elegans with only a single CO occurring per pair of homologous chromosomes. In C. elegans, it has been proposed that all COs are ZMM-dependent [22] and as such would not require MUS-81 or related structure-specific endonucleases for resolution. However, the study of rtel-1 anti-recombinase mutants revealed that there are two classes of COs in C. elegans: ZMM-dependent COs; and ZMM-independent COs, which require MUS-81 for resolution [23]. Other structure-specific endonucleases also play a role in CO formation in C. elegans; loss of XPF-1 (XPF/MEI-9 ortholog) results in a decrease in the number of COs compared to wild type animals. Furthermore, loss of both MUS-81 and HIM-18, the C. elegans Slx4 ortholog, results in an increase in mitotic and meiotic RIs [24]. Together, these data suggest that MUS-81 and HIM-18 play overlapping, non-redundant roles in processing RIs. Unlike MUS-81, XPF-1, or HIM-18, and in contrast to yeast, the related endonuclease GEN-1 (Yen1 ortholog) does not appear to have a role during C. elegans meiosis [25]. What remains unclear is the relative contribution of each of the structure-specific endonucleases to CO formation and where in the CO pathway each endonuclease functions. Are the endonucleases acting early to process RIs to NCO outcomes or are they acting late to resolve RIs as COs? Here, we investigated the relative roles of structure-specific endonucleases in processing RIs during meiosis in C. elegans with the goal of identifying the enzyme(s) responsible for processing JMs to produce COs. Unexpectedly, we found that MUS-81 performs both early and late roles in processing RIs during meiosis I in C. elegans and that its loss results in an overall decrease in the total number of COs. We show that MUS-81 and HIM-6 (Sgs1 homolog) act early during pachytene to limit the accumulation and persistence of RAD-51-associated RIs during meiosis, a defect that also manifests as chromosome fragmentation at diakinesis. Surprisingly, we found that MUS-81 and XPF-1 endonucleases, but not GEN-1 or EXO-1, act redundantly to process late stage JMs to form COs. Loss of both MUS-81 and XPF-1 resulted in defective CO maturation and persistent JMs at diakinesis, which could be rescued by germline injection of the human Holliday junction resolvase GEN1. As human GEN1 is able to cleave HJs in vitro, these results strongly suggest that the persistent JMs in mus-81; xpf-1 double mutants are HJs. Together, these data support a redundant role for MUS-81 and XPF-1 in processing HJ intermediates to produce interhomolog crossovers in C. elegans. To examine the function of MUS-81 during meiosis in C. elegans, we characterized the meiotic phenotype of a mus-81 null mutant [26]. C. elegans mus-81(tm1937) mutants exhibited no obvious phenotypes attributed to meiotic defects such as a high frequency of embryonic inviability or an increased frequency of XO males. mus-81 mutants had reduced brood sizes (142+/−19, approximately 50% of wild type, Student's t-test p<0.005) (Table 1). mus-81 brood sizes were variable, ranging from rare animals that were completely sterile (1/20 broods scored) to animals that produced more than 200 progeny (5/20 broods scored). The brood size in C. elegans is dictated by the number of viable sperm, so defects in the germ line that lead to reduction in the number of viable sperm could result in a smaller brood size. A checkpoint in the C. elegans female germ line senses DNA damage or persistent recombination intermediates, triggering apoptosis of the damaged nuclei [27], [28]. To determine if this apoptotic checkpoint protects the mus-81 germ line by removing defective nuclei before they develop into oocytes, we constructed mus-81(tm1937); ced-4(n1162) double mutants. ced-4 is essential for the initiation of apoptosis in C. elegans [29]. mus-81; ced-4 double mutants had a significantly reduced brood when compared to either ced-4 or mus-81 single mutants (Table 1, Student's t-Test p<0.01) and 4/10 lines were completely sterile suggesting that the apoptotic checkpoint removes nuclei with DNA damage or aberrant meiotic recombination intermediates in the mus-81 mutant. This observation together with the reduced brood suggested that there are defects in the mus-81 germ line. We next assayed the distribution of early RIs in the C. elegans germ line using an antibody that recognizes the recombination protein RAD-51. In wild type animals, SPO-11-dependent RAD-51 foci are visible in early pachytene and are resolved by late pachytene (Figure 1; Figure S1; Figure S2). In mus-81 animals the average number of RAD-51 foci was slightly but significantly increased in all zones except diplotene when compared to wild type animals (Student's t-test p<0.05). Apart from a small but significant increase of RAD-51 foci in the mitotic zone (average 0.54 mus-81, 0.05 WT) and the persistence of foci in late pachytene, the same general pattern of RAD-51 staining was observed in mus-81(tm1937) and wild type backgrounds (Figure 1A, B). Consistent with this observation, mus-81 mutants did not exhibit high levels of chromosome non-disjunction or fragmentation in diakinetic oocytes (Figure 1C, D). In yeast, loss of Mus81 and the helicase Sgs1 confers synthetic lethality as a result of the accumulation of unresolved RIs [7], [9], [30], [31]. Meiosis-specific mutant alleles of mus81 and sgs1 indicate that Sgs1 limits JMs and that, in the absence of Sgs1, Mus81 resolves these structures to prevent the accumulation or persistence of JMs during meiosis [11], [12]. To test whether a similar relationship exists in C. elegans, we characterized the formation of RIs in animals lacking MUS-81 and the C. elegans Sgs1 helicase ortholog HIM-6 [32]. Similar to yeast mus81 sgs1 double mutants, C. elegans mus-81; him-6 double mutants exhibited a severely reduced brood size and ∼100% embryonic lethality (Table 1). DAPI staining revealed that germline nuclei progression in mus-81; him-6 double mutants was grossly normal with nuclei progressing from the mitotic zone to pachytene and finally diakinesis (Figure S1, Figure S2). To further examine this phenotype we stained germ lines with an anti-RAD-51 antibody to monitor the distribution of early meiotic RIs. Both mus-81 and him-6 single mutants exhibited increased RAD-51-associated RIs in mid-pachytene compared to wild type animals (5.94 and 7.92 respectively; WT 2.96, Student's t-Test p<0.01). However, the mus-81; him-6 double mutant accumulated significantly more RIs in mid-pachytene (21.1 per nucleus, Student's t-test p<0.01 compared to either single mutant) than would be predicted for the additive effect of the two mutations (Figure 1A, B). Consistent with the generation of large numbers of persistent RIs, mus-81; him-6 double mutant animals exhibited evidence of chromosome breakage with significantly more DAPI staining bodies in diakinetic oocytes than would be predicted for the additive effect of the two mutants (Figure 1C, D). Approximately 20% of oocytes contained more than 12 DAPI staining bodies and many of the DAPI-stained bodies were very small, suggesting that at least some of these DAPI staining bodies represented chromosome fragments rather than loss of chiasmata between homologs, which would produce 12 univalents. The increased RAD-51 foci in the mus-81; him-6 double mutant coincided with the meiotic zones in which SPO-11 is active and Agostinho et al. [33] demonstrated that the chromosome fragmentation phenotype of mus-81; him-6 is SPO-11-dependent, Therefore, these RAD-51 foci likely result from SPO-11 generated DSBs, though it is formally possible that these foci resulted from DNA damage arising in the transition zone and early pachytene. Collectively, these data support roles for HIM-6 and MUS-81 in limiting early RIs during meiosis. The increased persistent RIs in mus-81; him-6 double mutants raised the possibility that MUS-81 resolves RIs in the C. elegans germ line. Previously, we demonstrated that MUS-81 is required in rtel-1 mutants to resolve supernumerary meiotic RIs to produce COs [23]. In the absence of MUS-81 in rtel-1 mutants, large numbers of RAD-51 foci persist into late pachytene [23], which is similar to what we observed in the mus-81; him-6 double mutant. To determine if MUS-81 played a role in the formation of endogenous COs, we used visible genetic markers to measure recombination frequency in two genetic intervals, unc-45 dpy-17 and dpy-17 unc-64, that span 48.8 map units of chromosome III (approximately 98% of the genetic length). Unexpectedly, recombinant progeny were reduced in both intervals in mus-81 mutants compared to mus-81/+ heterozygotes (Figure 2A), demonstrating that MUS-81 promoted meiotic CO generation. The apoptotic checkpoint removes nuclei with DNA damage or meiotic defects and is only active in the female germ line [27], [28], allowing us to test whether this checkpoint was ameliorating the mus-81 mutant phenotypes. We therefore measured recombination frequencies between dpy-17 and unc-64 in mus-81 males. mus-81 mutant males exhibited a much greater effect on recombination distances than that observed for hermaphrodites (WT males 29.2 [95% CI 24.2–34.5], mus-81 males 14.1 [95% CI 11.6–17.3]). The activity of the apoptotic checkpoint in the female germ line could explain the differences in CO frequency observed in our data compared to those reported by Saito et al. [34] and Agostinho et al. [33], who used a method that only measured COs in oocytes. The recombination frequency phenotype of mus-81 mutants was surprising as COs in C. elegans require MSH-4/MSH-5 (ZMM-dependent) and are subject to strong crossover interference (Class I crossovers) and therefore would not be expected to require MUS-81 for resolution. Moreover, mus-81 mutants exhibited a near wild type number of ZHP-3 foci, which mark emerging COs [35] and suggested that MUS-81 acts after ZHP-3 foci formation (Figure 2B, Figure S3). Several recent studies have described redundancy between Mus81 and other related structure specific endonucleases in the resolution of meiotic and mitotic JMs in yeast [14], [36]–[38]. To test whether MUS-81 acted redundantly with other structure-specific nucleases in C. elegans, we constructed double mutants with mus-81 and the C. elegans orthologs of the endonucleases, gen-1 and xpf-1, and the endo/exonuclease exo-1. Animals lacking GEN-1 or EXO-1 did not show statistically significant differences in brood size, arrested embryos, or frequency of males compared to wild type (Figure 3, Table 1). In contrast to the gen-1 and exo-1 single mutants, xpf-1 mutants produced 9% arrested embryos and 2% males (Student's t-test p<0.005), most likely as a result of general chromosome non-disjunction [39]. Loss of gen-1 or exo-1 enhanced the phenotype of the mus-81 mutant, increasing the frequency of arrested embryos from ∼8% in mus-81 to 18% in mus-81; gen-1 and 39% in mus-81; exo-1 (Table 1). In contrast, loss of xpf-1 in the mus-81 mutant resulted in a dramatic increase in the frequency of inviable embryos to levels far greater than would be expected for the additive effects of the two mutations (74% vs 16% expected for additive effects of the mus-81 and xpf-1). Furthermore, the mus-81; xpf-1 double mutant could not be maintained as a homozygous strain. The effect on viability was increasingly more pronounced in the mus-81; xpf-1 F2 and F3 generations with the brood size decreasing significantly and ∼95% (F2) and ∼99% (F3) of the embryos arresting (Figure 3). In addition, the frequency of sterile animals or animals producing 100% inviable embryos increased from 2 of 23 lines in the F1 generation to 5 of 10 lines in the F2 and 9 of 10 lines in the F3. These data suggested that MUS-81 and XPF-1 have redundant roles in maintaining a functional germ line whereas GEN-1 and EXO-1 did not, either alone or in combination with loss of MUS-81. The inviability of mus-81; xpf-1 strains precluded measuring CO frequencies with visible markers in the double mutant so we opted to measure COs in animals homozygous for xpf-1 and heterozygous for mus-81. Strikingly, mus-81/+; xpf-1 animals exhibited a significant decrease in COs compared to mus-81/+ animals suggesting that MUS-81 and XPF-1 function redundantly to promote COs (Figure 2). This result is consistent with the decrease in COs observed in mus-81; xpf-1 double mutants by Saito et al. [34] and Agostinho et al. [33]. Although the number of COs was reduced in mus-81/+; xpf-1 animals, mus-81; xpf-1 mutants did not show a significant difference in the number of ZHP-3 foci in meiotic nuclei (Figure 2B; Figure S3) suggesting that MUS-81 and XPF-1 act downstream of ZHP-3 in meiotic progression. To further investigate the phenotypic consequence of losing both MUS-81 and XPF-1, we stained double mutant germ lines with an anti-RAD-51 antibody to observe the progression of early meiotic RIs. We observed that the appearance and subsequent disappearance of RIs was not as profoundly affected in mus-81; xpf-1 mutant animals compared to mus-81; him-6 double mutants (Figure 4A,B). RAD-51 foci were only slightly but significantly increased in the mid and late pachytene zones of the mus-81; xpf-1 double mutant compared to either single mutant (Student's t-test p<0.05). Consistent with these observations, there was no measurable increase in DAPI-stained bodies in the mus-81; xpf-1 double mutant beyond what would be expected from additive effects of the two mutants (Figure 4C,D). Although mus-81; him-6 and mus-81; xpf-1 double mutants produced severely reduced broods and increased embryonic arrest, the mechanisms underlying their respective phenotypes appear to be distinct. The lack of obvious early RI defects to account for the inviability of the mus-81; xpf-1 animals and the decrease in COs observed in mus-81/+; xpf-1 mutants raised the possibility that MUS-81 and XPF-1 act redundantly to resolve late JMs to produce meiotic COs. To test this hypothesis, we first examined whether loss of MUS-81 and XPF-1 affected the kinetics of CO resolution by monitoring the assembly/disassembly of the synaptonemal complex (SC) component SYP-1. Previous studies in C. elegans have shown that either COs or CO precursors trigger the asymmetric dissolution of the SC [40]. SYP-1 is first disassembled in the region between the CO (or CO precursor) and the most distant telomere of bivalent chromosomes in diplotene nuclei. This creates an asymmetry with a long arm of the bivalent that lacks SYP-1 and a short arm that contains SYP-1. Later in diakinesis, the remaining SYP-1 dissociates from the short arms and the Aurora-like kinase AIR-2 becomes concentrated on the short arms. Thus, the maturation of COs during diakinesis can be followed by observing the asymmetric diassembly of the SC and the appearance of AIR-2 on diakinesis oocytes. SYP-1 staining was normal in pachytene and diplotene of all single and double mutants consistent with proper chromosome synapsis and formation of early RIs (Figure 5A). In contrast, mus-81 and mus-81; xpf-1 animals exhibited defects later in meiotic progression with the timely disassembly of SYP-1 in diakinetic oocytes. In wild type animals the SC was disassembled in diakinetic oocytes with no visible SYP-1 remaining in the most proximal oocyte (diakinesis −1). In mus-81 animals, SYP-1 staining persisted in late diakinesis with 20% of −2 oocytes and ∼5% of −1 oocytes containing visible SYP-1 staining. This phenotype was exacerbated in the mus-81; xpf-1 double mutant with ∼100% of −2 oocytes and 25% of −1 oocytes containing SYP-1 staining (Figure 5B). In addition to the defects in SYP-1 disassembly, bivalents in mus-81; xpf-1 double mutants also exhibited a highly unusual morphology with DNA bridges present between the two DAPI-staining bodies in each bivalent (Figure 6B). We observed 12 univalents in mus-81; xpf-1; spo-11 animals, indicating that the DNA bridges in mus-81; xpf-1 animals were dependent on meiotic DSBs (Figure S4). In wild type animals, the axial element protein HTP-3 forms a cruciform pattern between sister chromatids along both the short and long arms of the diakinesis bivalent [41] (Figure 6D). In mus-81; xpf-1 mutants HTP-3 localization on most bivalents was highly disorganized suggesting that their structure was disrupted (Figure 6D). Further evidence for the disruption of bivalents in mus-81; xpf-1 mutants came from AIR-2 staining. In wild type animals, AIR-2 was localized between the short arms of the sister chromatids in the bivalent, appearing as a single distinct line, whereas in mus-81; xpf-1, AIR-2 appeared in two distinct spots on either side of the DNA bridge spanning the two DAPI bodies (Figure 6C). Three lines of evidence support the hypothesis that unresolved late RIs, possibly HJs, were responsible for disrupting bivalent maturation in the mus-81; xpf-1 double mutant: i) the retarded SYP-1 disassembly in late stage oocytes; ii) the disorganized structure of the axial element and AIR-2 staining in the bivalent; and iii) the presence of a fine DNA bridge between bivalents. To determine if these phenotypes were the result of persistent unresolved JMs, we examined the impact of germline injection of human GEN1 on the presence of DNA bridges between DNA bodies, presumably homologs, in late diakinesis bivalents. Human GEN1 has been previously shown to promote HJ resolution in vitro and in mus81 mutant S. pombe strains [8], [42]. Strikingly, germline injection of human GEN1, but not buffer control, significantly reduced the number of nuclei containing bivalents with DNA bridges from 100% to 19% (Figure 7). Taken together, these results suggested that the defects observed in mus-81; xpf-1 diakinesis oocytes were due to a failure to resolve HJs to produce COs. The appropriate resolution of meiotic recombination intermediates (RIs) is critical for chromosome segregation at the first meiotic division. Errors during meiotic chromosome segregation can lead to aneuploidy and compromise the faithful transmission of genetic material. Organisms have therefore evolved a number of proteins that can resolve joint molecules (JMs) and in some organisms these resolvases act redundantly to ensure that all JMs are processed before chromosome segregation. This redundancy has made it difficult to identify meiotic Holliday junction resolvases in vivo. The complexity associated with Holliday junction resolution is well illustrated by the endonuclease component SLX4 (BTBD12). SLX4 together with SLX1 can resolve Holliday junctions. However, SLX4 also binds to a number of helicases and endonucleases that possess JM processing activity including MUS81, GEN1, BLM, and XPF. This observation has lead to the proposal that SLX4 functions as a platform for the coordination of a number of JM resolving enzymes complicating the analysis of any individual component (for a review, see [43]). It is unknown why JM resolution is so modular, with a great deal of redundancy built into the system. Furthermore, different organisms rely predominantly on different subsets of enzymes to resolve JMs, and it is unclear whether this reflects differences in RIs or whether evolution has shaped the different nuclease preference. In this study, we set out to define the enzymes responsible for JM resolution during meiosis in C. elegans. Unexpectedly, our results implicate that the structure specific endonuclease MUS-81 processes both early and late RIs in C. elegans. We propose that MUS-81: 1) functions to limit the formation or persistence of early RIs marked by RAD-51 that form when the C. elegans Sgs1 homolog HIM-6 is absent; and 2) acts redundantly with the related XPF-1 endonuclease to resolve late JM intermediates required to produce COs. In budding yeast, Sgs1 regulates the processing of meiotic RIs. Loss of Sgs1 function in meiosis results in an accumulation of JMs that require Mus81 and other enzymes for resolution; when both Sgs1 and Mus81 are lost, unresolved JMs persist into anaphase and cause meiotic catastrophe and death [11], [12]. We found that in the absence of HIM-6, MUS-81 was required to prevent the accumulation and persistence of RIs during pachytene. Based on the large number of RAD-51 foci in mus-81; him-6 double mutants it appears that most RIs are processed during pachytene by either HIM-6 or MUS-81. It is estimated that there are between 30–65 recombination intermediates formed per nucleus during pachytene [44]–[46]. This assertion is based on RAD-51 foci in rad-54 mutants, which are compromised for the later steps of homologous recombination downstream of RAD-51 loading onto the processed DSBs. Therefore, 80–90% of all meiotic recombination intermediates are processed to form NCOs by HIM-6, and in the absence of HIM-6, by MUS-81. In the absence of both HIM-6 and MUS-81, early RIs are not resolved and persist resulting in chromosome breakage and inviability. In some cases, fine DNA bridges could be seen between late diakinesis bivalents (data not shown). However, given the large number of unresolved early RIs in mus-81; him-6 mutants, we could not ascertain whether these bridges resulted from early unresolved RIs that persisted to diakinesis such as multichromatid JMs or if they were interhomolog JMs, as seen in the mus-81; xpf-1 mutant. In contrast to MUS-81, XPF-1 does not appear to have a role in processing these early RAD-51-associated RIs. The meiotic phenotype of xpf-1; him-6 mutant animals was no worse than that expected for an additive effect of the two mutations (Figure 4A). MUS-81 and XPF-1 acted redundantly to resolve late RIs, but not the early RIs that arise in the absence of HIM-6. Moreover, mutation in mus-81, but not in xpf-1, was synthetic lethal when combined with mutations in the anti-recombinase rtel-1. We presume this synthetic lethality reflects a failure to resolve aberrant meiotic RI that form in the rtel-1 mutant [23]. Consistent with this hypothesis, mus-81 rtel-1 animals exhibited elevated numbers of RAD-51 foci [47] similar to mus-81; him-6. This data suggested that HIM-6 and RTEL-1 act to limit or remove recombination intermediates and in their absence MUS-81 is needed to process these RIs. It is likely that RTEL-1 and HIM-6 have different roles in the processing of early RIs. Loss of RTEL-1 results in an increase in COs, presumably because D-loops are not disassembled leading to an increase in CO-forming RIs, whereas loss of HIM-6 results in a decrease in the frequency of COs [48]. How HIM-6 promotes CO formation in C. elegans is not yet clear. In budding yeast, the HIM-6 homolog, Sgs1, is proposed to be the central regulator of JM resolution, directing ∼50% of RIs to NCO outcomes before they form stable JM intermediates and the remaining RIs to CO outcomes, perhaps by preventing intersister and multichromatid JMs thereby ensuring that the remaining RIs result in productive JMs that can be resolved as COs [10], [13]. It is possible that HIM-6 functions similarly in C. elegans, since loss of HIM-6 results in an increase in RAD-51-associated RIs in pachytene and as in budding yeast these RIs require MUS-81 for processing. Furthermore, loss of HIM-6 results in a decrease in the frequency of COs. Similar to MUS-81, HIM-18 (the Slx4 ortholog) is required for wild type levels of crossovers in C. elegans. Like mus-81 mutants, him-18 mutants exhibit synthetic lethality when combined with mutations in him-6, with evidence of increased meiotic recombination intermediates [24]. However, him-18; him-6 double mutants do not exhibit an increase in DAPI-stained bodies at diakinesis, unlike mus-81; him-6 mutants. In fact, loss of HIM-18 suppressed the increase in DAPI-stained bodies at diakinesis in him-6. This suppression of additional DAPI-stained bodies suggested that the chromosome disjunction phenotype associated with him-6 mutant animals may be the result of inappropriate processing of RIs by HIM-18-associated endonucleases, leading to premature chromosome disjunction. The multiple binding partners of HIM-18 could account for this difference. Saito et al. [34] report a physical interaction between HIM-18 and MUS-81, SLX-1, and XPF-1. Loss of HIM-18 could limit the activity of all three endonucleases resulting in persistent recombination intermediates and a reduction in the number of DAPI-stained bodies. Whereas loss of either MUS-81, SLX-1, or XPF-1 in him-6 would still allow for endonuclease activity from one of the other endonucleases resulting in inappropriate cleavage and chromosome fragmentation. Overall, our data suggests that HIM-6 and MUS-81 have roles similar to those of Sgs1 and Mus81 budding yeast in the processing of early RIs to prevent the formation of aberrant JMs. Mus81 has been implicated in ZMM-independent COs in a number of organisms and as such is required for most COs in fission yeast, which lack ZMM proteins and crossover interference. Mus81 is also required to resolve a subset of COs in budding yeast and for a subset of COs in the murine male germline [5]–[7], [19]. In C. elegans, COs are tightly regulated with one crossover occurring on each bivalent [49]. Consistent with strong crossover interference, most COs in wild type animals are thought to be ZMM-dependent [22]. In support of this assertion, six ZHP-3 foci, which mark emerging crossovers, are observed in wild type animals [35]. Previously, we found that ZMM-independent COs could occur in certain circumstances, such as in the rtel-1 mutant or after the generation of excess COs by ionizing radiation-induced breaks, and that these COs were dependent on MUS-81 [23]. Surprisingly, we have found that MUS-81 and XPF-1 are also required for wild type levels of COs. Our data suggest that there are either significant numbers of ZMM-independent COs in C. elegans or that MUS-81, XPF-1, and HIM-18 can resolve ZMM-dependent crossovers. Multiple endonucleases are capable of resolving Holliday junctions, complicating the search for eukaryotic resolvases in vivo. It is apparent that in many organisms the resolution of Holliday junctions is buffered by redundant resolvases. For example, in budding yeast there are at least three different endonucleases that can contribute to the resolution of meiotic RIs: Mus81, Yen1, and Slx1/Slx4 [10], [13], [14]. In mice there are at least two pathways for CO resolution: one dependent on Mlh1 and Mlh3 and another dependent on Mus81 [19]. We have found that in C. elegans, meiotic JM resolution depends on the redundant activities of MUS-81 and XPF-1. Both single mutants showed relatively minor reductions in COs and in the resolution of early RAD-51-associated meiotic RIs. However, mus-81;xpf-1 double mutants exhibited severe late meiotic phenotypes in diakinesis oocytes consistent with loss of Holliday junction resolution. Unlike mus-81; him-6 double mutants, the number of RIs in the mus-81;xpf-1 mutant germline was not elevated significantly compared to either single mutant, and defects in meiotic progression was not observed until late in meiotic prophase at the onset of diplotene. The asymmetric disassembly of SYP-1, which is triggered by COs or CO precursors, was significantly delayed in mus-81; xpf-1 double mutants compared to wild type animals (100% vs. 10% of −2 oocytes exhibiting SYP-1 staining, respectively). HTP-3 staining, which marks the axial element, was normal in early meiotic mus-81; xpf-1 nuclei but was highly disorganized in diakinesis oocytes. These data indicate that bivalent maturation, which occurs in response to CO maturation, is compromised. AIR-2, which is concentrated on the short arm of the bivalent at diakinesis, was also disrupted in the mus-81; xpf-1 mutant, further supporting the hypothesis that the CO or CO precursor is abnormal. The most striking observation was that the mus-81; xpf-1 double mutant exhibited fine DNA bridges between the two DAPI-staining bodies of a single bivalent. These bridges occurred between AIR-2 staining regions, supporting the hypothesis that these bridges represent a crossover intermediate; AIR-2 concentration in diakinesis is dictated by CO or CO precursors that act as symmetry breaking events in the C. elegans meiotic bivalent [40]. Finally, the most compelling evidence that these DNA bridges are unresolved JMs came from germline injection of the human Holliday junction resolvase GEN1 into mus-81; xpf-1 double mutants. GEN1 injections rescued the persistent SYP-1 staining on late diakinesis chromosomes (data not shown) and also eliminated the DNA bridges evident in the mus-81; xpf-1 double mutant. These results are consistent with the ability of human GEN1 to both resolve Holliday junctions in vitro [42] and to substitute for Mus81 in promoting crossover formation in fission yeast mus81 mutants [8]. It is interesting to note that endogenous GEN-1 cannot resolve these meiotic JMs in C. elegans. It remains to be determined if this is due to GEN-1 not being active at the appropriate time in meiosis or whether C. elegans GEN-1 lacks the ability to resolve these meiotic JMs. In summary, our data support the hypothesis that MUS-81 and HIM-6 act early in meiosis to limit the formation or accumulation of JMs and that the related structure-specific endonucleases MUS-81 and XPF-1 act redundantly to resolve late JMs, which are likely Holliday junctions, to produce crossovers. Further study into the specific contributions of HIM-6, MUS-81, XPF-1, SLX-1 and HIM-18 will shed light on the nature of joint molecule and recombination intermediate processing, control of crossovers, and evolution of HJ resolution. Strains were cultured as described previously [50]. The strains used in this work include: Wild type Bristol N2, VC193 him-6(ok412), FX1937 mus-81(tm1937), CB1487 xpf-1(e1487), FX2842 xpf-1(tm2842), FX2940 T12A2.8(tm2940), FX1842 F45G2.3(tm1842), DW395 mus-81(tm1937);T12A2.8(tm2940), DW402 mus-81(tm1937);F45G2.3(tm1842), DW116 mus-81(tm1937);xpf-1(e1487)/mIn II, DW485 mus-81(tm1937); him-6(ok412)/nT1[gfp], KR4825 unc-45(e286) dpy-17(e164), KR4821 dpy-17(e164) unc-64(e246), mus-81(tm1937); unc-45(e286) dpy-17(e164), mus81(tm1937); dpy-17(e164) unc-64(e246), xpf-1(e1487); dpy-17(e164) unc-64(e246). Individual animals heterozygous for visible markers and of the genotypes of interest were plated and transferred daily for four days. In each of the broods, wild type, Dpy, Unc and Dpy Unc phenotypes were scored. Recombination frequencies were calculated as in [48]. Individual male animals heterozygous for visible markers and of the genotypes of interest were mated to tester hermaphrodites homozygous for both visible markers and transferred daily for four days. In each of the broods that contained ∼50% male outcross progeny, wild type, Dpy, Unc, and Dpy Unc phenotypes were scored. Recombination frequencies were calculated as recombinants/total brood. RAD-51 immunofluorescence was performed as in [47]. RAD-51 foci in the germlines were assessed when the animals were adults. Foci were counted in 100 RAD-51 positive early to mid-pachytene nuclei. This was done in order to avoid counting earlier nuclei that may not have yet formed meiotic DSBs and later nuclei that may be undergoing apoptosis. Primary antibodies (guinea pig anti-SYP-1, chicken anti-HTP-3, rabbit anti-AIR-2, and guinea pig anti-ZHP-3 (pre-adsorbed against zhp-3 worms)) were all used under standard conditions (as in [47]) at 1∶250. All secondary antibodies were used at 1∶2500 (anti-rabbit Cy3, anti-guinea pig and anti-chicken FITC). DNA was stained with DAPI (0.5 mg/ml) at 1/500. All images were captured using Deltavision microscopy and images were deconvolved using SoftWorx software (Applied Precision). Active C-terminally truncated human GEN1 protein (kindly provided by Steve West), was microinjected into the germline syncytium of adult N2 wild type and mus-81; xpf-1 double mutant animals at 1 ng/µl. Twenty-four hours after injection, germlines were extracted from surviving worms, fixed, and immunostained as detailed above. 20 germlines were scored for each condition. The Mann Whitney test was used to analyze the different conditions. Gaussian approximation was used for calculation of the indicated P-value.

10.1371/journal.pgen.1001079

Identification of the Bovine Arachnomelia Mutation by Massively Parallel Sequencing Implicates Sulfite Oxidase (SUOX) in Bone Development

Arachnomelia is a monogenic recessive defect of skeletal development in cattle. The causative mutation was previously mapped to a ∼7 Mb interval on chromosome 5. Here we show that array-based sequence capture and massively parallel sequencing technology, combined with the typical family structure in livestock populations, facilitates the identification of the causative mutation. We re-sequenced the entire critical interval in a healthy partially inbred cow carrying one copy of the critical chromosome segment in its ancestral state and one copy of the same segment with the arachnomelia mutation, and we detected a single heterozygous position. The genetic makeup of several partially inbred cattle provides extremely strong support for the causality of this mutation. The mutation represents a single base insertion leading to a premature stop codon in the coding sequence of the SUOX gene and is perfectly associated with the arachnomelia phenotype. Our findings suggest an important role for sulfite oxidase in bone development.

Arachnomelia is a defect in skeletal development of cattle. Affected calves are born dead with elongated limbs and facial deformities. The causative mutation for this recessive condition had previously been mapped to a ∼7 Mb interval. We exploited the special structure of cattle families to identify the causative mutation by a purely genetic approach. The rich pedigree records in cattle breeding allowed us to identify the founder animal of arachnomelia, a Brown Swiss bull born in 1957. A few generations later several cattle received two copies of the same chromosome segment from the father of this bull due to inbreeding. One copy was passed through the founder animal and acquired the causative mutation, while the other copy was transmitted through a different line of animals and stayed in its ancestral state. Using next-generation sequencing, we sequenced the entire critical interval in one of these inbred animals. As expected, we found only one single heterozygous position, which consequently represents the causative mutation for arachnomelia. The mutation affects the gene for sulfite oxidase, thus indicating a previously unrecognized important role for this enzyme in bone development. Our findings can immediately be applied to remove this deleterious mutation from the cattle breeding population.

Arachnomelia is a genetic disease in cattle characterized by skeletal abnormalities. Affected calves are usually stillborn with a spidery appearance and an abnormally shaped skull (Figure 1). The bones of the limbs are prolonged (dolichostenomelia) with marked thinning of the diaphyses that fracture easily in the course of forced birth assistance. Additional dysmorphic features are variable, e.g. defects of the vertebral column and sometimes cardiac malformations [1]–[4]. Initially, it was assumed that the pathogenesis of bovine arachnomelia resembles that of human Marfan syndrome, which is caused by mutations in the FBN1 gene [1]. However, the identification of other cattle with a mutation in the FBN1 gene established that arachnomelia is phenotypically distinct from Marfan syndrome [5]. Arachnomelia affected calves lack some typical Marfan features such as joint laxity and aortic root dilation [3]. Bovine arachnomelia is inherited as a monogenic autosomal recessive trait with complete penetrance [2], [4]. Carrier animals do not present any clinical signs. An outbreak of arachnomelia with hundreds of cases occurred during the 1980s after extensive world-wide usage of a highly selected artificial insemination sire in the international Brown Swiss cattle population [2]. In the past few years, a comparable outbreak of arachnomelia occurred in German Fleckvieh cattle [4]. Different genes seem to be responsible for the arachnomelia disease in these two cattle breeds, as there is no relationship between the founder animals and recently two independent loci were genetically mapped to different chromosomes [6], [7]. We mapped the arachnomelia mutation in Brown Swiss cattle to a 7.19 Mb interval on bovine chromosome (BTA) 5 [6]. Due to the lack of suitable candidates the genetic basis of arachnomelia is not understood. Therefore, the spontaneous cattle arachnomelia mutants provide unique resources to gain further insights into the biology of bone development. Array enrichment and next-generation sequencing technology can be used to rapidly sequence targeted subsets of the genome [8], [9]. Sequence capture enrichment has already successfully been used to sequence the coding portion of the human genome [10]–[12] and also large genomic intervals [13], [14]. This technology offers great potential for positional cloning projects where the mapping resolution may be limited, e.g. in the case of recent mutations. Thus megabase sized regions can now be re-sequenced efficiently. Furthermore, the unique family structures in livestock populations greatly facilitate the discrimination of the causative variant from the many neutral polymorphisms that have to be expected from such re-sequencing projects. In this report we applied this approach to identify the causative mutation for bovine arachnomelia. We selected two individuals for an array-based sequence capture and massively parallel sequencing approach to re-sequence the entire critical interval. Our design included one arachnomelia affected calf assumed to be homozygous across the entire sequence interval including the causative variant. The other animal chosen for re-sequencing was a partially inbred non-affected cow. Based on pedigree and marker data this non-affected cow was identical-by-descent for the critical segment of BTA 5, except for the causative arachnomelia mutation, which we predicted to reside only on her paternally derived chromosome (Figure 2). We chose this cow for complete re-sequencing of the entire critical interval, as the detection of a heterozygous polymorphism in this animal should reveal the causative mutation. We enriched the ∼3.5 Mb non-repetitive sequence within the critical interval of BTA 5 in the two animals and collected about 30 million illumina reads per animal. After the alignment to the reference sequence, the mean coverage was 44-fold for the affected animal and 15-fold for the non-affected animal. The difference was most likely due to technical variations during the hybridization. The depth of coverage was variable across the targeted interval, similar to what had been described previously [15]. However, the two animals showed a similar distribution of the gaps across the targeted interval (Figure S1). In the affected calf 96% of the enriched bases had at least 4-fold coverage compared to 91% in the control sample. We called a homozygous variant when the respective position had ≥4-fold coverage and the observed difference between the experimental reads and the reference sequence occurred at ≥75% frequency. For the calling of a heterozygote variant we applied a threshold of ≥15-fold sequence coverage and a variant allele frequency between 25% and 75% based on published recommendations [16]. Using these criteria we recovered a total of 6,025 putative variants between the reads and the reference for the affected calf (4,848 homozygous and 1,177 heterozygous), and 4,318 putative variants for the control cow (3,818 homozygous and 500 heterozygous, Table 1). The high number of seemingly heterozygous variants detected in two animals supposed to be completely or almost completely homozygous underscores the challenges of aligning short reads from a complex mammalian genome to a draft quality reference sequence. About three-fourths of the heterozygous variant calls were located within an olfactory receptor (OR) gene cluster encompassing a 2.2 Mb segment with a total of 136 annotated loci [17]. The read coverage within this region was significantly higher than the average (60/37-fold) indicating segmental duplications. Other factors contributing to the high number of putative heterozygous variants may have been sequencing errors, non-specific hybridization of DNA during the enrichment, or amplification-mediated artifacts (i.e. polymerase errors during library preparation). Due to the recessive inheritance and the lethal effect of the mutation we hypothesized that most likely a loss of function mutation affecting the coding sequence of a gene would be responsible for arachnomelia. Therefore, we subsequently concentrated on variants that were located within the coding sequences or within the splice sites of the annotated genes in the targeted region of the bovine genome. A total of 79 predicted homozygous variants were located within coding sequences or adjacent splice sites of the affected calf (Table 1, Table S1). The comparison between the affected calf and the non-affected cow revealed that 68 of these variants had identical homozygous mutant genotypes in the control cow and could thus be excluded as causative variants. Eight variants had no illumina coverage in the control cow and were subsequently found to be homozygous mutant by Sanger sequencing and thus also excluded. The three remaining variants were putative heterozygous variants in the control cow with 10, 12, and 27-fold coverage, respectively. We validated these three potential variants by Sanger sequencing and found that two of them were false positives as the affected calf and the control cow shared identical homozygous genotypes for these variants. For the remaining variant we confirmed that the control cow was indeed heterozygous, whereas the affected calf was homozygous mutant compared to the reference sequence. This variant was a 1 bp insertion located in exon 4 of the bovine SUOX gene (c.363–364insG, Figure 3). We found perfect concordance between the presence of this insertion and the arachnomelia phenotype (Table 2). All 16 affected calves were homozygous mutant and all 11 available mothers for these animals were heterozygous. Genomic DNA samples of 25 artificial insemination carrier sires, which had recorded arachnomelia offspring and which were related to the assumed founder, were tested and 23 of them were also heterozygous. For each of the two remaining suspected carriers only one single arachnomelia suspicious calf was recorded by the breeding organization and the diagnoses of these calves had not been confirmed by veterinarians. Therefore, we think that these two reported arachnomelia calves represented phenocopies and their sires are indeed free of the arachnomelia mutation. Our material included Beautician, son of the assumed founder Lilason, who was responsible for spreading the mutation into the global Brown Swiss population. We confirmed that Beautician is heterozygous for the SUOX mutation supporting the hypothesis for the origin of the causative mutation. Three acknowledged carrier bulls, which had the same genetic constellation as the non-affected inbred cow of our mutation analysis and were homozygous for all tested markers in the critical interval, were also found to be heterozygous for the SUOX mutation (Figure S2). None of 309 unrelated healthy Brown Swiss cattle had the homozygous mutant genotype, but 10 of them were presumed carriers. Thus the allele frequency of the deleterious insertion within this sample of unrelated Brown Swiss cattle was 1.6%, which is about half of the frequency that was estimated 20 years ago [18]. We screened a genetically diverse panel of animals from 15 breeds widely used in commercial cattle production to confirm that the identified mutation does not occur outside the Brown Swiss population. None of the 150 chromosomes in this sample showed the causative insertion (Table 2). The c.363–364insG insertion is predicted to result in a frameshift beginning with amino acid residue 124 in the bovine SUOX protein sequence (p.Ala124GlyfsX42). While it is unclear whether the mutant protein of 164 residues is actually expressed, with more than 75% of the normal SUOX protein missing, it is very unlikely that the mutant protein fulfills any physiological function. Due to the frameshift and the premature stop codon, any mutant protein produced will contain 42 altered amino acids, and could potentially interfere with normal cellular function. The SUOX gene encodes the molybdohemoprotein sulfite oxidase, a terminal enzyme in the oxidative degradation pathway of sulfur-containing amino acids. Each monomer of the dimeric SUOX enzyme consists of three domains, the N-terminal heme domain, the central molybdenum domain and a C-terminal domain [19]. Deficiency of this enzyme in humans usually leads to recessive inherited sulfocysteinuria (OMIM 272300) characterized by major neurological abnormalities and early death [19]–[22]. The more severe cases, characterized by frequent seizures and death within a few days of birth, result from a complete SUOX loss of function [20], [22]. A single case of human sulfocysteinuria caused by a 1 bp deletion of human SUOX leading to a truncated protein has been reported, where some dysmorphic skeletal features were diagnosed in addition to severe neurodevelopmental anomalies [22]. The arachnomelia phenotype in cattle shows more severe skeletal defects and neonatal lethality compared to the human sulfocysteinuria patients. Another genetic disorder of human sulfur metabolism associated with a bone phenotype (arachnodactyly) is caused by CBS mutations resulting in cystathionine beta-synthase deficiency [23]. In cattle there are two virtually identical arachnomelia phenotypes in Brown Swiss cattle and in German Fleckvieh cattle. The mutation in German Fleckvieh was mapped to BTA 23, a region where the MOCS1 gene encoding molybdenum cofactor synthesis 1 is located [7]. A candidate causative mutation for German Fleckvieh arachnomelia has recently been identified in the bovine MOCS1 gene [Johannes Buitkamp, personal communication]. The MOCS1 protein is required for the synthesis of the molybdopterin cofactor, which forms the active site in SUOX. The involvement of SUOX and MOCS1 in the same biochemical pathway mutually supports the causality of these mutations in Brown Swiss and German Fleckvieh cattle. We think that we have established the causality of the SUOX mutation for arachnomelia in Brown Swiss cattle based on the following arguments: (1) the perfect association of the SUOX mutation to the arachnomelia phenotype, (2) the obvious functional impact of a frameshift mutation on the encoded protein, (3) four non-affected inbred animals, which were identical-by-descent for all tested markers across the critical interval, were heterozygous at the SUOX mutation. The recognition of these four inbred animals was a key element in our discovery and illustrates the potential of livestock specific population structures for genetic research. This study represents one of the first successful applications of microarray-based enrichment of megabase-sized genomic regions followed by massively parallel sequencing to unravel the causative mutation underlying a Mendelian trait. This technology significantly reduces the time and resources required for mutation identification by abrogating the need for high resolution genetic mapping and thousands of Sanger sequencing reactions. In summary, we have successfully applied a sequence capture strategy to identify SUOX as the causative gene for bovine arachnomelia, and thereby discovered an essential role for this gene during bone development. The knowledge of the causative mutation will allow direct genetic testing of Brown Swiss cattle and the elimination of this fatal genetic disease from the breeding population. This study highlights the enormous potential of spontaneous mutants in domestic animals to gain further insights into mammalian biology. The bovine genome assembly Btau 4.0 was used for all analyses. A custom tiling 385k sequence capture array targeting the arachnomelia region (BTA 5, 57,285,788–64,478,535 bp) was designed and manufactured by Roche NimbleGen. The reference sequence contained 179 gaps with a total of 49,558 bp (0.7% of the target sequence). The array was designed using NimbleGen's standard 15-mer frequency masking to minimize repeat content within capture probes. The probe spacing, tiling overlap, and probe length were determined using proprietary algorithms (NimbleGen). For the sequence capture library construction a total of 20 µg high-molecular weight genomic DNA was sheared to yield approximately 400 bp fragments using an ultrasound device and purified on QIAquick columns (QIAGEN). The genomic DNA was polished and repaired using T4 DNA polymerase and T4 PNK (Fermentas). Illumina adapters were added to each genomic DNA sample using T4 DNA ligase (Fermentas). Adapter ligated samples were purified and amplification competency was assessed by PCR with primers complementary to the ligated adapters and finally evaluated by agarose gel electrophoresis. Array hybridization was executed using an X1 mixer (Roche NimbleGen) and the NimbleGen Hybridization System for 3 days at 42°C following the manufacturer's recommended conditions. Human Cot-1 DNA (Invitrogen/Life Technologies) was used at a mass ratio of 5:1 vs. the library. Arrays were washed using the recommended protocol (Roche NimbleGen Arrays User's Guide v2.0). The captured molecules were eluted from the slides with elution reagent using a NimbleGen Elution Station. Eluted molecules were dried by centrifugation under vacuum, rehydrated and PCR amplified for 18 cycles with Phusion polymerase (Finnzymes). Enrichment of samples was assessed by quantitative PCR comparison to the same samples prior to hybridization. Following evaluation by agarose electrophoresis and purification, the amplified capture libraries were processed into sequencing libraries for the illumina GAII. A total of 33,676,855/31,356,905 single reads of different length (36 or 76 bp) comprising 2,245,122,900/2,241,558,140 bases raw data were produced for case and control, respectively. Repetitive elements of the 7.19 Mb genome sequence were masked using the Repeatmasker software and a total of 3,542,013 bp single copy sequence was used as reference for short read mapping. The SeqMan NGen v2.0 software (DNASTAR) was used for assembly with a minimal match percentage of 97%, a minimal match size of 19 nt, and a maximal coverage of 200. A total number of 4,984,147/1,440,459 reads were assembled for the case and the control, respectively. We required a minimal coverage of 4-fold for variant detection and ≥75% of the variant allele for calling a homozygous mutant genotype. Variants falling within the first 10 nt adjacent of masked repetitive sequences were excluded due to obvious alignment inconsistencies (e.g. high coverage, probably due to incomplete masking at the ends of repeats). The post-assembly processing of the variant data was carried out using the R statistical software package [24]. Some variants were genotyped by re-sequencing of targeted PCR products using Sanger sequencing technology. PCR products were amplified using AmpliTaqGold360Mastermix (Applied Biosystems). PCR products were directly sequenced on an ABI 3730 capillary sequencer (Applied Biosystems) after treatment with exonuclease I and shrimp alkaline phosphatase. Sequence data were analyzed with Sequencher 4.9 (GeneCodes).

10.1371/journal.pcbi.1003190

A General Model for Toxin-Antitoxin Module Dynamics Can Explain Persister Cell Formation in E. coli

Toxin-Antitoxin modules are small operons involved in stress response and persister cell formation that encode a “toxin” and its corresponding neutralizing “antitoxin”. Regulation of these modules involves a complex mechanism known as conditional cooperativity, which is supposed to prevent unwanted toxin activation. Here we develop mathematical models for their regulation, based on published molecular and structural data, and parameterized using experimental data for F-plasmid ccdAB, bacteriophage P1 phd/doc and E. coli relBE. We show that the level of free toxin in the cell is mainly controlled through toxin sequestration in toxin-antitoxin complexes of various stoichiometry rather than by gene regulation. If the toxin translation rate exceeds twice the antitoxin translation rate, toxins accumulate in all cells. Conditional cooperativity and increasing the number of binding sites on the operator serves to reduce the metabolic burden of the cell by reducing the total amounts of proteins produced. Combining conditional cooperativity and bridging of antitoxins by toxins when bound to their operator sites allows creation of persister cells through rare, extreme stochastic spikes in the free toxin level. The amplitude of these spikes determines the duration of the persister state. Finally, increases in the antitoxin degradation rate and decreases in the bacterial growth rate cause a rise in the amount of persisters during nutritional stress.

Bacterial persistence plays an important role in many chronic infections. Persisters are subpopulations of bacteria which are tolerant to biological stresses such as antibiotics because they are in a dormant, non-dividing state. Toxin-antitoxin (TA) modules play a pivotal role in persister generation and bacterial stress response. These small genetic loci, ubiquitous in bacterial genomes and plasmids, code for a toxin that slows down or halts bacterial metabolism and a corresponding antitoxin that regulates this activity. In order to further unravel the intricate autoregulation of TA modules and their role in persister cell formation, we built stochastic models describing the transcriptional regulation including conditional cooperativity. This is a complex mechanism in which the molar ratio between both proteins determines whether the toxin will behave as a co-repressor or as a de-repressor for the antitoxin. We found that the necessary protein production and therefore the energetic cost decreases with increased binding site number. Finally, these models allow us to simulate the formation of persister cells through rare, stochastic increases in the free toxin level. We believe that our analysis provides a fresh view and contributes to our understanding of TA regulation and how it may be related to the emergence of persisters.

Stress response is an important aspect of the physiology of bacteria, allowing them to deal with a continuously changing environment and exposure to altering and fluctuating food sources as well as life-threatening chemicals such as antibiotics. Among the elements involved in bacterial stress response are the type II toxin-antitoxin (TA) modules [1], [2]. These are found in prokaryotes as pairs of genes encoding a protein that interferes with basic metabolism (the toxin) and its regulator (the antitoxin). The toxins display a variety of three-dimensional folds and biochemical activities: CcdB and ParE family members inhibit gyrase [3], [4], although via different molecular mechanisms. MazF toxins are structurally similar to CcdB but function as ribonucleases that degrade specific mRNAs and/or modify ribosomes [5]–[7]. RelE toxins, however, are structurally related to ParE but bind at the A site of the ribosome and degrade mRNAs in a translation-dependent manner [8], [9]. Other toxins such as HipA and Doc arrest translation without RNA degradation, for example through phosphorylation of elongation factor Tu [10]. A variety of biological roles have been attributed to TA modules ranging from molecular parasites over the stabilization of genetic elements (plasmids, introns and labile chromosomal segments) to altruistic suicide and the generation of persister cells. Persisters are cells which exhibit multidrug tolerance, not because of a specific resistance mechanism like a mutation in an antibiotic target, but because they are in a dormant, slow-growing state. Cell-wall synthesis, translation and topoisomerase activity are slowed down in dormant cells, making it impossible for bactericidal antibiotics, whose targets are often implicated in these general metabolic processes, to kill the cells [11]. Persisters pre-exist in bacterial populations [12], they are subpopulations that allow survival of the bacterial colony in the case of severe environmental stresses. As such, they are involved in the multidrug tolerance of biofilms and the recalcitrance of bacterial infectious diseases [11]. As expression of TA toxins can bring cells in a “dormant state” or reversible stasis [13], TA modules have been linked to the development of persisters [14]. Each type of toxin is associated with one or more types of antitoxins, leading to a large number of TA families [15]. The antitoxins are typically two domain proteins consisting of a folded common DNA binding domain (helix-turn-helix, ribbon-helix-helix, AbrB fold, etc.) associated with an intrinsically disordered toxin-neutralizing segment that folds upon binding. Regulation of the toxin activity is achieved by balancing the synthesis and proteolytic degradation of the antitoxin [16]. Therefore TA modules are typically activated (for example during nutritional stress) by an increased activity of housekeeping proteases such as Lon and ClpXp [17], [18]. The neutralization function of the antitoxin is however not necessarily passive. In certain cases such as gyrase poisoning by CcdB, the intrinsically disordered domain of the antitoxin was shown to play an active role in reactivating the stalled gyrase molecules [19]. TA modules are further regulated at the transcription level by a mechanism termed “conditional cooperativity” [20]. Here the toxin acts as a co-repressor or anti-repressor depending on the ratio between toxin and antitoxin. When either an excess of toxin or antitoxin is available in the cell, transcription will occur until the cellular ratio is balanced and a repressing toxin-antitoxin complex is predominantly formed. Conditional cooperativity has been observed in all classic type II TA families where it was investigated, independent of the toxin or antitoxin fold, the operator size or the toxin target [21]–[24]. The molecular mechanisms leading to conditional cooperativity vary and involve a low-to-high affinity switch going from a repressing to a non-repressing toxin-antitoxin complex and/or steric exclusion principles [19], [23], [25]. Since the toxins interfere with the basic bacterial metabolism, free toxins can have an inhibitory effect on bacterial growth. This growth inhibition, in turn, can lead to changes in gene expression, as the RNA transcription rates and the protein dilution rates depend on the growth rate [26]. Although relevant results were obtained with models excluding these toxic effects (see for example [27]), it is clear that including the interaction of the toxin-antitoxin module with the host bacterium will lead to more realistic conclusions. The impact of gene circuits on host physiology can lead to drastic changes in the dynamics of the gene circuits themselves, as demonstrated by Tan et al. [28], who found that bistability in the expression of a mutant T7 RNA polymerase was caused by the reduction in growth rate due to the expression of this non-toxic protein. Furthermore, Nevozhay et al. recently showed how the interplay of individual cell growth rate and cellular memory jointly determine the overall cell population fitness in a bistable synthetic gene circuit when including variable division rates of single cells [29]. Several mathematical models have already been used in the study of persister cell formation. For example, Balaban et al. modeled the phenotypic switch between normally growing cell populations and persisters, discriminating two different types of persisters, one generated during the stationary phase and one which spontaneously arises during growth [12]. As TA modules are implicated in the formation of persister cells [14], [30], [31], the regulatory network of these systems has been modeled as the underlying cause of persister generation. Koh and Dunlop built a model for the hipBA TA module, including transcription, translation and repression of gene expression by the antitoxin and a toxin-antitoxin complex [27]. They argue that persistence is not caused by bistability, but by stochastic fluctuations in the expression of HipA and HipB, causing the free toxin level to exceed a threshold. The autoregulation of the relBE module was studied by Cataudella et al., who found that conditional cooperativity prevents random toxin activation in growing cells and promotes fast translational recovery by quickly removing the free toxin after a period of starvation [32]. Although these publications significantly contributed to our understanding of persister cells and TA modules, we believe that a general modeling framework that includes conditional cooperativity and that is applicable to several toxin-antitoxin families could help to answer several of the remaining questions in this field, such as the role of multiple binding sites on the operator and the effect of toxin-dependent cell growth rate modulation. This paper presents a theoretical analysis of transcription regulation by conditional cooperativity based upon parameters available for the ccdAB (F-plasmid), phd/doc (bacteriophage P1) and relBE (E. coli) modules, three TA modules that are well characterized. We study both the molecular mechanism observed in the relBE TA module, where the binding sites on the operator are considered independent, and present the first mathematical model for the mechanism observed in the ccdAB and phd/doc TA modules, where an interaction between the different binding sites on the operator exists as chains of alternating toxins and antitoxins can be formed on the DNA. We model TA modules based on all essential interactions in three well-studied systems: the F-plasmid ccdAB, bacteriophage P1 phd/doc and E. coli relBE operon (Figure 1). Common to these systems is that the toxin and antitoxin can form complexes with distinct stoichiometries and DNA binding properties. In the figure the free antitoxin (A) and the free toxin (T) correspond to the biologically relevant species and are typically dimers for the antitoxin, but can be monomers (, ) or dimers () for the toxin, depending on the TA module considered. The AT complex (corresponding to the molecular species -, - and -) has a higher affinity for the operator sites than the isolated antitoxin. The TAT species consists of two toxins flanking a single antitoxin dimer - corresponding to --, -- and -- species. As the DNA binding properties of this species are dependent on the TA module considered, they will be discussed below. The TA operator has one or more binding sites (denoted with in Figure 1A) for A, AT and/or TAT. It is assumed that transcription is halted when at least one molecule (A) or complex (AT or TAT) is bound on the operator. When no proteins are bound on the operator, the genes coding for the toxin and antitoxin are transcribed. Translation of the mRNA leads to the creation of toxin and antitoxin. In TA modules the translation rate for the antitoxin has been found to be larger than the one for the toxin [17]. Therefore, when the toxin-antitoxin operator is being freely expressed, initially more antitoxin than toxin will be created. However, the antitoxin is degraded faster than the much more stable toxin, influencing the steady state toxin∶antitoxin ratio. The degradation of the antitoxins also extends to those within the complexes AT and TAT. Although the bound toxins protect the antitoxin from proteolytic degradation, this protection is not complete and the decay of antitoxin within complexes allows for the release of the attached toxins. When it comes to the DNA binding interactions at the operator, there are fundamental differences between the three studied TA systems. The mechanism in the relBE system is the basis for the “independent binding site model” (Figure 1B). In this model, it is assumed that the binding sites on the operator behave independently, and either an individual antitoxin or a toxin-antitoxin complex can bind to each binding site. Conditional cooperativity is included in this model by ensuring that the AT complex has a higher affinity for the binding sites on the DNA than the antitoxin alone. Therefore, the toxin can act as a co-repressor for the antitoxin. Furthermore, we assume that the binding of an extra toxin to a DNA-bound AT complex will lead to the detachment of a TAT complex (shown in the final step of Figure 1B) from the promoter/operator, enabling mRNA transcription to proceed if all binding sites are unbound. The toxin can therefore function as a derepressor in the autoregulation of the operon by removing (through the formation of the secondary complex TAT) the bound proteins when the ratio of total toxin to total antitoxin (∶) is high. The second or “interacting binding site” model (Figure 1C) considers more complex binding processes on the DNA, experimentally observed for the ccdAB and the phd/doc modules [19], [23]. In this case, toxins can bridge different binding sites on the operator, forming a chain of alternating toxins and antitoxins. When the ∶ ratio increases, this complex can again be released from the DNA: An extra toxin comes in and the soluble TAT complexes are formed (shown in the final step of Figure 1C), as it is impossible for TAT complexes to occupy adjacent binding sites on the promoter/operator due to steric clashes. Finally, in the last two sections of the results, we add the toxic effect. We assume that above a certain threshold, free toxin inhibits the bacterial growth. As noted by Klumpp et al. [26], a decrease in the growth rate will be reflected by a decrease in the transcription rates, therefore, we include the effect of the free toxin levels on the transcription rates as well. In order to determine the influence of the parameter set used on the behavior of the toxin-antitoxin module, we performed stochastic simulations of a toxin-antitoxin module with two independent antitoxin binding sites on the promoter/operator, and the parameter sets for the ccdAB, phd/doc and relBE system (see Table 1 and Figure 2). Initially, mRNA is transcribed and toxin and antitoxin are translated (Figure 2A and B) as the operator DNA is initially unbound (Figure 2E). Once the operator gets bound, repression starts and mRNA transcription stops. From Figure 2E, it can be seen that after this initial response, the operator DNA is mostly occupied by one or more protein species. Pulses in the toxin and antitoxin level occur during short periods when the operator becomes unoccupied in a single cell. The free toxin population is retained at very low levels as is expected in a growing cell population [17]. For the relBE system, a maximum of eight free toxins is found for the simulated cell shown in Figure 2 and the average free toxin level is approximately one, whereas for the ccdAB and phd/doc systems, the average free toxin level is much lower than one (Figure 2B). The overall majority of toxin molecules are thus sequestered into toxin-antitoxin complexes AT (Figure 2C) or TAT (Figure 2D). Although there are slight differences in the protein and complex concentrations and the number of binding/unbinding events on the DNA, the behavior of the toxin-antitoxin model with two independent binding sites on the operator is qualitatively similar for the ccdAB and the phd/doc parameter sets. The average levels are similar for the unbound DNA, antitoxin, toxin and complexes AT and TAT although the plotted single cell behavior differs due to the stochastic nature of the simulations. The outcome of the simulations changes more significantly when the relBE parameter set is used. The most remarkable difference is the increase in the free toxin level for the relBE module. This is not illogical as the molecular mechanism for conditional cooperativity used by relBE is distinct from the mechanism employed by ccdAB and phd/doc and several parameters such as the DNA binding rates are very different. Considering the similarities in the output, for simplicity we only use the ccdAB parameter set in the simulations presented in the remainder of this work. Two mechanisms are responsible for managing the potentially lethal toxin-antitoxin modules in bacteria. At the protein level, free toxins can be neutralized by complex formation with a free antitoxin or a non-saturated toxin-antitoxin complex AT [14]. At the transcriptional level, the negative autoregulation of the operon by conditional cooperativity ensures that the production of antitoxins and toxins is repressed when more antitoxin than toxin is present. When an excess of toxin emerges, the transcription is derepressed and the antitoxin will be the main product of translation, as explained above. To study the role of both levels in the regulation of toxin-antitoxin modules, we performed a series of simulations with the “independent binding sites” model, in which either operator binding or the sequestration of the toxin in complex TAT or in both non-toxic complexes (AT and TAT) were eliminated. When DNA binding by both antitoxin and toxin-antitoxin complexes is abolished during the simulation, the free toxin level remains fully controlled (see Figure S1). This shows that the sequestration into the complexes AT and TAT without any gene regulation accounts for a complete suppression of the toxin, albeit with a higher level for antitoxin and complexes AT and TAT. Alternatively, in simulations where formation of the secondary complex TAT (and therefore also conditional cooperativity) is eliminated, the cell continues to control the toxin level, although more variability in the antitoxin level is observed. When the formation of both complexes AT and TAT is abolished, but DNA binding remains included, the cell does not manage to control the free toxin level and produces as much toxin as antitoxin. This suggests that AT formation is necessary for the control of the free toxin level and TAT formation helps to reduce the variability in the antitoxin level. Different toxin-antitoxin modules have different numbers of binding sites on their operator, ranging from two in the phd/doc and relBE system [20], [33] to eight in the ccdAB system [34]. We investigated the influence of this property on the levels of free toxin, free antitoxin and non-toxic complexes in a toxin-antitoxin system with independent binding sites on the operator. In Figure 3A, we plot the Probability Density Functions (p.d.f) for each of the protein components of the toxin-antitoxin system. The p.d.f. is constructed by simulating the time evolution of many cells and detecting the protein level at each point in time. Using this information we calculated the probability to find a certain number of protein components in a cell. It can clearly be seen that increasing the number of binding sites on the operator leads to decreased protein levels and variability for the free antitoxin and complexes (AT and TAT), while the free toxin level stays low and relatively constant. This decrease in the protein concentrations allows a more economical maintenance of the toxin-antitoxin system. The increase from one to two binding sites on the operator has the most profound effect on the protein levels. The mean value for each distribution is shown in Figure 3B and there is a linear relationship with the reciprocal of the number of binding sites () on the operator for A, AT and TAT. There seems to be a direct correlation between the free toxin variability and the number of binding sites, however the absolute magnitude of this phenomenon in comparison to the total amount of antitoxin and complexes makes this relationship negligible. The p.d.f. for both of the non-toxic complexes can be described by a normal distribution for a toxin-antitoxin system with one binding site on the operator. However, with an increasing number of binding sites these distributions become bimodal. The extra peak at low complex concentrations may be explained by the fact that the antitoxin level equals zero more often as the number of binding sites on the operator increases (see Figure S2 and S3). The effect of the number of independent binding sites on the operator on the time evolution of the antitoxin and toxin level and the binding on the DNA is shown in more detail in Figure S2 and S3. When the operator only consists of one binding site, many fast DNA binding and unbinding events are observed. This leads to an evenly distributed response around the average for the mRNA production and therefore the free toxin and antitoxin level. With an increasing number of antitoxin binding sites on the operator, the probability of the operator being bound by at least one antitoxin increases as well. This leads to localized bursts in time of mRNA creation and corresponding spikes in the free toxin and antitoxin levels. Conditional cooperativity is included in both the model for independent binding sites and the model with interacting binding sites, since the toxin can derepress the operon at high ∶ ratios. In the former model, this is due to the assumption that a TAT complex is unable to bind the DNA. In the latter model, this is due to the fact that “stripping” of a protein chain from the promoter/operator can occur when a low affinity interaction in this chain is replaced by a high affinity interaction with a new toxin, forming soluble TAT complexes that are unable to occupy adjacent binding sites on the operator due to steric hindrance. The role of conditional cooperativity in the regulation of TA modules is studied in the following simulations by abolishing the formation of TAT complexes on the DNA and their subsequent release (independent binding sites) or the stripping reaction (interacting binding sites). In both cases, the formation of TAT complexes in solution is still possible. When the operator consists of independent binding sites (Figure 4A and C), the unbinding rates of antitoxin and AT from the operator are large enough to free the promoter and allow mRNA creation. Therefore, conditional cooperativity has no profound effect on the system dynamics as the DNA binding reaction rates control the behavior of the toxin-antitoxin system. In the model with interacting binding sites (Figure 4B and D), however, conditional cooperativity is of essential importance to free the DNA promoter/operator from the chain of alternating toxins and antitoxins bound to it, so that transcription can occur and the antitoxin can be expressed (Figure 4B). In this case, the toxin level can be controlled. In the absence of conditional cooperativity, the promoter/operator remains bound as the unbinding rates are too slow to completely free the DNA from the protein chain. In this situation, no toxin or antitoxin is expressed. As the antitoxin will be degraded more rapidly, a large increase in the free toxin level occurs, inducing a cessation of cell growth or cell death (Figure 4D). Please note that the decrease in the toxin level after this spike is not necessarily found in vivo. This decrease is caused by toxin dilution due to cell division, as we assumed that the doubling time of E. coli is constant. Furthermore, the synchrony in the average antitoxin and toxin concentrations in Figure 4B is caused by the initial conditions being identical for all cells. These coherent oscillations disappear after longer simulation times, but reflect the presence of a well-defined time between spikes in the free toxin and antitoxin level. Such coherence is not likely to be found in an actual bacterial population though due to the lack of similar synchronous initiation of the different cells. The translation rates for the antitoxin and the toxin, and , are hard to determine experimentally but are important parameters for the behavior of the toxin-antitoxin module. The translation rates in this article are based on the average translation rate in E. coli, on the lengths of the proteins, on the fact if monomers or dimers are formed in solution (immediate dimerization is assumed and therefore the translation rate is halved in the case of dimers) and on the translational coupling, ensuring that toxins are produced at a lower rate than antitoxins. In order to show the influence of variations in both translation rates, Figure 5 shows the free antitoxin and the free toxin level in the parameter plane (, ), using the model for independent binding sites on the operator. Two regions are clearly visible in this parameter space: One in which the free antitoxin level is high and the free toxin level is controlled (on average less than one free toxin per cell is present) and one in which the free toxin level is very high with negligible amounts of free antitoxin present, corresponding to a non-culturable cell population. The latter region is indicated as [K] in Figure 5. There is a clear threshold between these two cell populations, which is crossed when the toxin translation rate exceeds twice the antitoxin translation rate. In all currently investigated TA modules, the synthesis rate for the antitoxin was higher than the one for the toxin [17]. Therefore, these modules can be safely maintained in a cell population. The lower panels in Figure 5 show the effect of on the ∶ ratio, and on the total protein number (keeping as in all previous simulations). From these plots it can be seen that in controlled, stable cells the total amount of toxin is always lower than twice the total amount of antitoxin. This can be explained by the fact that one antitoxin can maximally neutralize two toxins for the investigated TA modules [25], [34], [35]. At the boundary , the critical ∶ value of 0.5 is reached. If the toxin translation rate is further increased, the total level of toxin is larger than twice the antitoxin level and free toxins can accumulate. When approaching the boundary, the total protein level in the cell also becomes increasingly large. This boundary can be found analytically from the deterministic version of the “independent binding sites” model under certain assumptions (see Text S1). TA modules are involved in the emergence of persister cells [14], [30], [31]. In the following paragraphs, we check which parameters and assumptions are necessary to allow a persister to be formed, and reveal one possible avenue to persistence. A parameter scan in the translation rate of the antitoxin and the toxin was performed, both for the model with three independent binding sites on the operator as for the model with three interacting binding sites. The top panels in Figure 6 show the percentage of cells that reach a toxin level higher than 100 during a time interval of 500 minutes. In accordance with Figure 5, a sharp transition from 0% to 100% can be observed when the translation rate of toxin exceeds twice the translation rate of antitoxin for the independent binding site model. Below this boundary, the toxin level is controlled in every cell (see for example Figure 4A). When crossing this boundary, the toxin level continuously grows to large values (see Figure S5B). Of course, the in vivo response may differ from the shown simulation once the toxin level reaches a sufficiently high level, as the toxic effect is not explicitly modeled in this simulation. However, as high toxin levels would be present in every cell, growth of a bacterial cell population would be impossible in this region of parameter space as indicated above (Figure 6 region [K]). In the case of three interacting binding sites on the operator, extra effects come into play. For every cell still experiences a continuously growing toxin level (see Figure S5B). However, in the experimentally most relevant case, where the translation rate of toxin is smaller than the translation rate of antitoxin [17], the observed response differs from cell to cell. In this region, two types of response are possible with different probabilities. The cell can have a stable low toxin level, controlled by regular oscillations in the antitoxin level (see Figure 4B and Figure 6A). In this case, each increase in the toxin∶antitoxin ratio is followed by the release of the protein chain from the DNA, causing a spike in the mRNA, antitoxin and complex levels, respectively, and keeping the free toxin level close to zero. This response is similar to the one in the case of independent binding sites on the operator. The other possible response is that the cell produces a large pulse of toxin (see Figure 6B). The toxin level does not continuously grow, but its growth is arrested after some time. However this is abated since after this occurrence the system quickly returns to its controlled state, because no toxic effect was included in this simulation. This rare event can be stochastically initiated if a TAT complex is still bound on the DNA when the ∶ ratio reaches the level of two, this is when the level of the antitoxin and the AT complex are very low or zero. In this case, the full chain of alternating toxin and antitoxin dimers can no longer be formed on the DNA. As conditional cooperativity is unable to free this complex from the DNA, the free toxin level will rise as long as the TAT complex does not unbind from the DNA. Please note that this rise in free toxin level is caused by the degradation of the antitoxin within the complexes AT and TAT, and the concomitant release of toxin molecules. Therefore, degradation of antitoxin within the toxin-antitoxin complexes is necessary to obtain persister cells in this framework. The probability of having toxin spikes, and therefore the potential of persisters occurring in the population, increases as one approaches the line. The toxin spike becomes increasingly high with increasing values of (see also Figure S4). In the region every cell will reach toxin levels higher than 100, but the response can either be a toxin spike or a continuously growing toxin level (see Figure S5A and B respectively). The percentage of the cells responding with continuously growing toxin levels increases (to 100%) as one approaches the line. The bottom panel of Figure 6 shows a more detailed analysis of the probability of obtaining large amplitude toxin spikes for the normal parameter values as presented in Table 1. The number of toxin spikes having an amplitude larger than 10 is numerically detected. One observes two characteristic scales. The first one is associated to regular stochastic fluctuations of the toxin amplitude under normal operation (see Figure 4B and Figure 6A). The probability of finding toxin spikes of increasingly high amplitude decreases exponentially. The second scaling can be attributed to the different mechanism where a TAT complex remains bound to the DNA for a certain time, as mentioned above. Provided that the binding affinity of TAT to the DNA operator site and the toxin translation rate are large enough, rare high amplitude toxin spikes can be observed (see also Figure S4). To obtain a more realistic view of persister cell formation, the duration of persistence and the influence of free toxin levels on the growth rates, we introduced toxic feedback effects into both the independent and interacting binding site models. Once the free toxin level crosses a threshold , the growth rate and transcription rates decrease. This decrease is modeled by a Hill function where the Hill factor determines how sharp the transition is around the threshold value . A minimal growth rate is defined to ensure that the cell can always recover after a (potentially long) time (see also Materials and methods). In the independent binding site model toxic feedback effects have only marginal impact and no long-term persister dynamics are found (see Figure S6). Panels A and B in Figure S6 show a direct comparison between simulations carried out without feedback (see also Figure 3) and with feedback. Even in the case of a very low threshold, the protein levels remain very similar and almost no difference is observed in the p.d.f. Panel C in Figure S6 shows the effect of the threshold level chosen and its impact on the normalized individual fitness or growth rate (where in the case of no growth rate reduction). The fitness is detected at each point in time of the simulation and used to calculate its probability at any given time. At high threshold values, free toxin levels remain too low to be able to cause a noticeable reduction in fitness. When decreasing the threshold, the fitness landscape is broadened due to stochastic excursions of the free toxin level, allowing for lower growth rates. We have calculated the average fitness R by taking the first moment of the probability distribution of the individual fitness (R = ) and it is displayed in the legend. Although a modulation of the growth rate can be obtained, at no point is the dynamics altered and no clear switch to a persister state is observed. The simulations using the interacting binding site model with toxic feedback effects are shown in Figure 7, where we have decreased the translational coupling by a factor of three () such that toxin spikes are more likely to be found. Panels A and B show a simulation without and with the inclusion of toxic feedback effects, respectively. When no feedback is included the system responds to the toxin spike by complex sequestration that causes a return to nominal levels of toxin. However, with feedback a toxin spike of significant size can cause the system to switch to a persistent state for multiple cell cycles where there is no antitoxin present to neutralize the toxin levels. The duration of this persister state is closely related to the spike amplitude, as the recovery time to switch back to normal operation is mainly determined by the time it takes for the toxin level to drop due to (slow) dilution. This close relation between toxin spike amplitude and duration is shown in Panel C. Without toxic feedback the red cluster of points shows a clear correlation between spike amplitude and duration (see inset). When only introducing a toxic effect on the transcription rates, this cluster of points is split in two separate ones (see clusters a and b in green). If one also introduces a toxin-dependent growth rate modulation (see blue points), cluster (a) remains similar, but the second cloud of points (b) shifts to duration times that are orders of magnitude larger. This is immediately reflected in the fitness landscape shown in Panel D for the three different cases. Including cell growth modulation, one can now observe that it is most probable to find the cell in a state with fitness . However, there is a clear second peak in the probability distribution at a much reduced fitness . The shape of this bimodal distribution function (such as for the relative heights of both peaks) can be controlled by changing the various system parameters. Similarly the average fitness R can be controlled. The bimodal response is qualitatively very different from the case in the independent binding site model and originates from the possibility to create the persister states (where the fitness can be decreased for longer periods of time). Similar bimodal effects have been studied in other papers [26], [29], [36]. However, no bistability is present in our model when including the toxic feedback, provided the minimal fitness is non-zero. The system remains monostable, but the bimodal response results from stochastically triggered transient excursions during which the individual fitness is very low. Panel E shows a sketch of normal exponential cell population growth (left) and a reduced growth of a cell population due to persister cell creation (right). The average fitness of a population of cells can be decreased through the presence of persister cells which are highlighted in red. Since these cells have their growth arrested at points they do not divide on the usual time scale as the normal (black) cells. This is why the population which has these persisters (right) can have a lower population number or slower growth in comparison to a population without persister cells (left). During nutritional stress, the antitoxin degradation rate increases due to the activation of cellular proteases like Lon [37]. Furthermore, the rate of protein synthesis decreases to approximately 5% of the pre-starvation level [38]. We thus investigated the influence of the antitoxin degradation rate and the antitoxin and toxin translation rates ( and ) on the free toxin level in the independent binding sites model (see Figure S7). It can be observed that the boundary for the viability of a cell population does not change when increasing the antitoxin degradation rate. For the viable cells, the increase in the antitoxin degradation rate is mainly responsible for the increase in the average free toxin level associated with nutritional stress [38]. A decrease of both the toxin and antitoxin translation level with the same factor will not heavily affect the free toxin level. We further investigated if the increase in antitoxin degradation and the decrease in translation rates and in growth rates associated with nutritional stress also affect the formation of persister cells. In Ref. [12], Balaban et al. outlined a model for persisters created through normal growth (type II) and showed a switch from normal behavior to persistent activity in a population. The model has two states, normal (N) and persister (P), the switching rate from N to P and P to N are defined as a and b, respectively, while the growth rate of both states are given by and :(1)(2) The same model can be used to analyze the growth of cell populations in our case. The mentioned switching rates a and b can be directly estimated from our simulations and changes in these rates can be linked to underlying system parameters and the corresponding dynamics. Both growth rates and correspond to both peaks in the bimodal fitness distribution (, ). We used our model to analyze the fitness landscapes and estimate corresponding switching rates resulting from the changes in growth, translation and antitoxin degradation rates. In the top panels of Figure 8A, using the standard parameter set, no well-separated persister population is found in the scatterplot showing the toxin spike amplitude vs. the spike duration. This absence of a clear family of persister cells is also reflected in the fitness landscape where no bimodal response is found. However, when the growth rate is halved, there are two distinct populations of cells (see bottom panels A). In addition to the population dividing at a normal growth rate, there is a fraction dividing at a fitness . The estimated switching rates show that the transfer from normal to persister state occurs at a faster speed (a>b) than the return. This difference in speed becomes more pronounced when also including a reduction in translation rate and antitoxin degradation rate (see Figure S8). Using Eqs. (1)–(2), the time evolution of both cell populations can be simulated in time. The resulting persister fraction is shown in Figure 8B for the switching rates as estimated from the case with halved growth rate. Independently from the initial conditions, the persister cell fraction evolves to a steady state solution after about 300 minutes. A similar two-state population dynamics model has been used in Ref. [29] to understand how the combination of cellular memory and individual fitness jointly define the overall distribution of cell populations. In this work, phenotypic switching rates were estimated in a bistable system of high and low expressers, and it was shown how cell lineage statistics can be different from population snapshot statistics. The authors concluded that cells tend to switch predominantly to the high expression state and switch back much more rarely. This translates to a>b, which agrees with our findings in the presence of nutritional stress. An example of this behavior can be seen in Figure 7B, where a typical time series is shown of an individual cell lineage. It is clear that the cell spends most of its time in the persister state. Looking at the persister fraction of the overall cell population, however, only a minority of the cells are in a persister state. The persister fraction in the overall cell population is greatly determined by the switching rates to get into a persister state and to escape from it. This escape time is essentially determined by the reduced fitness in such a persister state. Although we have found that during nutritional stress a>b, the normal cell population still dominates due to its much larger individual fitness with respect to the persisters. Figure 8C shows an analysis of the dependence of the persister fraction on both switching rates a and b. One can clearly see that the switching rate to get into persistence strongly controls the persister fractions, such that its increase in nutritional stress conditions immediately leads to an increased persister fraction. The return rate to normal operation (b) has practically no influence on the persister fraction, provided that it is slower than the decay rate due to dilution of the normal cell population (related to its growth rate). Strong evidence has been accumulating that various types of bacterial toxin-antitoxin modules are implicated in persister cell formation [14], [30], [31]. In the present paper we investigated how the peculiar type of gene regulation called “conditional cooperativity”, that seems to be a common feature of TA modules, is capable of controlling the cellular free toxin levels and might control the formation of persisters. We successfully constructed two models for the autoregulation of toxin-antitoxin modules by conditional cooperativity, which mirror two molecular mechanisms that allow for conditional cooperativity [19], [23], [25]. In the first model, we consider the binding sites on the operator as independent entities on which antitoxins and AT complexes can bind, whereas in the second model, the toxins can bridge the antitoxin-bound binding sites on the DNA. Stochastic simulations based upon these models showed several essential characteristics of TA modules, such as very low free toxin levels and high free antitoxin levels in non-starvation conditions, and this for three different parameter sets derived from experimental data available for F-plasmid ccdAB, bacteriophage P1 phd/doc and E. coli relBE. We found that sequestration of toxins in toxin-antitoxin complexes and not gene regulation is responsible for the main control of the free toxin level as a viable toxin-antitoxin balance is still maintained in absence of any regulation (removing the DNA binding properties of the antitoxin from the model). However, when the DNA binding reactions are included in the “interacting binding site” model, the “stripping” reaction (binding of T to AT to obtain a TAT species that quickly dissociates from the operator) is still necessary to allow fresh antitoxin synthesis and therefore maintain viable free toxin levels. The stripping reaction, which has a pivotal role in the conditional cooperativity, allows the toxin to function as a derepressor for the operon by releasing the chain of alternating toxins and antitoxins from the DNA at high ∶ ratios. In the independent binding site model, such a chain cannot be formed. Therefore, the dissociation rates of the antitoxin and the AT complex from the DNA are sufficiently high to free the operator, allowing antitoxin synthesis and subsequently toxin neutralization. We further found that the toxin level can be controlled in the presence of lower amounts of antitoxin and toxin-antitoxin complexes if the number of binding sites on the DNA increases. Therefore, the maintenance of a TA module becomes more economical for the cell as the amount of binding sites on the operator increases. This may be the reason why the ccdAB module has evolved to have as much as eight binding sites on the operator. When considering independent binding sites on the operator, parameter scans reveal a clear threshold between healthy, antitoxin dominated, and non-culturable, toxin dominated cell populations, which is crossed when the toxin translation rate is more than double the antitoxin translation rate. In the model with interacting binding sites on the operator, toxin accumulation also occurs in all cells above this boundary. In all studied TA modules, the antitoxin translation rates are higher than the toxin translation rates [17]. In this region in parameter space, most cells have a low free toxin level, but in very rare cases the free toxin level spikes, which can lead to the formation of a persister cell. This steep increase in the free toxin level can occur when the operator is occupied and no new antitoxin can be made at a moment when the free antitoxin level is very low. In this case, the degradation of antitoxin in toxin-antitoxin complexes leads to the accumulation of free toxins, which can perform their toxic activity. This toxic activity is added in certain simulations by decreasing the growth rate and the transcription rate once the free toxin concentration exceeds a certain threshold. In this case, the level of free toxin determines how long the toxin spike lasts and how long the cell resides in the persister state. A similar result was obtained by Rotem et al., who found that bacteria go into a dormant state once the toxin level crosses a threshold, and that this toxin level determines the length of the dormancy [36]. In reality, more complex toxic feedback effects can also take place, dependent on the TA module considered. For example, in the case of RelE or Doc, translation would be inhibited in vivo. As multiple TA modules can be present in one bacterium, the inhibition of translation by one toxin could lead to an increase in the concentration of other toxins as suggested by Keren et al. [14]. In order to obtain persister cells during our simulations, it was necessary to assume that antitoxins can be degraded within the toxin-antitoxin complexes. It was also previously shown that such degradation can play an important role in TA modules, as a switch from an antitoxin dominated state to a toxin dominated state upon amino acid starvation was only possible for the relBE system when the active degradation of RelB within toxin-antitoxin complexes was taken into account [32]. Moreover, we found that the increase in the amount of persisters during starvation is mainly caused by the increase in the antitoxin degradation rate and the decrease in the growth rate, rather than by the decrease in the translation rates of the toxin and the antitoxin. As toxin-antitoxin modules are very complex systems, even more interactions could be integrated in the models. For example, it would be interesting to examine the influence of the mechanism for the toxicity on the dynamics of a TA module. This mechanism is specific for every toxin-antitoxin module, for example mRNA degradation in the relBE TA module and inhibition of translation in phd/doc. Our model also assumes that the operator of a TA module consists of binding sites with identical affinity. It will be of interest to investigate the dynamics of a TA module with an operator that contains several binding sites with different affinities for the antitoxin. Finally, we would like to develop a more general interacting binding site model, removing the need of simulating all DNA interactions separately. Such a model would allow a more in-depth investigation of the dynamical mechanism leading to the described rare toxin spikes. So far, it seems that these spikes are triggered stochastically and do not exist in the deterministic system, being always monostable. In most systems where pulsed dynamics have been observed, however, they often rely on underlying deterministic bifurcations leading to for instance bistability, oscillations and excitability (for an overview, see Ref. [39]). One such example is for instance the genetic competence in Bacillus subtilis under stress conditions, where a transient cellular state is also initiated stochastically [40]. Three parameter sets were built up, one for the phd/doc, one for the ccdAB and one for the relBE toxin-antitoxin module (Table 1). mRNA transcription is assumed to only take place when the promoter/operator region is unbound, hence , the transcription rate for bound DNA, is zero. The transcription rate for unbound DNA, , is based on a transcription rate of 70 nucleotides/second [41], [42] and the transcript lengths. The translation rates are based on a translation rate of 20 amino acids per second [43]. Furthermore, the parameter set accounts for the fact that CcdA, CcdB, Phd and RelB form dimers in solution, whereas Doc and RelE are monomers. The in vivo translation rates for the antitoxins () are higher than the ones for the toxins due to translational coupling. As such, in order to evaluate the translation rate for the toxin (), the translation rate based on the length was divided by the translational coupling factor (c). The volume factor (V) allows us to convert molar units to molecules/cell, using an E. coli volume of 0.6 (μm)3 [44]. The decay rate of the mRNA () is based on a half life of 5.7 minutes in vivo. Cell division is not explicitly included in the model, but it is implicitly present in the decay rate for the toxin and the complexes AT and TAT. These values were chosen so that the amount of proteins in the cell is halved every generation and the doubling time of E. coli was set at 40 minutes. As the antitoxins are always degraded faster than the corresponding toxins, the antitoxin decay rate was fixed as four times , corresponding to a half life of approximately 15 minutes for CcdA [16]. We include antitoxin degradation in AT and TAT complexes. This is described by the parameter F, which is set at a certain percentage of . Both in the ccdAB and in the phd/doc system, the antitoxin can bind to a high affinity and a low affinity binding site on the toxin. The for the interaction of CcdA at the high affinity binding site, , was determined by Surface Plasmon Resonance (SPR) by De Jonge et al. [45]. The () for this interaction is calculated from this and the for the high affinity toxin-antitoxin interaction, determined by Drobnak et al. [46] using ITC. These kinetic parameters are based on SPR results (Loris and Garcia-Pino, unpublished data) for the phd/doc operon and on SPR results by Overgaard et al. [47] for the relBE operon. The for the interaction of CcdA with one binding site on the operator, , was determined as 3510 (Loris et al., unpublished data); the for this interaction, , is based on a of 2.5 μM [48], and . For the phd/doc operon, the of the antitoxin from the DNA is based on a half life of 30 seconds for a complex of Phd and a single binding site on the operator [49]. The for this interaction is based on the , determined by Garcia-Pino et al. [23] using ITC and this . It was assumed that the for a toxin-antitoxin complex () is equal to the for an antitoxin alone. The higher affinity of this complex for the operator DNA, derived from EMSA experiments, is therefore reflected in the () alone. For the relBE operon, the dissociation rates of the antitoxin and the toxin-antitoxin complex AT from the DNA were determined by Overgaard et al. [20] using SPR, while the corresponding association rates are based on the dissociation constants reported [50]. The of a TAT and an AT complex from the DNA are assumed to be equal in the model with interacting binding sites on the operator, whereas the TAT complex immediately unbinds in the independent binding sites model. When one protein or protein complex interacts at two different sites with proteins or DNA with known affinity, for example when a toxin forms a bridge between two bound antitoxin molecules by binding one antitoxin at the high affinity and one antitoxin at the low affinity binding site, these affinities are multiplied. We assume the for the binding of all proteins and complexes to the DNA or to a DNA-bound protein complex to be equal to the of the antitoxin to DNA, unless a supplementary high affinity toxin-antitoxin interaction is formed in the process. In this case, the is multiplied by ten. In certain simulations, we introduce toxic feedback effects (see also Figure 1). Firstly, we describe a decrease in transcription rate as a function of the free toxin level: Secondly, we consider the decrease in the growth rate (modeled by an equivalent decrease in the dilution rate ) as a function of the free toxin level: Each of these effects is implemented by a Hill-type function, with the toxin threshold, n the Hill factor describing how sharp the transition takes place around and defined as the lowest possible normalized growth rate at very high levels of free Toxin T. We define , the normalized growth rate, as: We use and and unless otherwise stated, we use . In order to obtain a high number of persisters, we decreased the translational coupling, , to 1 instead of 3 in Figure 7. The outlined models were simulated using a Gillespie algorithm which is based on treating the chemical reactions as discrete stochastic events [51]. At each time step, the state of the system is given by the number of molecules (or equivalently: the concentration) of mRNA (M), antitoxin (A), toxin (T), primary complex (AT) and secondary complex (TAT). The operator was defined as having n binding sites, denoted by , being 0 if bound and 1 if unbound. The total operator site is unbound if and bound if . The chemical reactions at the protein level with the rates determined by the parameters specified in Table 1 lead to the changes in the number of molecules as outlined in Table 2.